ARTIFICIAL

INTELLIGENCEA EUROPEAN PERSPECTIVE

EUR 29425 EN Joint

Research
Centre
This publication is a Science for Policy report by the Joint Research Centre (JRC), the European Commission’s
QAGCLACȩ?LBȩILMUJCBECȩQCPTGACȩ'Rȩ?GKQȩRMȩNPMTGBCȩCTGBCLAC@?QCBȩQAGCLRGȏAȩQSNNMPRȩRMȩRFCȩ#SPMNC?LȩNMJGAWK?IGLEȩ
NPMACQQȩ2FCȩQAGCLRGȏAȩMSRNSRȩCVNPCQQCBȩBMCQȩLMRȩGKNJWȩ?ȩNMJGAWȩNMQGRGMLȩMDȩRFCȩ#SPMNC?Lȩ!MKKGQQGMLȩ,CGRFCPȩRFCȩ
European Commission nor any person acting on behalf of the Commission is responsible for the use that might be
made of this publication.

Contact information
Max Craglia (Ed.)
#SPMNC?Lȩ!MKKGQQGML
ȩ(MGLRȩ0CQC?PAFȩ!CLRPC
ȩ2.
ȩ4G?ȩ$CPKG
ȩȩ'QNP?ȩ4
ȩ'2*7
massimo.craglia@ec.europa.eu
Tel.: +39 0332 786 269

EU Science Hub
https://ec.europa.eu/jrc

An online version of this publication is available at: GSSORDB DTQNO@ DTIQBDMOTAKHB@SHNMDTQRBHDMSHɅB@MC
SDBGMHB@KQDRD@QBGQDONQSR@QSHɅBH@KHMSDKKHFDMBDDTQNOD@MODQRODBSHUD

JRC113826
EUR 29425 EN

Print ISBN 978-92-79-97219-5 ISSN 1018-5593 doi:10.2760/936974

PDF ISBN 978-92-79-97217-1 ISSN 1831-9424 doi:10.2760/11251

*SVCK@MSPEȩ.S@JGA?RGMLQȩ-ȑACȩMDȩRFCȩ#SPMNC?Lȩ3LGML
ȩ

The reuse policy of the European Commission is implemented by Commission Decision 2011/833/EU
MDȩȩ"CACK@CPȩȩMLȩRFCȩPCSQCȩMDȩ!MKKGQQGMLȩBMASKCLRQȩ-(ȩ*ȩ
ȩ
ȩNȩȩ0CSQCȩ
is authorised, provided the source of the document is acknowledged and its original meaning
or message is not distorted. The European Commission shall not be liable for any consequence stemming
from the reuse. For any use or reproduction of photos or other material that is not owned
by the EU, permission must be sought directly from the copyright holders.

All content © European Union, 2018 except: cover © European Union, 2018 - graphic elaboration from ©wvihr
- AdobeStock; ©Oksancia - AdobeStock; ©Gstudio Group – AdobeStock.; p. 11 ©metamorworks - AdobeStock;
p. 18 ©zapp2photo - AdobeStock; p. 24 ©metamorworks - AdobeStock; p. 36 ©Jakub Jirsák - AdobeStock;
p. 39 ©adam121 - AdobeStock; p. 42 ©peshkova - AdobeStock; p. 44 ©Christian Horz - AdobeStock; p.
ȩm@WK?LBCQGELQȩȩBM@C1RMAIȩNȩȩm%MPMBCLIMȎȩȩBM@C1RMAIȩNȩȩmȏXICQȩȩBM@C1RMAIȩNȩȩ
©enterlinedesign - AdobeStock; p. 58 ©THANANIT - AdobeStock; p. 62 ©NicoElNino - AdobeStock; p. 67 ©nitsawan
- AdobeStock; p. 70 ©zinkevych - AdobeStock; p. 76 ©phonlamaiphoto - AdobeStock; p. 83 ©metamorworks -
BM@C1RMAIȩNȩȩm(?AI$ȩȩBM@C1RMAIȩNȩȩm%MPMBCLIMȎȩȩBM@C1RMAIȩNȩȩmX?NNNFMRMȩȩBM@C1RMAIȩȩNȩ
ȩmȩTAF?JSNȩȩBM@C1RMAIȩNȩȩm%MPMBCLIMȎȩȩBM@C1RMAIȩNȩȩmQCTCLRWDMSPȩȩBM@C1RMAIȩNȩȩm2GCPLCWȩ
- AdobeStock; p. 102 ©fgnopporn - AdobeStock; p. 114 ©porbital - AdobeStock; p. 120-122 ©Fabian- AdobeStock,
and where otherwise stated.

How to cite this report: Craglia M. (Ed.), Annoni A., Benczur P., Bertoldi P., Delipetrev P., De Prato G., Feijoo C.,
$CPL?LBCXȩ+?AG?Qȩ#
ȩ%MKCXȩ#
ȩ'EJCQG?Qȩ+
ȩ(SLIJCUGRXȩ&
ȩȩ*ŁNCXȩ!M@Mȩ+
ȩ+?PRCLQȩ 
ȩ,?QAGKCLRMȩ1
ȩ,?RGTGȩ1
ȩ
.MJTMP?ȩ
ȩ1?LAFCXȩ'
ȩ2MJ?Lȩ1
ȩ2SMKGȩ'
ȩ4CQLGAȩJSHCTGAȩ*
ȩ"SUJҨDJBM
*OUFMMJHFODF

"
&VSPQFBO
1FSTQFDUJWF, EUR
ȩ#,
ȩ.S@JGA?RGMLQȩ-ȑAC
ȩ*SVCK@MSPE
ȩ
ȩ'1 ,ȩ
ȩBMG 
ȩ(0!
 ARTIFICIAL
INTELLIGENCE
A EUROPEAN PERSPECTIVE
 Table of contents 4

TABLE
#VCASRGTCȩQSKK?PW
OF CONTENTS 8

Part 1: Introduction and AI international landscape 15

1 +MRGT?RGMLȩ?LBȩM@HCARGTCȩMDȩRFGQȩPCNMPR 16

2 About AI 19
2.1 A brief history of AI 19
2.2 Recent developments in machine learning 20
2.3 Recent developments in social robots 21
2.4 Current challenges 22
2.5 Summary and conclusions 23

3 EU in the AI competitive global landscape 25

3.1 2FCȩGLRCPL?RGML?JȩNMJGAWȩAMLRCVRȩ ȩ
3.1.1 USA 25
ȩ -RFCPȩAMSLRPGCQȩ 
3.2 Analysing the key features of the AI landscape 29
ȩ -TCPTGCUȩ 
3.4 Technological capacity 34
3.5 Summary and conclusions 35

4 AI in the EU 37
4.1 Strategies and plans 37
4.1.1 The European Union 37
4.1.2 France 40
4.1.3 United Kingdom 41
4.1.4 Finland 43
ȩ -RFCPȩ#3ȩAMSLRPGCQȩ 
4.2 Summary and conclusions 43

5 The AI ecosystem in China 45

ȩ !FGL?QȩCAMLMKGAȩ?LBȩNMJGAWȩAMLRCVRȩ 
ȩ %MTCPLKCLRȩNMJGAGCQȩ?LBȩGLGRG?RGTCQȩ 
5.3 Regional/local initiatives 49
ȩ 2FCȩQNCAGȏAGRGCQȩMDȩRFCȩ'ȩGLLMT?RGMLȩCAMQWQRCKȩGLȩ!FGL?ȩ 
5.5 Summary and conclusions 51
5 Table of contents

Part 2: Multi-dimensional perspectives 53

6 Ethical and societal perspective 55

6.1 Introduction 55
ȩ -TCPTGCUȩMDȩGLBGTGBS?Jȩ?LBȩAMJJCARGTCȩGKNJGA?RGMLQȩMDȩ'ȩ 
6.2.1 Challenges at individual level 56
6.2.2 Challenges at societal level 57
6.3 Summary and conclusions 60

7 Legal perspective 63
ȩ +?GLȩJCE?JȩAF?JJCLECQȩGBCLRGȏCBȩGLȩ#SPMNC?Lȩ'ȩQRP?RCEGCQȩ 
ȩ -ULCPQFGN
ȩ?AACQQȩ?LBȩQF?PGLEȩMDȩB?R?ȩ 
7.3 The protection of AI inventions/creations
by intellectual property rights 66
7.4 Regulatory approach 68
7.5 Summary and conclusions 69

8 Educational perspective 71
8.1 Introduction 71
8.2 AI impact on skills demand, learning and teaching 72
8.2.1 Direct AI impact on advanced skills demand 72
8.2.2 Impact of AI on learning 72
8.2.3 Impact of AI on teaching 73
8.3 AI skills and academic supply 74
8.4 Summary and conclusions 75

9 Economic perspective 77
ȩ .MRCLRG?JȩGKN?ARȩMDȩ'ȩMLȩHM@Qȩ 
9.2 Potential impact of AI on growth 80
9.3 Potential impact of AI on inequality 82
9.4 Summary and conclusions 84

10 Cybersecurity perspective 87
10.1 Background: AI and cybersecurity 87
10.2 Applications of AI in cybersecurity 88
ȩ "CRCPPCLACȩ?LBȩȏEFRȩ?E?GLQRȩAPGKCȩ 
10.4 Robustness of AI algorithms against malicious action 89
10.5 Summary and conclusions 91
 Table of contents 6

11 Computer processing and energy perspective 95

11.1 Introduction 95
11.2 Assessment of data centre (DC) energy consumption 96
ȩ -NRGMLQȩRMȩGKNPMTCȩRFCȩCLCPEWȩCȑAGCLAWȩMDȩRFCȩGLAPC?QGLEȩ
demand for HPC. 96
11.3.1 CPU advancements: energy saving and parallelisation
computing 97
ȩ 'LLMT?RGTCȩ?LBȩKMPCȩCȑAGCLRȩAMMJGLEȩQWQRCKQ CLEGLCCPGLEȩ
solutions 98
ȩ 'LLMT?RGTCȩGLDP?QRPSARSPCQȩ?PAFGRCARSPCQȩ?LBȩAMLȏESP?RGMLQȩ 
11.4 Summary and conclusions 100

12 Data perspective 103

12.1 The law and economics of access to data 103
12.1.1 The economic characteristics of data 103
12.1.2 Applying economic reasoning to data access 104
12.1.3 Policy intervention in the data market 105
12.2 Summary 108
12.3 Towards a data strategy for public administrations 108
12.3.1 Setting the scene 108
12.3.2 About platforms and APIs 109
12.3.3 From opening data to smart sharing 110
12.4 Summary and conclusions 113

13 Societal resilience perspective 115

13.1 The need for resilience 115
13.2 Measuring resilience to AI 116
13.3 Lessons from the analysis of AI for understanding resilience 119

14 Summary and conclusions 120

Footnotes 122
References 125
ȩ *GQRȩMDȩ@MVCQȩ 136
ȩ *GQRȩMDȩȏESPCQȩ 137 D
Acknowledgements 138
 7 Table of contents

DEEP

LE
ARNING
AI
IC S
OT

ROB

R SECU
BE
RIT
CY

S
HIC
ET
TA

DA

T
CEN ERE
HUMAN

D
AI
 #VCASRGTCȩQSKK?PW 8

EXECUTIVE
SUMMARY
This report presents a European view of PRGȍAG?J however. Among them, current ML algorithms display
'LRCJJGECLAC ' based on independent research QMKCȩMDȩRFCȩAF?P?ARCPGQRGAQȩMDȩ?ȩ@J?AIȩ@MVȩUCȩ?A-
and analysis by the European Commission Joint cess the inputs and outputs but do not understand
Research Centre to inform the debate at the Euro- fully what happens in-between, and how certain
pean level. outputs, including decisions and actions, are
dePGTCBȩ 2FGQȩ A?JJQȩ fMPȩ ?ȩ EPC?RCPȩ CȎMPRȩ RMȩ SLBCPQRand
5Cȩ ȏPQRȩ GLRPMBSACȩ 'ȩ ?Qȩ ?ȩ ECLCPGAȩ RCPKȩ RF?Rȩ PC- RFCGPȩ RFCMPWȩ ?LBȩ RMȩ BCTCJMNȩ CVNJ?GL?@JCȩ ?LBȩ ?Acount-
fers to any machine or algorithm that is capable able algorithms. We also need strong evaluation
of observing its environment, learning, and based
MLȩRFCȩILMUJCBECȩ?LBȩCVNCPGCLACȩE?GLCB
ȩR?ICȩGL-
telligent actions or propose decisions. Autonomy
of decision processes and interaction with other
machines and humans are other dimensions that
need to be considered.

Although many of the methodological develop-

ments in AI date back more than 50 years, the
reason why we now pay so much attention to AI
in general and machine learning (ML) in particular
is that the recent advances in computing power,
availability of data, and new algorithms have led to
K?HMPȩ@PC?IRFPMSEFQȩGLȩRFCȩJ?QRȩQGVȩRMȩQCTCLȩWC?PQȩ
The many applications of AI/ML have started to en-
ter into our everyday lives, from machine transla-
tions, to image recognition and music generation,
?LBȩ ?PCȩ GLAPC?QGLEJWȩ @CGLEȩ CVNJMGRCBȩ GLȩ GLBSQRPW
ȩ
government and commerce TFF
$IBQUFS

.
frameworks that can assess not only the per-
It is likely that we are only at the beginning of this formance but also the quality of AI and build trust
process because the development of ubiquitous sen- in this disruptive technology $IBQUFST



BOE

.
QMPȩLCRUMPIQ
ȩRFCȩ'M2
ȩUGJJȩGLAPC?QCȩCVNMLCLRG?JJWȩRFCȩ The overview of the global and European AI land-
sensing capabilities of AI, the volumes of data on scape shows that there is an intense competition on
which to train the algorithms, and their reach in soci- AI taking place world-wide with three main leaders:
ety through decisions and actions. the USA, Europe, and China $IBQUFS

. Each region
has about one quarter of all key players1 in the AI
The opportunities are many, and in some cases ȏCJB
ȩGLAJSBGLEȩ@MRFȩPCQC?PAFȩ?LBȩGLBSQRPW
ȩ@SRȩF?Qȩa
not yet foreseen. There are also many challenges, BGQRGLARGTCȩBGȎCPCLRȩKGVȩMDȩNJ?WCPQȩUFGJCȩ#SPMNCȩGQȩ
 9 ȩ#VCASRGTCȩQSKK?PW
well balanced in the number of research and
non-PCQC?PAFȩNJ?WCPQ
ȩRFCȩ31ȩF?Qȩ?NNPMVGK?RCJWȩRFree over the need for an agreed ethical framework and
times as many industrial/corporate players as re-
QC?PAFȩMLCQ
ȩ?LBȩ!FGL?ȩF?Qȩ?@MSRȩQGVȩRGKCQȩ?QȩK?LWȩ and uphold the European values enshrined in the
research players as industrial ones. The strength of the
.ȩ 2FCȩ &GEF*CTCJȩ #VNCPRȩ %PMSNȩ
corporate world in the USA is also indicated by its
dominance in the number of start-ups (almost half
of the total worldwide) and venture capital (more
than one third of the total). China, on the other hand,
GQȩK?IGLEȩ?ȩQRPMLEȩCȎMPRȩRMȩRSPLȩPCQC?PAFȩinto pa-
tents, and accounts for almost 60 % of the world
total. It has also put in place a strongly coordinated
approach to AI, including government policy,
GLBSQRPG?Jȩ?NNJGA?RGMLQȩ?LBȩPCQC?PAFȩUGRFȩRFCȩM@HCARive
of becoming the world leader in AI by 2030. This is
an ambitious but achievable target $IBQUFS

.
Europe is currently well positioned in the quality
of its research production, with more than 30 %
guide the development of AI, with shared concerns
?NNJGA?RGMLQȩ RF?Rȩ AJC?PJWȩ @CLCȏRȩ #SPMNC?Lȩ QMAGCRWȩ
Treaties $IBQUFS

established in 2018 by the European Commission
is elaborating a framework for subsequent develop-
ments linked to these values. We also note the high
level of awareness that data is crucial to the devel-
opment of AI, with policy documents at both nation-
al and European level putting strong emphasis on
the need to share data better among all the stake-
holders: the public sector, industry, and the public.
Finding the right way of doing it is challenging, as
analysed in $IBQUFS

of the report.
After the overview and analysis of the AI global and
European landscapes we discuss AI from multiple per-
MDȩ?JJȩN?NCPQȩMLȩ'ȩNS@JGQFCBȩGLȩRMNȩQAGCLRGȏAȩHMSP-
L?JQ
ȩ HSQRȩ @CFGLBȩ RFCȩ 31ȩ ȩ Ϥ
ȩ ?LBȩ GQȩ AMLQGB-
erably ahead of China. Key areas of strength in
Europe on which to build upon are automated and
connected vehicles and robotics.


We note that many European countries as well as
the EC are developing strategies and programmes to
QNCARGTCQȩRMȩ?BBȩQMKCȩBCNRFȩ?LBȩCVNJMPCȩQWLCPEGCQȩ
As we highlight at multiple stages in this report, a
key characteristic of the European way to AI has
to be a strong ethical framework. There is con-
sensus on this principle, and we report $IBQUFS

on many initiatives at both national and European
levels to develop ethical guidelines to frame the
 #VCASRGTCȩQSKK?PW 10

development of AI. We review the main dimensions ?LBȩ ECMEP?NFGAȩ BGQRPG@SRGMLȩ MDȩ RFCȩ ?A?BCKGAȩ MȎCPȩ
that such guidelines should consider and highlight of study topics related to AI. This is relevant to
the potential implications of AI at the level of indi- the discussion in DIBQUFST

BOE
 on possible
viduals and society. We conclude that to build and strategies to start preparing society, and the most
retain trust in AI we need a multi-layered approach vulnerable regions, to the challenges that AI will
that includes the critical engagement of civil so- bring.
ciety to discuss the values guiding and being em-
@CBBCBȩGLRMȩ'
ȩNS@JGAȩBC@?RCQȩGLȩBGȎCPCLRȩDMP?ȩRMȩ
translate these values into strategies and guide-
lines, and responsible design practices that encode
these values and guidelines into AI systems, so that
they are ethical-by-design $IBQUFS

.
The digital
In reviewing the European legal framework for AI,
transformation
particularly with respect to the fundamental rights,
data ownership, and intellectual property, we note of society
the tensions between protecting rights of individuals
?LBȩȏPKQȩ?LBȩCLAMSP?EGLEȩGLLMT?RGMLȩUGRFȩRPWGLEȩRMȩ has just begun:
K?VGKGQCȩMNCLLCQQȩ?LBȩRP?LQN?PCLAWȩ5CȩAMLAJSBCȩ
however that Europe is well placed to establish a
distinctive form of AI that is ethically robust and
AI is central
NPMRCARQȩRFCȩPGEFRQȩMDȩGLBGTGBS?JQ
ȩȏPKQ
ȩ?LBȩQMAGCRWȩ
?Rȩ J?PECȩ $MPȩ CV?KNJC
ȩ RFCȩ %CLCP?Jȩ "?R?ȩ .PMRCARGMLȩ
to this change
Regulation, opposed by many during preparation, is
now perceived as a European asset and is inspiring
and offers major
QGKGJ?Pȩ ?NNPM?AFCQȩ MSRQGBCȩ #SPMNCȩ #VRCLBGLEȩ RFGQȩ
notion, we should consider the high standards of the opportunities
European legal and regulatory landscape in a similar
way to those in environmental quality which are an to improve
asset for Europeans and their future generations to
build upon, not a barrier $IBQUFS

. our lives.
From an educational perspective, we observe
$IBQUFS


that AI has potential positive impacts
on shortages of skills, learning, and teaching.
Three crucial points stand out from the review: The potential impact of AI on the labour market and
ȏPQRJW
ȩRFCȩLCCBȩRMȩSLBCPQR?LBȩ@CRRCPȩFMUȩRFCȩGL- inequality raises concerns in the media, research
teraction with AI impacts human intelligence in and public debates. We analyse the literature and
cognitive capacities in both adults, and even more evidence available to date on the potential impacts,
importantly, children. Secondly, we need to think both positive and negative, with respect to work,
beyond current needs and practices, and consider EPMURF
ȩ?LBȩGLCOS?JGRWȩ'LȩPCJ?RGMLȩRMȩUMPI
ȩUCȩȏLBȩ
how AI is likely to change the relationship between
education and work, and human development.
Thirdly, we highlight possible risks related to AI in
education, particularly privacy and ethical ones.
In this chapter, we also provide an initial overview
that neither theory nor evidence are very conclusive
at the present time. AI could complement and en-
hance human activity, replace an increasing num-
ber of routine tasks, or both. Studies measuring
RFCȩQF?PCȩMDȩHM@Qȩ?RȩFGEFȩPGQIȩMDȩ?SRMK?RGMLȩCVFG@-
 11 ȩ#VCASRGTCȩQSKK?PW

GRȩ?ȩFGEFȩT?PGCRWȩGLȩRFCGPȩȏLBGLEQȩBCNCLBGLEȩMLȩRFCȩ through many channels. The most discussed con-

BCȏLGRGMLȩ ?LBȩ JCTCJȩ MDȩ EP?LSJ?PGRWȩ ?Rȩ UFGAFȩ R?QIQȩ ACPLȩ HM@ȩ NMJ?PGQ?RGML
ȩ GCȩ GLAPC?QCBȩ BCK?LBȩ DMPȩ
?LBȩMAASN?RGMLQȩ?PCȩBCȏLCBȩ $IBQUFS

. FGEFJCTCJ
ȩ FGEFJWȩ N?GBȩ HM@Qȩ MLȩ RFCȩ MLCȩ F?LBȩ ?LBȩ
JMUJCTCJ
ȩ NMMPJWȩ N?GBȩ HM@Qȩ MLȩ RFCȩ MRFCP
ȩ PCBSA-
LMRFCPȩ?PC?ȩMDȩSLACPR?GLRWȩGQȩRFCȩCVRCLRȩRMȩUFGAFȩ RGMLȩ GLȩ HM@ȩ OS?JGRWȩ ?Rȩ RFCȩ JMUCPQIGJJCBȩ CLB
ȩ ?LBȩ
AI has the potential to spur economic growth. ?JQMȩ EPC?RCPȩ BGȑASJRWȩ DMPȩ JMUCPQIGJJCBȩ UMPICPQȩ RMȩ
When considered as a general-purpose technology, ?BHSQRȩ RMȩ AF?LECȩ ?LBȩ ȏLBȩ LCUȩ HM@Q
ȩ UGRFȩ JMLECPȩ
periods of unemployment than those with higher
ȩ GDȩ LMȩ ?ARGMLȩ
'ȩ AMSJBȩ QNPC?Bȩ ?APMQQȩ K?LWȩ HM@Qȩ ?LBȩ GLBSQRPG?Jȩ
sectors, boosting productivity, and yielding strong
OS?JGȏA?RGMLQȩ?LBȩQIGJJQȩ2FCQCȩNMRCLRG?JJWȩLCE?RGTCȩ
consequences on the labour market have a geo-
graphical footprint, as regions and subregions
RF?Rȩ?PCȩ?JPC?BWȩCVNCPGCLAGLEȩEPC?RCPȩBGȑASJRGCQȩGLȩ
terms of unemployment and low level of skills are
JGICJWȩ RMȩ @Cȩ RFCȩ MLCQȩ QSȎCPGLEȩ KMQR
is taken. We clearly need to monitor closely and
research the multiple impacts of AI on the econo-
my in the coming years. We also need to consider
a more proactive strategy to build the resilience
of regions across Europe with a particular regard
to the most vulnerable ones. We return to this in

$IBQUFST

BOE
.
When it comes to cybersecurity, AI is a double-edged
QUMPBȩGRȩA?Lȩ@CȩEPC?RJWȩ@CLCȏAG?JȩRMȩGLAPC?QCȩRFCȩ
security of devices, systems and applications, but
can also empower those who seek to attack sys-
tems and networks and thus become an advanced
tool in the arsenal for cyber-attacks. Moreover, the
robustness of AI against malicious action itself be-
comes an issue, posing the most immediate danger
for the security of cyber-physical systems, in which
AI will be increasingly deployed. There are lines of
PCQC?PAFȩDMASQGLEȩMLȩSLBCPQR?LBGLEȩRFCȩQNCAGȏAȩTSJ-
nerabilities of AI and related attacks, and ways of in-
creasing AI robustness and interpretability and safe-
ty by design. We also need shared, large, high-quality
datasets to train and test algorithms, and agreed
NMQGRGTCȩ EPMURFȩ 2Mȩ RFCȩ CVRCLRȩ RF?Rȩ +*ȩ ECLCP-
.
ates new insights from data, it may also con-
tribute to the production of new ideas and innov-
ation itself. Economic growth models are starting
RMȩCVNJMPCȩT?PGMSQȩQACL?PGMQȩ@SRȩRFCPCȩGQȩLMȩCKNGPG-
cal evidence yet that favours one or the other.
5FCLȩ GRȩ AMKCQȩ RMȩ GLCOS?JGRW
ȩ UCȩ ȏLBȩ RF?Rȩ 'ȩ A?Lȩ ognition that the increasing energy consumption of
?ȎCARȩ SLD?TMSP?@JWȩ RFCȩ BGQRPG@SRGMLȩ MDȩ GLAMKCȩ
frameworks to evaluate them
$IBQUFS

We discuss at several points in the report that the
K?HMPȩNPMEPCQQȩUCȩ?PCȩQCCGLEȩGLȩRFCȩBCTCJMNKCLRȩMDȩ
AI is linked to the rapid advances in computing and
in data availability. With respect to computing, we
are starting to see a paradigm shift due to the rec-
data centres and data transfers will become unsus-
 #VCASRGTCȩQSKK?PW 12

R?GL?@JCȩGLȩRFCȩCP?ȩMDȩRFCȩ'LRCPLCRȩMDȩ2FGLEQȩ?LBȩ%ȩ an opportunity to bootstrap social and economic

LCRUMPIQȩ"?R?ȩRP?ȑAȩ?LBȩNPMACQQGLEȩJM?BQȩ?PCȩJGICJWȩ
to be unprecedented when billions of additional de-
vices will be connected to the internet, sensing the
.
environment, and constantly sending and receiving
data. With this in mind, we are seeing a trend to-
wards more decentralised frameworks of edge and
fog computing where processing is done closer to
the sensors capturing and displaying data, includ-
ing mobile phones and tablets. Industry has started
moving in this direction but there is still a window
of opportunity for European investment, regulatory
frameworks, and standards to shape these develop-
ȩ
KCLRQȩQMȩRF?RȩRFCWȩ@CLCȏRȩ#SPMNC
$IBQUFS
. SMEs in particular, in combination with the planned
With respect to digital data, we review its key eco-
nomic characteristics including economies of scale,
MDȩQAMNC
development in peripheral regions, leveraging the
richness of European diversity and culture to the
full $IBQUFS

We put forward some elements of a possible ap-
proach centred on the network of Digital Innova-
tion Hubs. There are already several hundred hubs
UFGAFȩ ?PCȩ CVN?LBGLEȩ P?NGBJWȩ UGRFȩ ?ȩ R?PECRȩ MDȩ MLCȩ
for each region in Europe. The Communication on
AI (EC, 2018a) foresees a specialised subset of
these hubs on AI to facilitate access to technology
?LBȩILMUFMUȩRMȩNS@JGAȩ?BKGLGQRP?RGMLQȩ?LBȩȏPKQ
AI-on-demand-platform. We suggest putting these
hubs at the centre of local ecosystems comprising
public administrations, local enterprises, education-
ȩ?LBȩLMLPGT?JPWȩ2FCȩȏPQRȩRUMȩNMGLRȩRMȩ?Lȩ al and training establishments, and civil society.
increase in the concentration of data, and hence
information and power, in the hands of a few actors These ecosystems could create local pools of shared
in the internet economy, while non-rivalry creates data among the key actors so that the AI skills de-
potential tensions between opening access to the veloped/upgraded locally could be put to good use in
B?R?ȩQMȩRF?RȩQMAGCRWȩ@CLCȏRQȩRFCȩKMQR
ȩ?LBȩPCQRPGAR- developing algorithms and solutions based on local
GLEȩ ?AACQQȩ QMȩ RF?Rȩ RFCȩ B?R?ȩ FMJBCPȩ @CLCȏRQȩ KMQRȩ
2FCQCȩRCLQGMLQȩ?PCȩPCȐCARCBȩGLȩRFCȩJCE?JȩDP?KCUMPI
ȩ F?QȩGRQȩMULȩQNCAGȏAȩKGVRSPCȩMDȩNPGMPGRGCQȩGLȩPCJ?RGMLȩ
as discussed earlier, and make the development of
data strategies in an open and globalised environ-
ment a particular challenge. We argue, however,
ȩ
that if we apply the lessons of successful internet
companies to the European public sector and de-
velop ecosystems based around public platforms, it
.
is possible to create large pools of shared data that
are semantically well structured and labelled, and
data to address local needs. Each European region
to its environment, economy, demography, health,
and so on. Therefore, this approach could provide
?LȩCVACJJCLRȩMNNMPRSLGRWȩRMȩF?PLCQQȩJMA?JȩAPC?RGTGRW
culture, and knowledge of the territory to create so-
cially relevant and people-centred AI and develop di-
verse and inclusive AI systems $IBQUFST

BOE

From this multi-disciplinary analysis, we draw the
A?LȩDSCJȩLCUȩ'ȩ?NNJGA?RGMLQȩGLȩBGȎCPCLRȩBMK?GLQȩ'Lȩ following conclusions:
this way we can open access to the data, develop
the market, serve the public, and enrich the data We are only at the beginning of a rapid period of
commons at the same time $IBQUFS

. transformation of our economy and society due to
the convergence of many digital technologies. AI is
We conclude our multi-perspective analysis of AI ACLRP?JȩRMȩRFGQȩAF?LECȩ?LBȩMȎCPQȩK?HMPȩMNNMPRSLGRGCQȩ
focusing on the concept of resilience which is use- to improve our lives.
ful to frame a European approach to AI addressing
BGȎCPCLRȩ NF?QCQȩ NPCTCLRGML
ȩ ?LRGAGN?RGML
ȩ NPCN There is strong global competition on AI among the
aration, but also adaptation and transformation to USA, China and Europe. The USA leads for now but
@MSLACȩ@?AIȩDPMKȩRFCȩCȎCARQȩMDȩ?ȩQFMAIȩ1CCLȩDPMKȩ China is catching up fast and aims to lead by 2030.
this perspective, AI could not only be an engine for For the EU, it is not so much a question of winning
growth and change in Europe but also become MPȩJMQGLEȩ?ȩP?ACȩ@SRȩMDȩȏLBGLEȩRFCȩU?WȩMDȩCK@P?AGLE
 13 ȩ#VCASRGTCȩQSKK?PW
RFCȩMNNMPRSLGRGCQȩMȎCPCBȩ@Wȩ'ȩGLȩ?ȩU?WȩRF?RȩGQȩFS-
man-centred, ethical, secure, and true to our core
values.
The EU Member States and the European Commis-
sion are developing coordinated national and Euro-
pean strategies, recognising that we can only suc-
ceed together. We can build on our areas of strength
GLAJSBGLEȩ CVACJJCLRȩ PCQC?PAF
ȩ JC?BCPQFGNȩ GLȩ QMKCȩ to national and European, across institutions, indus-
industrial sectors like automotive and robotics, a sol-
id legal and regulatory framework, and very rich cul-
tural diversity at both regional and subregional level.
'RȩGQȩECLCP?JJWȩPCAMELGQCBȩRF?Rȩ'ȩA?LȩȐMSPGQFȩMLJWȩGDȩ
supported by a robust computing infrastructure and
good quality data:
• 5GRFȩ PCQNCARȩ RMȩ AMKNSRGLE
Ethical and secure-by-design algorithms are crucial
to build trust in this disruptive technology, but we
also need a broader engagement of civil society on
the values to be embedded in AI and the directions
for future development.
This social engagement should be part of the effort
to strengthen our resilience at all levels from local,
try and civil society. Developing local ecosystems of
skills, computing, data, and applications can foster
the engagement of local communities, respond to
their needs, harness local creativity and knowledge,
and build a human-centred and socially driven AI.
We still know very little about how AI will impact the
way we think, make decisions, relate to each other,
ȩ UCȩ GBCLRGȏCBȩ ?ȩ ?LBȩFMUȩGRȩUGJJȩ?ȎCARȩMSPȩHM@Qȩ2FGQȩSLACPR?GLRWȩA?Lȩ@Cȩ
window of opportunity for Europe to invest a source of concern but is also a sign of opportunity.
in the emerging new paradigm of computing The future is not yet written. We can shape it based
distributed towards the edges of the network on our collective vision of what future we would like
in addition to centralised facilities. This will to have. But we need to act together and act fast.
?JQMȩQSNNMPRȩRFCȩDSRSPCȩBCNJMWKCLRȩMDȩ%ȩ?LBȩ
the IoT.

• With respect to data, we argue in favour of

learning from successful internet companies,
opening access to data and developing interac-
Uncertainty
RGTGRWȩUGRFȩRFCȩSQCPQȩP?RFCPȩRF?LȩHSQRȩ@PM?BA?QR-
ing data. In this way, we can develop ecosys-
about the impact
RCKQȩMDȩNS@JGAȩ?BKGLGQRP?RGMLQ
ȩȏPKQ
ȩ?LBȩAGTGJȩ
QMAGCRWȩCLPGAFGLEȩRFCȩB?R?ȩRMȩK?ICȩGRȩȏRȩDMPȩ'ȩ
of AI can
applications responding to European needs. be a concern
5Cȩ QFMSJBȩ CK@P?ACȩ RFCȩ MNNMPRSLGRGCQȩ ?ȎMPBCBȩ @Wȩ
'ȩ@SRȩLMRȩSLAPGRGA?JJWȩ2FCȩ@J?AIȩ@MVȩAF?P?ARCPGQRGAQȩ
but is also
of current ML make them obscure even to special-
ists. AI systems are currently limited to narrow and
an opportunity:
UCJJBCȏLCBȩ R?QIQ
ȩ ?LBȩ RFCGPȩ RCAFLMJMEGCQȩ GLFCPGRȩ
imperfections from their human creators, such as the future
RFCȩ UCJJPCAMELGQCBȩ @G?Qȩ CȎCARȩ NPCQCLRȩ GLȩ B?R?ȩ 5Cȩ
should challenge the shortcomings of AI and work to- is not yet written.
wards strong evaluation strategies, transparent and
reliable systems, and good human-AI interactions. We can, and should,
shape it.
 PART 1:
INTRODUCTION AND
INTERNATIONAL AI
LANDSCAPE
 Part I - Introduction and AI International Landscape 16

MOTIVATION
AND OBJECTIVE
OF THIS REPORT
PRGȍAG?J'LRCJJGECLAC'UGJJ@PGLEK?LWMN- AMLRCVRQ
ȩ CEȩ RMȩ ESGBCȩ SQȩ RMȩ ?ȩ BCQRGL?RGML
ȩ FCJNȩ
NMPRSLGRGCQx SQȩ ȏLBȩ RFCȩ LCVRȩ QMLEȩ RMȩ JGQRCLȩ RM
ȩ ȏJRCPȩ MSPȩ QN?Kȩ
'ȩGQȩ?ȩRCAFLMJMEGA?JȩȏCJBȩRF?RȩNMUCPQȩK?LWȩ?NNJG- CK?GJQ
ȩMPȩFCJNȩSQȩRCVRGLEȩKCQQ?ECQȩ'LȩRFCȩAMKGLEȩ
cations we use daily, often without knowing it. As it years, AI-supported medical diagnostics may be-
develops further, it has the potential to bring many AMKCȩ GLAPC?QGLEJWȩ AMKKMLȩ #T?JS?RGLEȩ HM@ȩ A?LBG-
changes to both economy and society. This might B?RCQȩMPȩNPCBGARGLEȩPCMȎCLBGLEȩAMSJBȩ@CȩQSNNMPRCBȩ
unfold in a similar way to the invention of the World by the responsible use of transparent algorithms
Wide Web 25 years ago which, combined with the helping to overcome personal or institutional
QS@QCOSCLRȩBGȎSQGMLȩMDȩKM@GJCȩNFMLCQ
ȩMNCLCBȩRFCȩ @G?QCQȩQCC
ȩDMPȩCV?KNJC
ȩ$?JG?EI?ȩCRȩ?J
ȩȩ?LBȩ
way to the massive increase the use of the internet
across the world.
Looking back over these 25 years, we can see how
early views about the death of distance (&DPOPNJTU,


ȩ?LBȩRFCȩBCKGQCȩMDȩAGRGCQȩ%GJBCPȩGLȩ)?PJE??PBȩ
and Malone, 1995) as a result of the internet were
T?QRJWȩ CV?EECP?RCBȩ AGRGCQȩ F?TCȩ AMLRGLSCBȩ RMȩ EPMUȩ
and in fact increased their role as the hubs of in-
novation. The internet has profoundly changed the
way we work, learn, and socialise, created entirely
Tollenaar et al., 2013). Robots able to understand
RFCȩAMKNJCVGRWȩMDȩQMAG?Jȩ?LBȩFSK?Lȩ@CF?TGMSPȩF?TCȩ
the potential to improve human cognitive perfor-
mance (Belpaeme et al., 2018), and support the
development of skills, e.g. in children with autism
(Scassellati et al., 2018).
x?LBK?LWAF?JJCLECQ
AI can solve a wide range of problems, but we still
know very little about how it will impact the way
we think, make decisions, or relate to each other
new products, services, and business models, and %MKCXȩCRȩ?J
ȩȩ'ȩGQȩQRGJJȩLMRȩUCJJȩSLBCPQRMMBȩ
reshaped international relations and markets. We even by specialists. AI systems are currently limited
have seen new challenges emerging, including cy- RMȩ L?PPMUȩ ?LBȩ UCJJBCȏLCBȩ R?QIQ
ȩ ?LBȩ RFCGPȩ RCAF-
ber-threats, misinformation, loss of privacy, and the nologies inherit imperfections from their human
BGQNJ?ACKCLRȩ MPȩ JMQQȩ MDȩ HM@Qȩ GLȩ K?LWȩ QCARMPQ
ȩ @SRȩ APC?RMPQ
ȩ QSAFȩ ?Qȩ RFCȩ UCJJPCAMELGQCBȩ @G?Qȩ CȎCARȩ
UCȩF?TCȩ?JQMȩQCCLȩK?LWȩ@CLCȏRQȩQSAFȩ?QȩS@GOSG- present in data and algorithms. We should there-
tous access to information and services. fore be aware of the potential shortcomings of AI
and work towards agreed ways to evaluate algo-
The recent developments in AI are the result of rithms, build transparent and reliable systems, and
increased availability of processing power, im- design good human-AI interactions.
NPMTCKCLRQȩ GLȩ ?JEMPGRFKQȩ ?LBȩ RFCȩ CVNMLCLRG?Jȩ
growth in the volume and variety of digital data,
and increased funding. We use AI today in many
 17 Introduction and AI International Landscape

5FWRFGQPCNMPRx UCJJȩ?AFGCTCȩRFCȩM@HCARGTCȩMDȩ@CAMKGLEȩRFCȩBMKGL
Hundreds of reports are published each year on AI ant AI player in the world by 2030, overtaking the
?LBȩ PCJ?RCBȩ RCAFLMJMEGCQ
ȩ GLȩ ?BBGRGMLȩ RMȩ QAGCLRGȏAȩ USA. Whilst we are more familiar with USA compa-
papers. This clearly shows the high level of interest nies, technologies and applications, we know little
GLȩRFGQȩQS@HCAR
ȩ@SRȩBMȩUCȩPC?JJWȩLCCBȩ?LMRFCPȩPCNMPR about developments in China and for this reason
we pay comparatively more attention to China in
Most of these publications come from industrial or this report, taking advantage of bespoke research
national interests. Therefore, an independent Euro-
pean perspective underpinned by original research
and not focused on a single policy domain is miss-
ing. Hence the value of this report from the Joint
Research Centre (JRC), the science and knowledge
management service of the European Commission
(EC). This report is also timely because many Euro-
pean countries are developing national strategies
for AI (see Chapter 4) to embrace this technologic-
?JȩȏCJBȩ?LBȩNPCN?PCȩQMAGCRWȩDMPȩRFCȩAF?LECQȩGRȩUGJJȩ
bring. The EC is working with the Member States
on a coordinated Action Plan for AI. A High-Level
BMLCȩMLȩRFGQȩQS@HCARȩDMPȩRFCȩ(0!
In.?PR we look at AI from multiple perspectives:
ethical, legal, educational, economic, cybersecurity,
computing and energy, data-related, and societal
resilience. We discuss each perspective in turn but
they are intertwined in multiple ways and should be
considered all together in framing a European way
to the development and use of AI.
5F?RGQLMRGLRFCPCNMPR
The report does not include any analysis of AI appli-
#VNCPRȩ %PMSNȩ MLȩ 'ȩ F?Qȩ ?JQMȩ @CCLȩ CQR?@JGQFCBȩ @Wȩ A?RGMLQȩGLȩBGȎCPCLRȩQCARMPQȩMDȩRFCȩCAMLMKWȩ2FGQȩUGJJȩ
the EC to make recommendations on policy and @CȩRFCȩQS@HCARȩMDȩ?ȩQCN?P?RCȩ(0!ȩPCNMPRȩGLȩNPCN?P
investment, and set guidelines on the ethical de- ation, as well as the focus of an AI-Watch facility
velopment of AI. The latter are crucial as there is to monitor and analyse AI developments launched
a strong consensus in Europe that our AI should be
ȩ
human-centric, with ethics and security by design,
and supporting the core values of the European
Union: human dignity, freedom, democracy, equal-
ity, rule of law and human rights.
x?LBGRQM@HCARGTC
To provide a balanced assessment of opportunities
and challenges for AI from a European perspective,
and support the development of European action in
GLȩ"CACK@CPȩȩ-RFCPȩ(0!ȩPCNMPRQȩMLȩPCQGJGCLAC
fairness, mobility, cybersecurity, hybrid threats,
and China will be published soon, complementing
and adding details to some of the arguments ad-
dressed in this report.
5F?RWMSQFMSJBR?IC?U?W
Currently, the global competition on AI is largely be-
tween the USA and China at the present time. For the
EU, it is not so much a question of winning or losing a
RFCȩEJM@?Jȩ'ȩAMLRCVRȩ P?ACȩ@SRȩMDȩȏLBGLEȩRFCȩU?WȩMDȩCK@P?AGLEȩRFCȩMNNMP-
RSLGRGCQȩMȎCPCBȩ@Wȩ'ȩGLȩ?ȩU?WȩRF?RȩGQȩFSK?LACLRPCB
ȩ
5F?RGQGLRFCPCNMPR ethical, secure, and true to our core values.
We have organised the report into two sections:
in .?PR  we give a brief introduction to AI, and 5Cȩ F?TCȩ K?LWȩ ?PC?Qȩ MDȩ QRPCLERF
ȩ GLAJSBGLEȩ CV-
analyse the EU positioning in the global AI land- ACJJCLRȩ PCQC?PAF
ȩ CVRPCKCJWȩ PGAFȩ ASJRSP?Jȩ BGTCPQGRW
ȩ
scape, including key features and leading players, and leadership in some industrial sectors, such as
research and technological capacity, and areas automotive and robotics.
MDȩ QRPCLERFȩ 5Cȩ RFCLȩ AMLQGBCPȩ ?ȩ DCUȩ CV?KNJCQȩ MDȩ
Member States’ policies and strategies to assess
points of synergy and where European action can
best add value. This is followed by an analysis of AI
in China since this country has set out, and could
We should therefore not feel intimidated or fearful
about the future. However, we need to act and shape
this future because EU countries are stronger to-
gether in this global competition than acting alone.
 2. About AI 18

"*
BVHNFOUT
PVS
WJTJPO
PG
UIF
XPSME

SUMMARY

AI is a generic term that refers to any machine or algorithm that is capable of obser-
TGLEȩGRQȩCLTGPMLKCLR
ȩJC?PLGLE
ȩ?LBȩ@?QCBȩMLȩRFCȩILMUJCBECȩ?LBȩCVNCPGCLACȩE?GLCB
ȩ
R?IGLEȩGLRCJJGECLRȩ?ARGMLȩMPȩNPMNMQGLEȩBCAGQGMLQȩ2FCPCȩ?PCȩK?LWȩBGȎCPCLRȩRCAFLMJMEGCQȩ
RF?RȩD?JJȩSLBCPȩRFGQȩ@PM?Bȩ'ȩBCȏLGRGMLȩRȩRFCȩKMKCLR
ȩML techniques are the most
widely used.

In this chapter, we introduce the basic concepts of AI and highlight recent develop-
ments. The opportunities opened up by AI are many, and in some cases not yet fore-
QCCL
ȩ@SRȩQMȩ?PCȩRFCȩAF?JJCLECQȩ2FCQCȩUGJJȩ@CȩCVNJMPCBȩGLȩN?PRGASJ?PȩGLȩ.?PRȩȩMDȩRFCȩ
report, but in this chapter, we highlight that the algorithms powering AI display all the
AF?P?ARCPGQRGAQȩMDȩ?ȩ@J?AIȩ@MVȩUCȩA?Lȩ?AACQQȩRFCȩGLNSRQȩ?LBȩMSRNSRQ
ȩ@SRȩBMȩLMRȩSL-
derstand fully what happens in-between, and how certain outputs, including decisions
?LBȩ?ARGMLQ
ȩ?PCȩBCPGTCBȩ%GTCLȩRFGQȩK?HMPȩBP?U@?AI
ȩGRȩGQȩGKNCP?RGTCȩRMȩincrease our
QAGCLRGȏAȧSLBCPQR?LBGLE
ȧBCTCJMNȧRP?LQN?PCLRȧKCAF?LGQKQȧRMȧ?QQCQQȧRFCȧOS?JGRWȧ
?LBȧNCPDMPK?LACȧMDȧ'ȧQMȧRF?Rȧ?Qȧ?ȧQMAGCRWȧUCȧ?PCȧ?@JCȧRMȧ@SGJBȧ?LBȧPCR?GLȧRPSQR
in this disruptive technology.
 19 2. About AI

ABOUT AI
2P?BGRGML?JJW
ȩ PRGȏAG?Jȩ 'LRCJJGECLACȩ 'ȩ PCDCPQȩ RMȩ
machines or agents that are A?N?@JCȧMDȧM@QCPTGLEȧ
Although AI
RFCGPȧ CLTGPMLKCLR
ȧ JC?PLGLE
ȧ ?LBȧ @?QCBȧ MLȧ RFCȧ
ILMUJCBECȧ ?LBȧ CVNCPGCLACȧ E?GLCB
ȧ R?IGLEȧ GLRCJ-
has a long history
JGECLRȧ?ARGMLȧMPȧNPMNMQGLEȧBCAGQGMLQ.
of development,
2MȩB?RC
ȩRFCPCȩGQȩLMȩ?EPCCBȩBCȏLGRGMLȩMDȩUF?RȩAML-
QRGRSRCQȩ'ȩ5FCLȩPCȐCARGLEȩMDȩ'ȩ?LBȩK?AFGLCQ
ȩRFCȩ
recent major
general public often thinks about humanoid robots
which have always been part of the public’s per-
breakthroughs
ception of intelligent machines, as popularised in
QAGCLACȩ ȏARGMLȩ LMTCJQȩ ?LBȩ ȏJKQȩ 'Lȩ D?AR
ȩ PM@MRGAQȩ GQȩ
are due to the
MLJWȩMLCȩQS@ȏCJBȩMDȩ'ȩ?LBȩK?ICQȩSQCȩMDȩK?LWȩ+*ȩ
techniques, including neural networks.
convergence
The origins of AI date back to the 1940s and 50s,
of increased
but it is only recently that it has entered into our
CTCPWB?Wȩ JGDC
ȩ DPMKȩ ȏJRCPGLEȩ MSRȩ RFCȩ QN?Kȩ GLȩ MSPȩ
computing
K?GJ@MVȩ QCCȩ MVȩ 
ȩ RMȩ ?SRMK?RGA?JJWȩ R?EEGLEȩ MSPȩ
loved ones in family photos, to real-time machine
capacity,
translation, and many more applications in busi-
ness, industry, defence, and social media.
availability
of data, and
Ɏɛɤ #?PJWQSAACQQA?QCMD+*GL
Ɏɛɤǩ
AW@C
AW@CPQCASPGRW1N?KȍJRCPGLE
new algorithms.
Wee are all familiar with the sense of annoyance when our inbox
W
DDBQPȭIIBATFQEGRKHJ>FILOPM>J'KQEBB>OIVP
QEB>MMIF
DB QPȭI  @PGCDFGQRMPWMD'
cation
ca
a of a ML technique called Bayesian Networks brought a
new generation
ge of JLOBBȬB@QFSBBJ>FIȭIQBOP that were able to There have been ups and downs in the history of
adapt themselves to their environment, learning to detect spam AI with MPHJDCBTFE approaches in the 1950s and
BJ>FI
BJ>FI?V>K>IVPFKDI>ODBNR>KQFQFBPLCP>JMIBP2KPLIF@FQBABJ>FI early 60s, LOPXMFEHFCBTFE
FYQFSU
TZTUFNT in the
has become
be not just a nuisance but a potential security threat as 1970s and 80s, and EBUBESJWFO approaches (from
mail messages
m are used to distribute malware, conduct phishing 2000 onwards) with periods of disillusionment and
@>JM>
@>JM>FDKP LO @>OOV LRQ Q>ODBQBA >QQ>@HP +LABOK BJ>FI ȭIQBOP reduced funding in-between. We are now at the
@LKQFKR
@LKQFKRB QL ȭDEQ QEBPB QEOB>QP BJMILVFKD a new generation of @CEGLLGLEȩ MDȩ ?ȩ LCUȩ NF?QCȩ MDȩ FGEFȩ CVNCAR?RGMLQ
ȩ
algorithms and AI methods QL@OB>QBJLOBBȬB@QFSB>KAOBPFIFBKQ
algorith fuelled by vastly increased computing processing
counte
countermeasures. capabilities and data. Their combination supports
new developments such as Machine Learning (ML)
 2. About AI 20

RF?RȩF?QȩCKCPECBȩ?Qȩ?ȩQS@ȏCJBȩMDȩ'ȩRMȩNPCBGARȩ?ȩ@C- TGQGRȩ MJBȩ ?PRGȏAG?Jȩ LCSP?Jȩ LCRUMPIȩ KMBCJQ

ȩ RP?GLGLEȩ
F?TGMSPȩDPMKȩK?LWȩCV?KNJCQȩEGTCLȩ?QȩGLNSRȩ
ML represents a paradigmatic shift in computing.
Traditionally, a programmer would write computer
code setting the rules needed to process data inputs
to get an answer as output. In ML, the AMKNSRCPȧPC-
ACGTCQȧGLNSRȧB?R?ȧ?QȧUCJJȧ?QȧRFCȧ?LQUCPQȧCVNCARCBȧ
DPMKȧRFCȧB?R?
ȧ?LBȧRFCȧ+*ȧ?ECLRȧLCCBQȧRMȧNPMBSACȧ NPMACQQȩKGJJGMLQȩMDȩQCLRCLACQȩRMȩȏLBȩMSRȩRFCȩAML-
RFCȧPSJCQ (see Fig. 1). These rules can then be applied
to new data to produce original answers. Lȧ+*ȧQWQ-
RCKȧGQȧRP?GLCBȧP?RFCPȧRF?LȧCVNJGAGRJWȧNPMEP?KKCB.
Figure 1. Paradigm shift in ML2
Source: after Chollet, 2017
them with very large datasets to address some
F?PBȩNPM@JCKQȩQSAFȩ?QȩM@HCARȩPCAMELGRGMLȩDPMKȩBGE-
ital images and machine translation. Whereas it
was previously believed that computers needed to
understand language and its structures before they
AMSJBȩRP?LQJ?RCȩRCVRȩ?LBȩQNCCAFȩDPMKȩMLCȩJ?LES?ECȩ
to another, for many practical uses it is enough to
RCVRQȩ UFCPCȩ UMPBQȩ ?NNC?Pȩ ?LBȩ K?ICȩ RP?LQJ?RGMLQȩ
without linguistic knowledge. This has considerably
speeded up the process, leading to real-time trans-
lation applications now available on most mobile
phone platforms.
Third, there have been new advances in ML al-
gorithms. In addition, specialised open source ML
software libraries have become available that
make the creation and testing of ML algorithms
easier. These trends started to came together
?PMSLBȩȩQCCȩ MVȩ
Whilst ML is the generic class of algorithms that
learn from the data, their accuracy depends very
much on the quality of the training dataset, and
how well they have been structured, semantical-
ly labelled, and cleaned by humans to make them
representative of the problem to tackle, and reduce
the number of parameters in the data.

0CACLRBCTCJMNKCLRQGL+*
The recent interest in AI, and ML in particular, re-
sults from three parallel developments. First, in-
creasingly realistic computer games have required
 > JRIQFI>VBO KBQTLOH QO>FKBA RPFKD Nvidia’s graphics
specialised graphic processors. When the PC graph-
ics card manufacturer Nvidia published the CUDA
programming interface to its graphics accelerator
cards in 2007, it became possible to take advan-
tage of fast parallel computing at low cost.
ɎɛɤǪ2SPLGLENMGLRGLGK?ECPCAMELGRGML
'K 
processor cardsPELTBALRQPQ>KAFKDMBOCLOJ>K@BFK>TBIIHKLTK
image recognition competition. The competition was based on
the ImageNet database that contains about  JFIIFLK ERJ>K
annotated digital images. The ImageNet Large Scale Visual
Recognition Challenge (ILSVRC) is now one of the main benchmarks
for progress in AI.'QPL?GB@QABQB@QFLK>KA@I>PPFȭ@>QFLK@E>IIBKDB
Second, very large amounts of data have become uses JFIIFLK images for training, with AFȬBOBKQ
available as computers and their users have been QVMBPLCL?GB@QP'K
QEB?BPQKBRO>IKBQTLOHPTBOB>?IBQL
networked. The digitalisation of images, videos, guess the correct object category with MBO@BKQ|QLM
TMGACȩ?LBȩRCVRȩF?QȩAPC?RCBȩ?LȩCLTGPMLKCLRȩGLȩUFGAFȩȩ 5’ accuracy, meaning that the correct object class was among
ML thrives. This has allowed AI researchers to re- QEBȭSBJLPQMOL?>?IB@I>PPBP>PBPQFJ>QBA?VQEBKBQTLOH
 21 2. About AI

• %FFQ
-FBSOJOH
(DL) is a subset of ML that can

AMNCȩUGRFȩLMGQGCPȩB?R?ȩ@WȩGLAPC?QGLEȩQGELGȏA?LR-
ly the number of neural layers and neurons3,
and the amount of data used for the training.
"*ȩF?QȩD?AGJGR?RCBȩM@HCARȩPCAMELGRGMLȩGLȩGK?ECQ
ȩ to analyse and synthesise given its expressive character and
video labelling, and activity recognition, and is
K?IGLEȩQGELGȏA?LRȩGLPM?BQȩGLRMȩMRFCPȩ?PC?QȩMDȩ
perception, such as audio, speech, and natural
language processing.
• 3FJOGPSDFNFOU
-FBSOJOH
(RL) is another set
MDȩ ?JEMPGRFKQȩ RF?Rȩ DMASQȩ MLȩ CVNCPGCLACBPGT-
en sequential decision-making, i.e. they make
QMDRU?PCȩ?ECLRQȩR?ICȩ?ARGMLȩRMȩK?VGKGQCȩQMKCȩ
notion of cumulative reward. RL combined with
DL is the basis of many recent successes in
Ɏɛɤǫ"*DMPQGLEGLENPMACQQGLE
The singing voice has traditionally been a challenging instrument
ment
the variety of timbre and expressive resources that characterise
erise
styles and singers. However, OB@BKQA>Q>AOFSBK+* techniques,
ques,
especially DL, have substantially boosted the quality and
accuracy of singing processing techniques.
Singing synthesisers might soon make them indistinguishable from
human singing. As ML algorithms can also compose new music
and imitate the style of classic composers of the past, we may
PLLK?B@LKCOLKQBATFQE>J>GLOPEFȱFKLRORKABOPQ>KAFKDLCTE>Q
TE>Q
music is, how we experience it, as well as new models of music
MOLAR@QFLK
AFPQOF?RQFLK>KA@LKPRJMQFLK%LJBWBQ>I


AMKNJCVȩ E?KCQ
ȩ QSAFȩ ?Qȩ %M
ȩ .MICPȩ ?LBȩ "MR?
ȩ
where computers have been able to beat lead-
GLEȩFSK?LȩCVNCPRQȩGLȩRFCȩȏCJBȩ.P?ARGA?Jȩ?NNJG-
cations of these algorithms are starting to be
applied in many domains, such as autonomous
ȩ UFGJCȩ +GAPM-
driving, unmanned aerial vehicles, stock mar-
kets, and defence.
0C?JRGKCȩ RP?LQJ?RGMLȩ MDȩ @MRFȩ RCVRȩ ?LBȩ TMGACȩ F?TCȩ
@CCLȩ BCTCJMNCBȩ @Wȩ %MMEJCȩ ?LBȩ ?GBS
soft researchers have created algorithms that read
a document and answer questions about it almost
'Lȩ?BBGRGMLȩRMȩGK?ECȩ?LBȩRCVRȩPCAMELGRGMLȩ?LBȩE?K-
GLE
ȩ RFCȩ ȏCJBȩ MDȩ FSK?LAMKNSRCPȩ GLRCP?ARGMLȩ F?Qȩ
seen many impressive improvements recently as
a result of the AI applications. A long-standing
goal has been to enable people to have a natural
?QȩUCJJȩ?Qȩ?ȩFSK?Lȩ2FGQȩGQȩ?ȩK?HMPȩKGJCQRMLCȩGLȩRFCȩ
push to have search engines such as Bing and intelli-
gent assistants such as Cortana interact with people
and provide information in more natural ways, much
as people communicate with each other.
conversation with computers, as they would with
each other. In recent years, the application of DL
and the combination of multiple AI components,
including natural language processing, voice syn-
thesis and voice translation, have revolutionised the
interaction we have with computers, as discussed
 0CACLRBCTCJMNKCLRQGLQMAG?JPM@MRQ
Advances in robotics have allowed AI to be embed-
ded in physical entities that take into consideration
the characteristics of their ‘body’ as well as those of
the physical and social environment in which they
GLȩ1CARGMLȩȩ@CJMUȩ?LBȩCVCKNJGȏCBȩ@WȩRFCȩTGPRS- operate. This raises the potential for socially intel-
al assistants now available on most mobile phone JGECLRȩ PM@MRGAȩ ?ECLRQȩ RMȩ AMCVGQRȩ UGRFȩ FSK?LQ
ȩ ?Qȩ
platforms, such as Apple’s Siri, Microsoft’s Cortana in the case of co-workers, personal companions or
?LBȩK?XMLQȩJCV? self-driving vehicles.

2FCȩ CVRCLRȩ MDȩ NPMEPCQQȩ GLȩ AMKNSRCPECLCP?RCBȩ Advances in perception systems make it possible to
speech is such that even in singing it is increasing- translate low-level signals into high-level abstract
JWȩBGȑASJRȩRMȩBGQRGLESGQFȩ?ȩFSK?LȩDPMKȩ?ȩK?AFGLCȩ concepts. For instance, natural language processing
QCCȩ MVȩȩ (NLP) techniques connect sounds from speech with
language semantics. This contributes to creating a
perception-cognition-action loop for autonomous
 2. About AI 22

robots that makes them able to understand and in-
which would involve the development
teract with the surrounding environment.
These developments in robotics bring additional
perspectives to the integration of AI in our every-
day life: (i) the agents are becoming increasingly
(FOFSBM
"*
of consciousness, is many years away. The current
AMLQCLQSQȩ MDȩ RFCȩ NPGT?RCQCARMPȩ CVNCPRȩ AMKKSLGRWȩ
is that general AI will not be achieved for several
decades (NSTC 2016a).
A?N?@JCȩMDȩNCPACGTGLEȩRFCȩAMLRCVRȩUGRFGLȩUFGAFȩRFCWȩ  !SPPCLRAF?JJCLECQ
act by developing an understanding of the human
NFWQGA?Jȩ CLTGPMLKCLR
ȩ GLAJSBGLEȩ AMKNJCVȩ FSK?Lȩ Each specialised sub-area of AI has its own specif-
social behaviour; (ii) the embodied form of AI can ic challenges, but here we focus on those related
physically act upon and shape human environment to algorithm design and evaluation in ML as this
?LBȩNMRCLRG?JJWȩ?ȎCARȩFSK?Lȩ@CF?TGMSPȩ?QȩUCJJȩ?Qȩ is one of the main areas of development, as dis-
human to human social interaction not only at the ASQQCBȩ ?@MTCȩ -RFCPȩ AF?JJCLECQ
ȩ GLAJSBGLEȩ CRFGA?J
ȩ
individual level but also in small groups; and (iii) the legal and cybersecurity, are discussed in Chapters
interaction between humans and intelligent robot- 
ȩ ȩ ?LBȩ 
ȩ ?LBȩ @Wȩ BGȎCPCLRȩ QAGCLRGQRQȩ QCCȩ ?JQMȩ
ic systems increasingly involves human emotions %MKCXȩCRȩ?J
ȩ
which might stimulate empathic behaviour towards
?PRGȏAG?Jȩ QWQRCKQȩ ?LBȩ F?TCȩ ?Lȩ GKN?ARȩ MLȩ FSK?Lȩ
development over time.

'LȩRFGQȩAMLRCVR
ȩ?ȩPCJ?RGTCJWȩLCUȩQRP?LBȩMDȩPCQC?PAFȩ
has emerged, which investigates the ways in which
humans interact with robots and how various
We need
PM@MRȩ @CF?TGMSPQȩ ?LBȩ BCQGELQȩ ?ȎCARȩ FSK?Lȩ AME-
nitive and socio-emotional development. A seam-
to advance
less human-robot interaction has the potential to
AMLRPG@SRCȩ RMȩ RP?LQN?PCLRȩ ?LBȩ CVNJ?GL?@JCȩ '
ȩ ?LBȩ
the explainability,
for vulnerable populations too (e.g. children and the
elderly), because of its intuitive nature even with
accountability
minimalist design characteristics.
and transparency
$MPȩ RFGQȩ PC?QML
ȩ UFGJCȩ RFCȩ ȏCJBȩ MDȩ PM@MRGAQȩ GQȩ ?B-
T?LAGLE
ȩ PCQC?PAFCPQȩ GLȩ RFCȩ ȏCJBȩ MDȩ FSK?LPM@MRȩ of algorithms
interaction create scenarios in which they test
various robot characteristics in real-life settings.
In this way, they investigate the impact of so-
in general and
cially intelligent robots on human cognitive and
socio-emotional development by involving multiple
DL systems
stakeholders. This type of research allows society
to actively shape the ways in which robots might be
in particular.
integrated into our everyday life and contribute to a
more inclusive and sustainable use of AI.

Whilst these developments are important because

RFCWȩ @PGLEȩ 'ȩ GLRMȩ MSPȩ CTCPWB?Wȩ CVNCPGCLAC
ȩ RFCWȩ
PCNPCQCLRȩ UF?Rȩ GQȩ MDRCLȩ BCȏLCBȩ ?Qȩ RFCȩ /BSSPX
"*.
 23 2. About AI

Nowadays, many ML techniques display some char- and the algorithm development process which may
?ARCPGQRGAQȩMDȩ?ȩ~@J?AIȩ@MVȩKMBCJ
ȩGCȩUCȩILMUȩUF?Rȩ CLAMBCȩ@G?QCQȩ@WȩRFCȩBCTCJMNKCLRȩRC?KȩMPȩAMLRCVR
goes in and what comes out of the algorithm, but
we do not have a full understanding of its inner 1SKK?PW?LBAMLAJSQGMLQ
UMPIGLEQȩ 2FGQȩ JGKGRQȩ RFCȩ QAGCLRGȏAȩ SLBCPQR?LBGLEȩ
of algorithms, the capability to recover from ad- AI is a generic term that refers to any machine
TCPQ?PG?JȩCV?KNJCQ
ȩ?LBȩAMKNJGA?RCQȩFSK?LȩQSNCP- or algorithm that is capable of observing its en-
vision in practical applications (see Chapter 10 on vironment, learning and, based on the knowledge
cybersecurity). ?LBȩ CVNCPGCLACȩ E?GLCB
ȩ R?IGLEȩ GLRCJJGECLRȩ ?ARGMLȩ
MPȩ NPMNMQGLEȩ BCAGQGMLQȩ 2FCPCȩ ?PCȩ K?LWȩ BGȎCPCLRȩ
5CȩLCCBȩRMȩBCȏLCȩCT?JS?RGMLȩDP?KCUMPIQȩRF?Rȩ?PCȩ RCAFLMJMEGCQȩRF?RȩD?JJȩSLBCPȩRFGQȩ@PM?Bȩ'ȩBCȏLGRGMLȩ
meaningful and in realistic settings to match prac- At the moment, Machine Learning techniques are
RGA?Jȩ?NNJGA?RGMLȩAMLRCVRQȩ'LȩRFGQȩPCQNCAR
ȩUCȩQFMSJBȩ the most widely used.
consider engineering best practices, impact as-
sessment methods, user satisfaction and business The methodological developments in AI date back
metrics, in order to develop smart and transparent more than 50 years but the recent breakthrough has
@CLAFK?PIGLEȩQRP?RCEGCQȩ5CȩQFMSJBȩRP?GLȩRFCȩLCVRȩ come from advances in computing power, availability
generation of ML developers to apply and commu- MDȩB?R?
ȩ?LBȩAMKNJCVGRWȩMDȩRFCȩ?JEMPGRFKQȩ,MUȩRFCȩ
nicate these strategies and follow best practices many applications of AI are entering into our every-
for AI evaluation. B?Wȩ JGTCQ
ȩ DPMKȩ ȏJRCPGLEȩ QN?K
ȩ RMȩ K?AFGLCȩ RP?LQJ?-
tions, image recognition, and music generation, and
5CȩLCCBȩRMȩ?BT?LACȩRFCȩCVNJ?GL?@GJGRW
ȩ?AAMSLR?@GJGRW are increasingly used in industry, government and
and transparency of algorithms in general, and DL commerce. We are probably only at the beginning of
systems in particular, both from the perspective this process because the development of ubiquitous
MDȩ+*ȩPCQC?PAFȩ?LBȩDPMKȩRF?RȩMDȩSQCPQȩGLȩBGȎCPCLRȩ QCLQMPȩLCRUMPIQ
ȩ?LBȩRFCȩ'M2
ȩUGJJȩGLAPC?QCȩCVNMLCL-
?NNJGA?RGMLQȩ5CȩLCCBȩRMȩCVRCLBȩB?R?ȩ?LBȩ?JEMPGRF-
mic literacy across society to increase the ability of
critical thinking with respect to machine intelligence
and challenge it when necessary.
Further research is also needed to understand fair-
tially the sensing capabilities of AI, the volumes of
data available to train the algorithms, and the reach
of AI through decisions and actions.
The opportunities are many, and in some cases not
yet foreseen, but so are the challenges. These will be
LCQQȩGLȩRFCȩAMLRCVRȩMDȩ?SRMK?RCBȩBCAGQGMLK?IGLEȩ CVNJMPCBȩGLȩN?PRGASJ?PȩGLȩ.?PRȩȩMDȩRFCȩPCNMPR
ȩ@SRȩGLȩ
An algorithm or decision is fair when it does not this chapter we have highlighted that the algorithms
discriminate against people because of their mem- powering AI display all the characteristics of a black
@CPQFGNȩ RMȩ ?ȩ QNCAGȏAȩ EPMSNȩ QSAFȩ ?Qȩ ECLBCP
ȩ P?AC
ȩ @MVȩUCȩA?Lȩ?AACQQȩRFCȩGLNSRQȩ?LBȩMSRNSRQ
ȩ@SRȩBMȩ
QCVS?JȩMPGCLR?RGMLȩ'RȩGQȩSL@G?QCBȩGDȩGRȩGQȩLMRȩ?ȎCAR- not understand fully what happens in-between, and
ed by the membership of a person to a protected how certain outputs, including decisions and actions,
group. Attempts to formalise algorithmic discrimi- ?PCȩBCPGTCBȩ%GTCLȩRFGQȩK?HMPȩBP?U@?AI
ȩGRȩGQȩGKNCP?-
nation leads to a long list of fairness criteria that RGTCȩRMȩGLAPC?QCȩMSPȩQAGCLRGȏAȩSLBCPQR?LBGLEȩBCTCJMNȩ
are always harmful to one group of stakeholders transparent mechanisms to assess the quality and
?LBȩ @CLCȏAG?Jȩ RMȩ RFCȩ MRFCP
ȩ K?IGLEȩ GRȩ BGȑASJRȩ RMȩ performance of AI so that we are able as a society
arrive at an agreed set. Moreover, these criteria to build and retain trust in this disruptive technology.
are helpful to identify if there has been discrimi-
L?RGMLȩGLȩBCAGQGMLK?IGLE
ȩ@SRȩRFCWȩBMȩLMRȩCVNJ?GLȩ
the source of unfair treatment. This is best done by
looking at biases in the training data, its labelling,
 3. EU in the AI competitive global landscape 24

&VSPQF
IBT
FYDFMMFOU
SFTFBSDI
JO
"*
BOE
MFBEJOH
JOEVTUSJBM
TFDUPST
PO
XIJDI
UP
CVJME

SUMMARY

The overview of the global and European AI landscape we report in this chapter is the fruit of
?ȩAMKNPCFCLQGTCȩ?L?JWQGQȩMDȩMTCPȩȩȩICWȩNJ?WCPQȩUMPJBUGBCȩ2FCȩ?L?JWQGQȩAMLȏPKQȩRFCȩ
intense competition in AI taking place worldwide with three main leaders: the USA, Europe,
?LBȩ!FGL?ȩ#?AFȩPCEGMLȩF?Qȩ?@MSRȩMLCȩOS?PRCPȩMDȩ?JJȩICWȩNJ?WCPQȩGLȩRFCȩ'ȩȏCJB
ȩGLAJSBGLEȩ@MRFȩ
PCQC?PAFȩ?LBȩGLBSQRPW
ȩ@SRȩ?ȩBGQRGLARGTCȩBGȎCPCLRȩKGVȩMDȩNJ?WCPQȩUFGJCȩ#SPMNCȩGQȩUCJJȩ@?J?LACBȩ
GLȩRFCȩLSK@CPȩMDȩPCQC?PAFȩ?LBȩLMLPCQC?PAFȩNJ?WCPQ
ȩRFCȩ31ȩF?Qȩ?NNPMVGK?RCJWȩRFPCCȩRGKCQȩ
?QȩK?LWȩGLBSQRPG?J AMPNMP?RCȩNJ?WCPQȩ?QȩPCQC?PAFȩMLCQ
ȩ?LBȩ!FGL?ȩF?Qȩ?@MSRȩQGVȩRGKCȩ?QȩK?LWȩ
research players as industrial ones. The strength of the corporate world in the USA is also
indicated by its dominance in the number of start-ups (almost half of the total worldwide)
and venture capital (more than one third of the total). China, on the other hand, is making a
AMLACPRCBȩCȎMPRȩRMȩRSPLȩPCQC?PAFȩGLRMȩN?RCLRQ
ȩ?LBȩ?AAMSLRQȩDMPȩ?JKMQRȩϤȩMDȩRFCȩUMPJBȩRMR?J

Europe is currently in a good position regarding the quality of its research, with more than 30 %
MDȩRMNȩ'ȩNS@JGA?RGMLQ
ȩHSQRȩ@CFGLBȩRFCȩ31ȩȩϤȩ2FGQȩPCQSJRȩGQȩRFCȩDPSGRȩMDȩRFCȩFGEFȩJCTCJȩMDȩ
research funding and the intense inter-country collaboration provided by the framework pro-
grammes. Key areas of strength in Europe on which to build upon are automated and connected
vehicles, and robotics. Continued policy support seems key mantaining and improving Europe’s
position in an ecosystem which is rather polarised around a few big powers.
 25 3. EU in the AI competitive global landscape

EU IN THE AI
COMPETITIVE
GLOBAL
LANDSCAPE
 2FCGLRCPL?RGML?JNMJGAWAMLRCVR
In the context
Dutton (2018) ȩF?QȩGBCLRGȏCBȩ?RȩJC?QRȩȩAMSLRPGCQȩ
of strong global
4

that have launched or are preparing national AI

strategies and programme in 2017-2018, of which
8 are in Europe, 3 in North America, and the rest in competition on AI,
QG?ȩ?LBȩ-AC?LG?ȩ&CPCȩUCȩQSKK?PGQCȩRFCȩBCTCJMN-
ments in the USA, Canada, India, Japan, and South QEB#2FPTBIIMI>@BA
Korea, leaving Europe to Chapter 4, and China to
Chapter 5. if it builds on its
 USA strengths, like
'Lȩ -ARM@CPȩ 
ȩ RFCȩ -@?K?ȩ ?BKGLGQRP?RGMLȩ NS@- research and robotics,
lished three reports that concluded that AI technol-
ogy will be the driving force behind transformations and works together.
?APMQQȩ@MRFȩRFCȩCAMLMKWȩ?LBȩL?RGML?JȩQCASPGRWȩ-LCȩ
of the reports (NSTC, 2016c) proposed a national AI
R&D strategy linked to some other initiatives with
QGVȩK?GLȩBGPCARGMLQ
1. Make long-term investments in AI research.
2. "CTCJMNȩ CȎCARGTCȩ KCRFMBQȩ DMPȩ FSK?L'
collaboration.
3. Understand and address the ethical, legal,
and societal implications of AI.
4. Ensure the safety and security of AI systems.
5. Develop shared public datasets and environ-
ments for AI training and testing.
6. Measure and evaluate AI technologies
through standards and benchmarks.
The current administration seems to be supporting
the unrestricted development of AI, and in the May
2018 Summit on AI at the White House with lead-
ing industrialists Michael Kratsios, Deputy Assistant
 3. EU in the AI competitive global landscape 26

to the President for Technology Policy, announced Science Foundation, the main agency funding R&D
the government’s goals: (1) maintain American on AI together with DARPA and the Department for
leadership in AI, (2) support the American worker, 2P?LQNMPR?RGMLȩGLȩRFCȩQNCAGȏAȩ?PC?ȩMDȩ?SRMLMKMSQȩ
(3) promote public R&D; and (4) remove barriers and unmanned systems), amounts to US$ 5.3 bil-
RMȩGLLMT?RGMLȩ2Mȩ?AFGCTCȩRFCQCȩM@HCARGTCQ
ȩ)P?RQGMQȩ lion in the 2019 budget.10
?LLMSLACBȩ?ȩLCUȩ1CJCARȩ!MKKGRRCCȩMLȩPRGȏAG?Jȩ'L-
telligence to advise the White House on interagency 2FCȧNMUCPȧMDȧRFCȧ31ȧRCAFȧEG?LRQȧGQȧMDȧAMSPQCȧLMRȧ
AI R&D priorities and to consider the creation of Fed- HSQRȧȏL?LAG?J
ȧ@SRȧKMPCȧAPSAG?JJWȧ?@MSRȧRFCȧT?QRȧ
eral partnerships with industry and academia.5 OS?LRGRGCQȧ MDȧ B?R?ȧ RFCWȧ ?PCȧ ?@JCȧ RMȧ E?RFCPȧ ?LBȧ
In September 2018, DARPA, the government de-
fence research agency, announced an investment
plan of over US$ 2 billion to overcome some of the
perceived limitations of current AI technologies, i.e.
their dependence on large amounts of high-quali-
ty training data, poor ability to adapt to changing
conditions, limited performance guarantees, and
ȩRFCȩAMLRCVRȩDMPȩRFCȩB?R?
ȩGQȩ
GL?@GJGRWȩRMȩNPMTGBCȩSQCPQȩUGRFȩCVNJ?L?RGMLQȩMDȩRFCGPȩ
?L?JWQCȧDPMKȧ?JJȧMTCPȧRFCȧUMPJBȩ%MMEJCȩ?AAMSLRQȩ
for 86 % of all internet searches in the world,11
processing some 40 000 queries per second, or
3.5 billion per day.12 Facebook has some 2.2 billion
active users per month.13 It is thus the largest
content manager in the world, although it does not
produce any of its own data. All the data, and more
GKNMPR?LRJWȩDMPȩ'ȩRP?GLGLE
NPMTGBCBȩ@WȩGRQȩASQRMKCPQȩGLȩCVAF?LECȩDMPȩRFCȩQCPTGACȩ
PCQSJRQȩ2FCȩ~'ȩ,CVRȩNPMEP?KKCȩQCCIQȩRMȩCVNJMPCȩ provided by the platform. Its Like button has turned
new theories and applications that could make it the company into the largest recommendation
possible for machines to adapt to changing situa- NJ?RDMPKȩ GLȩ RFCȩ UMPJB
ȩ ?@JCȩ RMȩ NPMȏJCȩ ?LBȩ R?PECRȩ
RGMLQȩBCTCJMNGLEȩAMLRCVRS?JȩPC?QMLGLEȩA?N?@GJGRGCQ6
These are signals of continued support from the
Trump administration towards the development of
AI, which will also be boosted by an undisclosed
part of the large USA defence budget. Compared to
other regions of the world where government plays
?ȩK?HMPȩN?PRȩGLȩ@MRFȩNMJGAWȩ?LBȩGLTCQRKCLR
ȩGLȩRFCȩ
small groups and individuals for both marketing,
and as seen recently with the Cambridge Analytica
case, also for political purposes. As indicated in
Chapter 2, the development of AI rests largely on
the ability to have large volumes of well-structured
and labelled data, and in this respect the USA giants
have a clear lead in the West.
USA, private companies play a central role in the  -RFCPAMSLRPGCQ
development of technology in general as well as AI.
$BOBEB
Canada announced its AI strategy in the
The USA stock market was worth some US$ 34 tril- 2017 budget, which allocates CAN$ 125 million
lion in April 2018, 43 % of the total world markets, MTCPȩȏTCȩWC?PQȩRMȩ?AFGCTCȩDMSPȩK?GLȩM@HCARGTCQ
compared to Europe’s 15 %, and Asia and Japan
15 %.7 Tech companies account for one quarter of • To increase the number of outstanding AI
this value8
ȩ?LBȩRFCȩRMNȩȏTCȩNNJC
ȩK?XML
ȩ%MMEJC
researchers and skilled graduates in Canada.
Alphabet, Microsoft and Facebook) account for
some US$ 4 trillion9, or 11 % of the total USA mar- • To establish interconnected nodes of scien-
ket, and 33 % of the total valuation of the Euro- ȩ RGȏAȩCVACJJCLACȩGLȩ!?L?B?QȩRFPCCȩK?HMPȩACL
pean market. Not surprisingly, the R&D investment tres for AI in Edmonton, Montreal and Toronto.
of companies like Amazon and Alphabet is very
QGELGȏA?LR
ȩ?Rȩ31ȩȩ@GJJGMLȩ?LBȩ31ȩȩ@GJJGMLȩ • To develop global thought leadership on the
respectively in 2017 (Knight, 2018), much more economic, ethical, policy and legal implica-
than the civilian investment of the government. tions of advances in AI.
In comparison, the total budget for the National
 27 3. EU in the AI competitive global landscape

ɎɛɤǬ 1MKCCV?KNJCQMD'BCTCJMNKCLRQGLRFC31

Google is an AI world leader. It researches and develops AI duct

Amazon AI empowers fake reviews detection, chatbots, product
products and services, has developed its own AI chip, the Tensor recommendations, big data management, etc. Amazon on
ÄOL@BPPFKD2KFQ1Ä2
>KA@OB>QBA1BKPLO$ILT
LKBLCQEBJLPQ Web Services are probably the biggest cloud system in the
widely used open source AI/ML libraries. DL algorithms are world providing infrastructure as a service to over JFIIFLK
LK
already a considerable part of several Google products including users. 

recommendation systems, Android, Gmail, Maps, and translation

services. The number of DL directories Google uses has gone +F@OLPLȱ has also been investing heavily in AI and infrastructure
ture
from zero in  to  in .'K+>V
 as a service. +F@OLPLȱ WROB is one of the three biggest gest
Google announcedGoogle duplex, an AI system for cloud (and AI) providers together with Google Cloud and
>@@LJMIFPEFKD OB>ITLOIA Q>PHP LSBO QEB MELKB Amazon web PBOSF@BP !LOQ>K> FP QEB +F@OLPLȱ SFOQR>I QR>I
Google, and other large operators, also use AI extensively to assistant that competes with Amazon Alexa, Apple Siri, Googleogle
manage their data centres and save energy. Duplex and others.

$>@B?LLH is the biggest social media company in the world. It has Apple is the most valuable company in the world worth over
BPQ>?IFPEBA>$>@B?LLH'OBPB>O@E$'/QB>JTEF@EFPLKBLC 20”QOFIIFLK. Apple’s main AI divisions are Siri team, and
QEBJLPQ>AS>K@BA
TFQEPBSBO>II>?P>OLRKAQEBTLOIA$>@B?LLH the Core ML team. Siri’s team focus is on NLP and computer uter
supports the open source AI libraries PyTorch and !>ȬB vision, both of which are necessary to power voice assistant tant
that compete with %LLDIB1BKPLO$ILT. The company has CB>QROBP >KA KBT
 JLOB @RQQFKDBADB QB@EKLILDV PR@E >P
also released plans to develop its own AI chips. augmented reality apps that rely on object recognition. Core MLL is
the ML Application Programming Interface (API) PI)
Amazon uses AI extensively in its recommendation systems and QE>QMMIBI>RK@EBAI>PQVB>OQLEBIM'Q>PHP>KA'CL@RPBA>MMP
MMP
ILDFPQF@P'Q>IPLLȬBOP?LQE@LKPRJBO>KA?RPFKBPPLOFBKQBA' >KAPBOSF@BPCOLJQEFOAM>OQVABSBILMBOPORKJLOBBȯ@FBKQIVLK
VLK
products and services. Amazon Echo brings AI into the Apple devices.
ELJB QEOLRDE QEB FKQBIIFDBKQ SLF@B PBOSBO
 IBU> $LO ?RPFKBPP
• To support a national research community
on AI.
The strategy is led by the Canadian Institute for
Advanced Research17 in close partnership with
the Canadian government and the three new AI
Institutes: the Alberta Machine Intelligence Insti-
tute in Edmonton, the Vector Institute in Toronto,
and MILA in Montreal. Compared to other nation-
al strategies that have a strong emphasis also on
sectoral developments, data, and supporting in-
dustry, the Canadian strategy is primarily focused
on research.
*OEJB
A discussion paper for an AI strategy was pub-
lished in June 2018 by NITI Aayog (a government
think-tank) focusing very much on social inclusive-
ness: #AIforAll.
The strategy follows three distinct, yet inter-related
components:
a) -NNMPRSLGRWȩRFCȩCAMLMKGAȩGKN?ARȩMDȩ'ȩDMPȩ
India.
b) 'ȩDMPȩ%PC?RCPȩ%MMBȩQMAG?JȩBCTCJMNKCLRȩ?LB
inclusive growth.
c) 'ȩ %?P?ECȩ DMPȩ ȩ Ϥȩ MDȩ RFCȩ UMPJBȩ QMJSRGML
provider of choice for the emerging and
developing economies across the globe.
 3. EU in the AI competitive global landscape 28

In relation to c) the strategy argues that India pro-
vides a perfect ‘playground’ for enterprises and
institutions globally to develop scalable solutions
ȩGRȩBMCQȩLMRȩNSRȩ?ȩȏESPCȩMLȩ
which can be easily implemented in the rest of the
developing and emerging economies. NITI Aayog
DMASQCQȩMLȩȏTCȩQCARMPQȩRF?Rȩ?PCȩCLTGQGMLCBȩRMȩ@CLC-
ȏRȩRFCȩKMQRȩDPMKȩ'ȩGLȩQMJTGLEȩQMAGCR?JȩLCCBQ
ȩ?LBȩ LMSLACBȩ ?ȩ ȏTCWC?Pȩ NJ?L
ȩ NJ?AGLEȩ 'ȩ ?LBȩ PM@MRGAQȩ
where only private-sector-led initiatives may not
lead to the desired societal outcomes:
a) Healthcareȩ GLAPC?QCBȩ ?AACQQȩ ?LBȩ ?ȎMPB? the USA and Chinese models. It is based on the 5th
bility of quality healthcare,
Although the discussion paper acknowledges that
QGELGȏA?LRȩ NS@JGAȩ GLTCQRKCLRȩ UMSJBȩ @Cȩ LCCBCBȩ RMȩ
GKNJCKCLRȩRFCȩQRP?RCEW
it, leaving the matter to the political discussion.
+BQBO In 2015, the Japanese government an-
as the stepping stone of its renewed strategy for
science, technology and innovation. Japan’s ap-
NPM?AFȩRMȩ?PRGȏAG?JȩGLRCJJGECLACȩGQȩ?ȩAMK@GL?RGMLȩMDȩ
1AGCLACȩ ?LBȩ 2CAFLMJMEWȩ ?QGAȩ .J?Lȩ zȩ
of the Japanese Council for Science, Technology
b) Agriculture: enhanced farmers’ income, in-
creased farm productivity and reduction of
wastage,
c) Education: improved access and quality
of education,
d) Smart Cities and Infrastructure: improved
connectivity for the rapidly increasing urban
ȩ
population, and
e) Smart Mobility and Transportation: smarter
and safer modes of transportation and bet-
?LBȩ'LLMT?RGMLȩ2FCȩNJ?LQȩM@HCARGTCȩGQȩRMȩCQR?@JGQFȩ
Japan as a super smart society (i.e. Society 5.0). The
programme is headed by the prime minister, and
GRQȩȏCJBQNCAGȏAȩNPMEP?KKCQȩ?PCȩFC?BCBȩ@WȩQCJCAR-
ed corporate heads. In April 2016, the government
established a Strategic Council for AI Technology
promoting research and development of AI tech-
nology, coordinating with industries related to the
GLBSQRPGCQȩRF?RȩSRGJGQCȩ'ȩQMA?JJCBȩ~CVGRȩGLBSQRPGCQ
and moving forward with the social implementa-
tion of AI technology. The Council coordinates the
activities of three leading research centres:
ȩ RCPȩRP?ȑAȩ?LBȩAMLECQRGMLȩK?L?ECKCLR 1) The Centre for Information and Neural
The strategy makes some 30 recommendations.
With respect to research, it proposes a two-tier
QWQRCKȩ LCUȩ !CLRPCQȩ MDȩ 0CQC?PAFȩ #VACJJCLACȩ GLȩ
AI will focus on fundamental research and act as
technology feeders for the International Centres
for Transformational AI, which will focus on cre-
ating AI-based applications in domains of societal
importance.
To facilitate adoption of AI, the strategy recom-
mends in particular:
1) Establishing a market place for AI, and
Networks and Universal Communication
Research Institute of the National Insti-
tute of Information and Communications
Technology.
2) The RIKEN Centre for Advanced Intelligence
ȩȩ ȩ .PMHCARȩ MDȩ RFCȩ 'LQRGRSRCȩ MDȩ .FWQGA?Jȩ ?LB
Chemical Research.
3) 2FCȩPRGȏAG?Jȩ'LRCJJGECLACȩ0CQC?PAFȩ!CLRPCȩMD
the National Institute of Advanced Industrial
Science and Technology.
The AI Technology Strategy was published in March
2) Facilitating the creation of large found-
ational annotated data sets.
ȩ'RȩGBCLRGȏCQȩDMSPȩNPGMPGRWȩ?PC?QȩDMPȩRFCȩBCTCJ-
opment and integration of AI: increased productiv-
ity through user-driven hyper customisation of ser-
vices, medical healthcare and welfare to support an
increasingly ageing population, mobility to support
 29 3. EU in the AI competitive global landscape

safe and environmentally friendly travel for all, and 2FCȩRFPCCȩACLRPCQȩGBCLRGȏCBȩ?@MTCȩUGJJȩ?ARȩ?QȩJC?BȩDMPȩ

information security. What is particularly interest- C?AFȩMDȩRFCȩNPGMPGRWȩ?PC?QȩGBCLRGȏCB
ȩ?LBȩ?ARȩ?QȩFS@Qȩ
ing in this strategy is that each of these four prior- for R&D in partnership with industry and academia.
ities is driven by an image of the society Japan is
?GKGLEȩDMPȩȩAJC?PȩCV?KNJCȩMDȩQMAG?JJWȩBPGTCLȩ'ȩ2FCȩ
strategy is articulated in three phases:
1) Utilisation and application of data-driven
AI developed in various domains (until 2020),
2) Public use of AI and data developed across
various domains (2020-2025), and
3) creation of an ecosystem built by connecting
the multiplying domains (2025-2030).
The strategy pays particular attention to the col-
lection, management and sharing of data across
government, academia and the private sector. With
respect to the latter, it recognises that there are
still many issues to be solved, including the han-
dling of personal data, but has set up a Data Dis-
RPG@SRGMLȩAACJCP?RGMLȩ5MPIGLEȩ%PMSNȩRMȩFCJNȩQF?PCȩ
best practice.
4PVUI
,PSFB
In 2016, the government introduced
a mid-to-long-term Master Plan for the Intelligent
'LDMPK?RGMLȩ 1MAGCRWȩ -LCȩ MDȩ RFCȩ NJ?LQȩ EM?JQȩ GQȩ RMȩ
foster the ‘establishment of the basis for AI tech-
nology’. It announced it would spend 1 trillion won
31ȩ ȩ KGJJGMLȩ @Wȩ ȩ RMȩ @MMQRȩ RFCȩ ?PRGȏAG?Jȩ

Agriculture, intelligence industry. The government plans to sup-

NMPRȩ RFCȩ CQR?@JGQFKCLRȩ MDȩ ?ȩ FGEFNPMȏJCȩ PCQC?PAFȩ

healthcare, centre that will serve as a ‘pivot’ of the nation’s

PCQC?PAFȩ?LBȩBCTCJMNKCLRȩGLȩRFCȩ'ȩȏCJBȩ1?KQSLEȩ
?LBȩ *%ȩ #JCARPMLGAQ
ȩ 1)2
ȩ )2
ȩ ,?TCP
ȩ ?LBȩ &WSLB?Gȩ
mobility +MRMPȩF?TCȩBCAGBCBȩRMȩHMGLȩRFCȩGLGRG?RGTCȩ#?AFȩAMK-
pany plans to invest about 3 billion won (Marquart,
and security 2016). In May 2018, the government added another
2.2 trillion won to its AI strategy focusing on three

are among priorities: training an additional 5 000 AI scientists,

developing applications in the areas of national de-

the most promising fence, medicine, and public safety and funding in-
frastructure and incubators for start-ups.

applications 
 L?JWQGLERFCICWDC?RSPCQ
MDRFC'J?LBQA?NC
of AI for many 2FCȩ ȏCJBȩ MDȩ 'ȩ GQȩ CVNCPGCLAGLEȩ ?ȩ NCPGMBȩ MDȩ GLRCLQCȩ
countries. progress, due to the consolidation of several key
technological enablers: faster processing, vastly in-
creased amounts of data, and better algorithms.
CA?SQCȩ'ȩGQȩCVN?LBGLEȩQMȩOSGAIJWȩ?APMQQȩKSJRGNJCȩ
sectors, and there has yet to be a single and agreed
BCȏLGRGMLȩ GLȩ UFGAFȩ RCAFLMJMEGCQȩ ?LBȩ ?NNJGA?RGMLQȩ
are included as ‘AI’, it is not possible to track its
CTMJSRGMLȩRFPMSEFȩMȑAG?JȩQR?RGQRGAQȩ?RȩRFGQȩQR?EC
ȩ?Qȩ
UMSJBȩ@CȩBMLCȩDMPȩKMPCCQR?@JGQFCBȩȏCJBQȩ
 3. EU in the AI competitive global landscape 30

Ɏɛɤǭ 2FC2CAFLM#AMLMKGA1CEKCLR2#1?NNPM?AFGL?LSRQFCJJ

Ɏɛɤ

1EBȭ
1EBȭOPQPQBMFKQEBJBQELALILDVFPQLABȭKBQEB?LRKA>OFBPLCQEB HBVTLOAP QE>Q EBIMBA QL FABKQFCV QEB EFDEBPQ KRJ?BO LC 'OBI>QBA
TES (AI in this case), by detecting players that focus on AI as their MI>VBOP>KA/">@QFSFQFBPFPPELTKFK$FDROB-K@BTBABȭKBAQEB
MOFJ>OV LO PB@LKA>OV >@QFSFQV ÄI>VBOP >OB ABȭKBA >P research
MOFJ> boundaries of the TES and mapped the players, we examined their
centres,
cent academic institutions, and companies which have FKQBO@LKKB@QFLKP QL ABP@OF?B QEB B@LPVPQBJ 4B @LKPFABOBA AFȬBOBKQ
participated in one or more of the following economic activities: R&D
partic QVMBP LC OBI>QFLKP PR@E >P @LM>OQF@FM>QFLK FK /" >@QFSFQFBP
 @L
MOL@BPPBP
FKARPQOF>IMOLAR@QFLK>KAJ>OHBQFKD
PMB@Fȭ@
MOL@ location, similarity in technology, etc. We used multilayer network
'OBI>QBAPBOSF@BP
'OB . The players´ activities may be explicitly stated, analysis to understand the relative importance of the players in the
e.g. in the description of the companies’ activities in business registers ecosystem and their positional advantages/disadvantages. We also
LOPBDJBKQPMB@Fȭ@ȭOJP}OBMLPFQLOFBP
LOABOFSBACOLJQEB>K>IVPFPLC
LOPB JLABIIBAQEBFKCLOJ>QFLKȮLTPQEOLRDELRQQEBKBQTLOH
>KAPQRAFBA
their R&D activities, e.g. text from patents, conference proceedings the network´s resilience over time. By clustering players by type of
and research projects. To carry out this analysis, we developed a technology, geographical location, and linkages to other players we
@LJMOBEBKPFSB
@LJM AF@QFLK>OV LC ALJ>FK PMB@Fȭ@ HBVTLOAP @LSBOFKD QEB >OB>IPL>?IBQLFABKQFCV@LRKQOVLOOBDFLK>IMOLȭIBPLCQB@EKLILDF@>I
technological
techn aspect of the TES (in this case AI). A selection of the specialisation and study their evolution over time.

2FCȩ(0!ȩF?QȩBCTCJMNCBȩ?ȩQNCAGȏAȩKCRFMBMJMEWȩRMȩ?Q- and how innovation spreads through the network.

sess emerging techno-economic segments (TES) like
AI, providing a synthetic overview of their ecosystem
and dynamics. The main aim is to understand the en-
ȩ @SQGLCQQȩ PCEGQRCPQ
ȩ QAGCLRGȏAȩ ?LBȩ RP?BCȩ NS@-
tire ecosystem as it emerges: who are the key play-
ers, in which technologies they specialise, how they
are distributed, how they are connected to each other,
The methodology (De Prato et al., 2018) makes
SQCȩMDȩBGȎCPCLRȩRWNCQȩMDȩD?ARS?JȩB?R?ȩQSAFȩ?QȩN?R-
CLRQ
lications, and industrial associations records.18 To
make sense of this heterogeneous data it applies
both statistical and ML techniques to detect early

Figure 2. Percentage of AI players detected by the most representative keywords (2000-18)

Note: The conceptual proximity of some of the keywords, and the overlap between techniques and applications, may result in the same
player been detected more than once.
 31 3. EU in the AI competitive global landscape

developments, map them, monitor their evolution, HSQRȩ MTCPȩ ȩ Ϥȩ UMPJBUGBC

ȩ %CPK?LWȩ ?@MSRȩ ȩ Ϥȩ MDȩ
and benchmark the ecosystem with the necessary the EU’s AI players and 3.7 % worldwide, and France
QAGCLRGȏAȩCTGBCLACȩ about 11 % of the EU’s and 2.7 % worldwide.

%GTCLȩ RFCȩ TMJSKCȩ ?LBȩ T?PGCRWȩ MDȩ QMSPACQȩ ?L?JWQCB

ȩ Consideration of the number of players relative to
and the amount of painstaking work needed to match RFCGPȩAMSLRPWQȩ%".ȩNPMTGBCQȩ?BBGRGML?JȩT?JS?@JCȩGL-
BGȎCPCLRȩ B?R?ȩ QRPSARSPCQȩ ?LBȩ GBCLRGDWȩ AMLLCARGMLQȩ formation about countries with a smaller economy
through the data sources, we believe this is the most
accurate and comprehensive database of the AI eco-
system to date. In the following section, we provide
a global overview, and highlight some characteristics
of the EU with respect to this global landscape.

ȩ RFCȩ GLBGA?RMPȩ PCJ?RGTCȩ RMȩ %".ȩ QA?JCQȩ SNȩ RMȩ
@SRȩ ?ȩ ȐMSPGQFGLEȩ 'ȩ GLBSQRPWȩ $GESPCȩ ȩ QFMUQȩ RF?Rȩ
'QP?CJȩ ?LBȩ 1GLE?NMPCȩ ?PCȩ K?IGLEȩ ?ȩ FSECȩ CȎMPRȩ RMȩ
keep up in the race for AI; while they already have
a non-negligible 2.1 % and 1.3 % of total world AI
NJ?WCPQ
2.45 and 1.63 players per billion euro respectively
 -TCPTGCU ?LBȩNMQGRGMLQȩRFCKȩ?QȩȏPQRȩ?LBȩQCAMLBȩGLȩRFCȩP?LI-
ing. The EU presents 0.59 AI players per billion euro,
!MLQGBCPGLEȧRFCȧ'ȧJ?LBQA?NCȧUMPJBUGBC
ȧRFCȧ2#1ȧ slightly higher than the USA (0.57) and China (0.43).
?L?JWQGQȧF?QȧBCRCARCBȧ?PMSLBȧȧȧNJ?WCPQȧGLȧ
RFCȧNCPGMBȧ, of which 16 000 were in- In the EU we can see a similar picture (Fig. 5).
volved in at least one research or innovation activ- 2FCȩ 3LGRCBȩ )GLEBMK
ȩ %CPK?LW
ȩ $P?LAC
ȩ 1N?GL
ȩ 'R-
ity, and 19 000 are players engaged solely in indus- aly, and the Netherlands have the largest number
trial activities. MDȩ NJ?WCPQȩ @SR
ȩ UFCLȩ UCGEFRCBȩ @Wȩ %".
ȩ SJE?PG?
ȩ
Estonia, Cyprus and Malta stand out, although when
The TES analysis shows (Fig. 3) that RFCȧ#3ȧGQȧ?KMLEȧ numbers are small, ratios have be treated with
RFCȧECMEP?NFGA?Jȧ?PC?QȧUGRFȧRFCȧFGEFCQRȧLSK@CPȧMDȧ some caution.
NJ?WCPQȧ?ARGTCȧGLȧ'
ȧHSQRȧ@CFGLBȧRFCȧ3LGRCBȧ1R?RCQȧ
?LBȧHSQRȧ?FC?BȧMDȧ!FGL?. The USA had about 28 % $GESPCȩȩGJJSQRP?RCQȩAJC?PJWȩRFCȩBGȎCPCLRȩKGVȩMDȩ0"ȩ
of the players, with the EU (25 %), and China (23 %) ?LBȩGLBSQRPG?JȩNJ?WCPQȩGLȩRFCȩBGȎCPCLRȩPCEGMLQȩ?LBȩ
following closely behind. All together these three re-
ȩRFCȩ@?PQȩ
gions account therefore for more than three quar-
ters of all AI players worldwide. Within the EU (Fig. 5)
ȩ
AMSLRPGCQȩ-LȩRFCȩPGEFRȩQGBCȩMDȩRFCȩEP?NF
indicate the number of players engaged in at least
MLCȩ 0"ȩ ?ARGTGRWȩ GLȩ RFCȩ NCPGMBȩ ȩ ȏPKQ
the largest economies have the most AI players, with PCQC?PAFȩGLQRGRSRGMLQȩ?LBȩ?A?BCKG?ȩ-LȩRFCȩJCDR
ȩUCȩ
the UK representing 25 % of the EU AI players and have the non-R&D, i.e. industrial players only.

Figure 3. AI players in the world by geographical zones (% over world total), 2009-2018
 3. EU in the AI competitive global landscape 32

Figure 4. Top 10 AI players in the world and relation with GDP by geographical zones, 2009-2018

As shown, while Europe is well balanced in the tries publishing strongly in their own language, Fig-
number of research and non-research players (ap- SPCȩȩQFMUQȩAJC?PJWȩRFCȩ!FGLCQCȩCȎMPRȩRMȩCLRCPȩRFCȩ
NPMVGK?RCJWȩ ϤȩC?AFȩMDȩRFCȩEJM@?JȩLSK@CP
ȩRFCȩ R&D worldwide network and possibly come to dom-
31ȩF?Qȩ?NNPMVGK?RCJWȩRFPCCȩRGKCQȩ?QȩK?LWȩGLBSQ- inate in the future. This overwhelming ratio of R&D
trial/corporate players as research ones (41 % of all versus non-R&D players is also observed in South
industrial players in USA vs. 13 % R&D), and China )MPC?ȩ ?LB
ȩ RMȩ ?ȩ JCQQCPȩ CVRCLR
ȩ GLȩ 0SQQG?ȩ 2FCȩ FGEFȩ
F?Qȩ ?@MSRȩ QGVȩ RGKCQȩ ?Qȩ K?LWȩ PCQC?PAFȩ NJ?WCPQȩ ?Qȩ
industrial ones (7 % industrial players and 42 %
R&D of the world total in each category). While we
ȩ UFGJCȩ RFCȩ @?J?LACȩ GLȩ RFCȩ #3ȩ PCȐCARQȩ
cannot forget that the TES analysis relies, so far,
proportion of non-R&D players in the USA, India,
'QP?CJȩ ?LBȩ !?L?B?ȩ PCȐCARQȩ RFCGPȩ TG@P?LRȩ GLBSQRPG?Jȩ
CAMQWQRCKQ
a very strong research and academic environment
only on R&D materials published in English and UGRFȩNMRCLRG?JȩRMȩLSPRSPCȩ?ȩȐMSPGQFGLEȩ'ȩGLBSQRPW
thus is possibly penalising China, and other coun- Figure 7 analyses the players that deserve special

Figure 5. AI players in the EU and relation with GDP by country, 2009-2018

 33 3. EU in the AI competitive global landscape

Figure 6. R&D versus non-R&D players in top 10 countries by number of AI players, 2009-2018

Figure 7. Number of AI players in selected types of activity (% over world total), 2009-2018
 3. EU in the AI competitive global landscape 34

?RRCLRGML
ȩJGICȩGLLMT?RGTCȩNJ?WCPQȩ?LBȩȏPKQȩAMLQGB-
ered to be start-ups or receiving venture capital.
The highly developed industrial ecosystem in the
USA puts this country at the forefront of AI-related
?ARGTGRGCQȩUGRFȩ ϤȩMDȩ?JJȩȏPKQȩPCACGTGLEȩTCLRSPCȩ
A?NGR?J
or research institutions. The importance of the EU
GLȩRFGQȩȏCJBȩQRCKQȩDPMKȩJGTCJWȩ'ȩQAGCLRGȏAȩPCQC?PAFȩ
activity supported by the EU framework pro-
grammes for R&D, which have fostered an intense
inter-country collaboration and high participation
ȩ ?LBȩ  Ϥȩ MDȩ QR?PRSNQȩ ?ARGTCȩ GLȩ RFCȩ ȏCJBȩ from most EU countries. Figure 8 shows the dis-
Both TCLRSPCȧA?NGR?Jȧ?LBȧQR?PRSNQȧ?PCȧAMLACL- tribution of players by country which participated
RP?RCBȧGLȧ?ȧDCUȧAMSLRPGCQ
ȧUGRFȧRFCȧ#3ȧDC?RSPGLEȧ GLȩ'PCJ?RCBȩPCQC?PAFȩNPMHCARQȩSLBCPȩRFCȩ$.ȩ?LBȩ
QRPMLEJWȧ UGRFȧ ȧ ϡȧ MDȧ QR?PRSNQȧ ?LBȧ ȧ ϡȧ MDȧ H2020 European research programmes in the peri-
TCLRSPCȧA?NGR?Jȧ od 2009-2018 with a good spread among the larg-
CPȩ ?LBȩ KMPCȩ PCQC?PAF?ARGTCȩ AMSLRPGCQ
ȩ %CPK?LW
ȩ
It is worth noting the number of start-ups in India UK, Italy, Spain, France and the Netherlands.
(9 %) and Canada (5 %), while the relatively low
LSK@CPȩGLȩ!FGL?ȩPCȐCARQȩ?ȩBGȎCPCLRȩDSLBGLEȩKCAF-
anism for new companies in that country.
China has a relatively strong position in venture
capital (14 %), behind the USA and the EU. Moreover,
in the past few years, China has conquered the top
position in patenting, with 57 % of patenting appli-
cants in AI worldwide coming from that country.
Behind the USA, RFCȧ#3ȧF?Qȧ?ȧNPMKGLCLRȧNMQGRGMLȧ
GLȧ RFCȧ PCQC?PAFȧ QACLCȧ DPMLRGCPȧ PCQC?PAF
ȧ UGRFȧ networks), and applications such as face recognition,
ȧϡȧMDȧN?NCPQȧQS@KGRRCBȧRMȧRFCȧRMNȧ'ȧGLRCPL?-
RGML?JȧAMLDCPCLACQ are coming from EU companies
 2CAFLMJMEGA?JA?N?AGRW
Research, both basic and applied, is the backbone
of the technological capacity needed to bring the
European industry to a leading position in the inter-
national landscape. In this section, we look at a few
technological sub-domains of the AI landscape to
analyse the EU’s key strong areas with respect to
other world regions. The largest number of AI play-
ers address AI domains related to ML (e.g. neural
speech recognition, computer vision or those encom-
passed by connected vehicles.

Figure 8. Distribution of EU players (%) in EU-funded AI research projects, 2009-2018

 35 3. EU in the AI competitive global landscape

1SKK?PW?LBAMLAJSQGMLQ
The overview of the global and European AI land-
scape we report in this chapter is the fruit of a
comprehensive analysis of over 35 000 key players
ȩ GLAJSBGLEȩ @MRFȩ PCQC?PAFȩ ?LBȩ GLBSQRPW
ȩ
ȩ RFCȩ 31ȩ F?Qȩ ?NNPMVG-
ȩ?LBȩ!FGL?ȩF?Qȩ?@MSRȩQGVȩ
ȩMLȩRFCȩMRFCPȩF?LB
ȩGQȩK?IGLEȩ?ȩQRPMLEȩCȎMPRȩ
Figure 9. Occurrence of AI topics (%) by geographical zone,
2000-2018
$PMKȩRFCȩ2#1ȩ?L?JWQGQ
ȩUCȩGBCLRGȏCBȩDMSPȩICWȩRMN-
UMPJBUGBCȩ2FCȩ?L?JWQGQȩAMLȏPKQȩRFCȩGLRCLQCȩAMK-
petition in AI taking place worldwide with three
main leaders: the USA, Europe, and China. Each
region has about one quarter of all key players in
RFCȩ 'ȩ ȏCJB
@SRȩ?ȩBGQRGLARGTCȩBGȎCPCLRȩKGVȩMDȩNJ?WCPQȩUFGJCȩ#S-
rope is well balanced in the number of research
?LBȩ LMLPCQC?PAFȩ NJ?WCPQ
mately three times as many industrial/corporate
NJ?WCPQȩ?QȩPCQC?PAFȩMLCQ
times as many research players as industrial ones.
The strength of the corporate world in the USA is
also indicated by its dominance in the number of
start-ups (almost half of the total worldwide) and
venture capital (more than one third of the total).
!FGL?
to turn research into patterns, and accounts for al-
most 60 % of the world total. China has also put
in place a strongly coordinated approach to AI, in-
cluding government policy, industrial applications
?LBȩ PCQC?PAFȩ UGRFȩ RFCȩ M@HCARGTCȩ MDȩ @CAMKGLEȩ RFCȩ
world leader in AI by 2030. This is an ambitious but
achievable target (Chapter 5).
Europe is currently in a good position in the quality
of its research, with more than 30 % of top AI pub-
JGA?RGMLQ
ȩHSQRȩ@CFGLBȩRFCȩ31ȩȩϤȩ2FGQȩPCQSJRȩGQȩ
ics in the AI technological landscape for 2000- the fruit of the good level of research funding and
2018: ML methods; connected and automated the intense inter-country collaboration provided by
vehicles; speech recognition and natural language the framework programmes. Key areas of strength
processing; and face recognition. Figure 9 shows in Europe on which to build upon are automated
the key areas of attention/specialisation of each and connected vehicles, and robotics. Continued
geographical area. As indicated, the USA and Ja- policy support seems key to maintain and improve
pan have focused on the area of speech recogni- the European position in an ecosystem rather po-
tion and natural language; ML has captured great larised around a few big powers.
GLRCPCQRȩGLȩRFCȩ#3
ȩRFCȩ31
ȩ?LBȩRMȩ?ȩJCQQCPȩCVRCLRȩ
in China, while face recognition is in the spotlight
in China and Japan. The EU is the region present-
ing a more balanced coverage of all four topics.
 4. AI in the EU 36

&VSPQF
BU
UIF
DSPTTSPBET
PG
TFDVSF
BOE
FUIJDBM
"*

SUMMARY

This chapter shows that K?LWȧ #SPMNC?Lȧ AMSLRPGCQȧ ?Qȧ UCJJȧ ?Qȧ RFCȧ #!ȧ ?PCȧ BCTCJMNGLEȧ
QRP?RCEGCQȧ?LBȧNPMEP?KKCQȧRMȧESGBCȧRFCȧBCTCJMNKCLRȧMDȧ', with shared concerns over the
LCCBȩDMPȩ?Lȩ?EPCCBȩCRFGA?JȩDP?KCUMPIȩ?LBȩ?NNJGA?RGMLQȩRF?RȩAJC?PJWȩ@CLCȏRȩ#SPMNC?LȩQMAGCRWȩ
?LBȩSNFMJBȩRFCȩ#SPMNC?LȩT?JSCQȩCLQFPGLCBȩGLȩRFCȩ2PC?RGCQȩ2FCȩ&GEF*CTCJȩ#VNCPRȩ%PMSNȩCQR?-
blished in 2018 by the European Commission is developing ethical and investment guidelines
that will provide a framework for subsequent developments.

+MQRȧL?RGML?JȧQRP?RCEGCQȧ?L?JWQCBȧQF?PCȧQGKGJ?Pȧ?RRCLRGMLȧRMȧQRPCLERFCLGLEȧRFCGPȧPCQC?PAFȧ
base, including the setting up of one or more national centres for AI, QSNNMPRȧDMPȧRFCGPȧGLBSQRPW
and SMEs and awareness of the need to share data better between all the stakeholders: the
public sector, industry, and the public. They also focus on applications aimed at KMBCPLGQGLEȧ
NS@JGAȧ?BKGLGQRP?RGMLQ
ȩ?QȩUCJJȩ?QȩQNCAGȏAȩQCARMPQȩQSAFȩ?QȩFC?JRF
 37 4. AI in the EU

AI IN THE EU
1RP?RCEGCQ?LBNJ?LQ Many European
2FC#SPMNC?L3LGML
Many activities related to AI are currently taking
place In the EU. They include national strategies
and programme, and coordination activities among
the Member States and the European Commission.
Many of the enablers for the development of AI
F?TCȩ@CCLȩ?BBPCQQCBȩBSPGLEȩRFCȩJ?QRȩȏTCȩWC?PQȩSL-
der the Digital Single Market strategy. This is aimed
at ensuring access to online activities for individ-
uals and businesses under conditions of fair com-
petition, consumer and data protection, removing
geo-blocking and copyright issues.
In April 2018, the EU Member States signed a
%FDMBSBUJPO
PG
DPPQFSBUJPO
PO
"SUJҨDJBM
*OUFMMJHFODF19
in which they agreed to work together on the most
important issues raised by AI, from ensuring Europe’s
competitiveness in the research and deployment
of AI, to dealing with social, economic, ethical and
legal questions.
During the same month, the EC issued a Communi-
cation on AI for Europe (EC, 2018a) with three main
M@HCARGTCQȩ
• Boosting the EU’s technological and industrial
capacity and AI uptake across the economy,
both by the private and public sectors.
This includes investments in research and
innovation and better access to data.
countries
as well as the EC
are developing strategies
and programmes to guide
the development of AI
PLQE>QFQ?BKBȭQP
all Europeans.
market transitions and adaptation of social
protection systems.
• Ensuring an appropriate ethical and legal
framework, based on the Union’s values
and in line with the Charter of Fundamental
Rights of the EU. This includes forthcoming
ȩ ESGB?LACȩ MLȩ CVGQRGLEȩ NPMBSARȩ JG?@GJGRWȩ PSJCQ
a detailed analysis of emerging challenges,
and cooperation with stakeholders through
a European AI Alliance to develop AI ethics
guidelines.
The Communication earmarked EUR 1.5 billion to
support AI research for 2018-2020, having already
invested some EUR 2.6 billion in AI-related research
in the Horizon 2020 Framework Programme (EC,
@ȩ&MUCTCP
ȩRFCȩ?GKȩGQȩRMȩHMGLȩDMPACQȩ@CRUCCLȩ
• Preparing for socio-economic changes European, national, and private-sector invest-
brought about by AI by encouraging the ments to reach some EUR 20 billion per year for
modernisation of education and training sy- RFCȩLCVRȩBCA?BCȩQȩN?PRȩMDȩRFGQȩN?PRLCPQFGNȩCȎMPR
ȩ
stems, nurturing talent, anticipating changes the EC established the European AI Alliance as a
in the labour market, supporting labour multi-stakeholder forum for engaging in a broad

and open discussion of all aspects of AI develop-
ACPRGȏA?RGMLȩQWQRCKQȩDMP
ment and its impact on the economy and society.
'RȩGQȩQRCCPCBȩ@WȩRFCȩ&GEF*CTCJȩ#VNCPRȩ%PMSNȩMLȩ'ȩ
'ȩ&*#%
ȩUFGAFȩAMLQGQRQȩMDȩȩCVNCPRQȩUFMȩF?TCȩ
been selected by the Commission for this task.
2FCȩ'ȩ&*#%ȩGQȩLMUȩDMASQGLEȩMLȩRUMȩK?GLȩPCNMPRQȩ
with input from the members of the European AI
Alliance:
(1) Draft AI ethics guidelines, which will
ȩ MȎCPȩESGB?LACȩMLȩFMUȩRMȩGKNJCKCLRȩCRFGA?Jȩ
principles when developing and deploying
AI, building on the work of the European
ȩ %PMSNȩ MLȩ #RFGAQȩ GLȩ 1AGCLACȩ ?LBȩ ,CUȩ
Technologies and the European Union
Agency for Fundamental Rights; and
(2) Mid- and long-term policy recommendations
4. AI in the EU 38
ȩ B?R?ȩ?LBȩT?JGB?RGML
algorithms.
• Access to know-how support and shared
testing facilities such as smart hospitals or
precision-farming solutions.
• Networks of Digital Innovation Hubs to
facilitate the uptake of AI.
• Support to start-ups.
• Skilling and upskilling of the workforce and
action to attract and retain talent in Europe.
• Improved public services through numerous
AI-enabled applications, such as citizen-
ȩ EMTCPLKCLRȩ GLRCP?ARGMLȩ GLȩ RFCȩ AMLRCVRȩ MD
on AI-related challenges and opportunities,
which will feed into the policy development
process, the legislative evaluation process
ȩ ?LBȩ RFCȩ BCTCJMNKCLRȩ MDȩ ?ȩ LCVRECLCP?RGML
ȩ %
ȩ &GEF.CP-
digital strategy.
The EC is currently working with Member States to
have a coordinated plan on AI by the end of 2018. The
ȩ QK?PRȩAGRGCQȩ?LBȩR?VDP?SBȩBCRCARGML
The plan for AI will also link to the ecosystem of
QSNNMPRGLEȩ RCAFLMJMEGCQȩ 'ȩ AFGNQ
formance Computing (HPC) that are already in the
pipeline or are planned. In this respect, it is worth
noting that in June 2018 the EC proposed a Digital
Europe programme20 with a budget of EUR 9.2 bil-
K?GLȩ ?GKȩ GQȩ RMȩ K?VGKGQCȩ RFCȩ GKN?ARȩ MDȩ GLTCQRKCLRȩ
at the EU and national levels, encourage cooperation
?APMQQȩRFCȩ#3
ȩCVAF?LECȩ@CQRȩNP?ARGACQ
ȩ?LBȩQSNNMPRȩ GLE
ȩ ?PRGȏAG?Jȩ GLRCJJGECLAC
ȩ AW@CPQCASPGRW
ȩ ?BT?LACBȩ
the EU’s global competitiveness in this sector.
The plan should last until 2027, i.e. the remaining
RGKCȩ SLBCPȩ RFCȩ ASPPCLRȩ ȏL?LAGLEȩ DP?KCUMPIȩ ?LBȩ
RFCȩCLRGPCȩNCPGMBȩMDȩRFCȩLCVRȩ+SJRG?LLS?Jȩ$GL?LAG?Jȩ
Framework 2021-2027 to support better European
and national strategies.
The areas currently listed as priorities for collabor-
ation among the Member States include:
•ȩ +?NNGLEȩ PCQC?PAFȩ CVACJJCLACȩ ?LBȩ HMGLGLEȩ
JGMLȩDMPȩRFCȩNCPGMBȩȩ2FCPCȩ?PCȩȏTCȩ?PC?Qȩ
proposed as the focus: high-performance comput-
digital skills, and ensuring their wide use across the
economy and society.
Within the Digital Europe Programme, the EC pro-
poses to develop common ‘European libraries’ of
algorithms that would be accessible to all to help
the public and private sectors identify and acquire
whichever solution would work best for their needs.
-NCLȩNJ?RDMPKQȩ?LBȩ?AACQQȩRMȩGLBSQRPG?JȩB?R?ȩQN?ACQ
DMPȩ ?PRGȏAG?Jȩ GLRCJJGECLACȩ UGJJȩ @Cȩ K?BCȩ ?T?GJ?@JCȩ
across the EU in Digital Innovation Hubs, providing
testing facilities and knowledge to small business-
ȩ CȎMPRQȩRMȩGLAPC?QCȩN?PRLCPQFGNQȩGLȩ#SPMNC es and local communities.

• Interoperable dataset repositories (or ‘data In addition to the above regulatory package, oth-
spaces’) making quality data available for a er aspects are currently been addressed such
broad range of users, including trusted as: Liability (including the revision of the Product
 39 4. AI in the EU

Machinery Directive), and an Action Plan on Disin-

formation.

The European action is not limited to the development

of an appropriate legislation (and standards) but
Coordinated also includes a set of relevant initiatives to speed
up technology development and its uptake. Some
Plan for the PCJCT?LRȩ CV?KNJCQȩ DMPȩ 'ȩ GLAJSBC
ȩ DMPȩ CV?KNJCȩ %
ȩ
%?JGJCM
ȩTCPWȩFGEFQNCCBȩ?LBȩS@GOSGRMSQȩAMLLCARGTG-

Development ty, European Processor Initiative, Joint Undertaking

for High Performance Computing, and the Big Data

of AI in Europe Public-Private Partnership.

aims to maximise Moreover, sectoral policies are being discussed to

regulate the use of the ICT technologies (including
'ȩGLȩQNCAGȏAȩQCEKCLRQ
ȩDMPȩCV?KNJC
the impact • The Third Mobility Package aims to allow
of investment ȩ AGRGXCLQȩ RMȩ @CLCȏRȩ DPMKȩ Q?DCPȩ RP?ȑA
ȩ JCQQȩ
polluting vehicles and more advanced tech-
at EU and national nological solutions, while supporting the
competitiveness of the EU industry.

levels, and promote Particular focus is given to autonomous mo-

bility that has the potential to make tran-

cooperation. sport safer, more accessible, inclusive and

sustainable. The Commission’s proposal en-
compasses a strategy aiming to make
Europe a world leader for fully automated
and connected mobility systems.
 4. AI in the EU 40

• The Clean Energy for all Europeans package In the following sections, we provide a summary of
was presented in November 2016 and aims selected national strategies on AI.
to provide the stable legislative framework
needed to facilitate the clean energy transition. $P?LAC

• The Commission Directive 2018/844 aims to France is one of the more active countries in devel-
ȩ GKNPMTCȩ CLCPEWȩ CȑAGCLAWȩ RFPMSEFȩ BGEGR?J oping a national AI strategy. The government has
solutions such as building automation and recently published three main reports which are re-
electronic monitoring of technical building viewed below:
systems.
5IF
'SFODI
"*
QMBO
'*"



MȎCPQȩ?ȩQRP?RCEWȩDMPȩ
WȧL?RSPC
ȧ'ȧRCAFLMJMEWȧUGJJȧNCLCRP?RCȧ?JJȧGLBSQ- the coming years focusing on research and educa-
RPG?JȧQCARMPQ
ȧBGEGR?JȧQCPTGACQȧ?LBȧQMAG?JȧJGDCȧ'RȧGQȧ tion, innovation, and the social and economic im-
RFCPCDMPCȩCVRPCKCJWȩBGȑASJRȧRMȧPCAMLAGJCȧRFCȧLCCBȧ pacts of AI.
DMPȧ ?ȧ AJC?Pȧ AMKKMLȧ DP?KCUMPIȧ ?LBȧ RFCȧ P?NGBȧ
?BMNRGMLȧGLȧQMKCȧBMK?GLQ when the technology is 5IF
SFQPSU
CZ
UIF
0ҪDF
PG
UIF
1BSMJBNFOU
GPS
4DJ
already mature to enter the market. To monitor the FODF
BOE
5FDIOPMPHZ
01&$45


‘Toward a
BCTCJMNKCLRȩ?LBȩSNR?ICȩMDȩ'ȩGLȩBGȎCPCLRȩQCARMPQȩ !MLRPMJJCB
ȩ 3QCDSJȩ ?LBȩ "CKWQRGȏCBȩ PRGȏAG?Jȩ 'LRCJ-
and its impacts on work and society, in December ligence’ focuses on social and regulatory aspects
2018, the EC is launching an AI-Watch facility with and raises important issues to be addressed in de-
RFCȩDMJJMUGLEȩM@HCARGTCQȩ veloping a socially useful AI, whilst being realistic
1) Develop an overview and analysis of the
European AI ecosystem,
ȩ ?BBPCQQCQȩ QGVȩ
2) Monitor the uptake of AI applications across
the economy,
3) Monitor the progress of AI technology,
4) Assess the evolution of service robotics over
the last 10 years,
5)ȩȩ%?RFCPȩGLDMPK?RGMLȩMLȩ?JJȩ#3ȩ+CK@CPȩ1R?RCQȩ
?@MSRȩCVNCAR?RGMLQȩMLȩNMRCLRG?JȩGKN?ARQȩ
5IF
.JTTJPO
7JMMBOJ
SFQPSU

main areas:
1. An innovative and ambitious industrial and
economic policy: the mission opted to con-
centrate on some key sectors: health, trans-
portation, environment, and defence-secur
ity. The policy will require public support but
along with some structuring and precise
challenges.
national initiatives on AI, 2. 2MȩBCTGQCȩ?ȩB?R?ȩNMJGAWȩȏRRCBȩDMPȩRFCȩQR?ICQ
ȩ AACQQȩ RMȩ ?ȩ QSȑAGCLRȩ ?KMSLRȩ MDȩ B?R?
6) Provide an overview on the use of AI in
public services,
7)ȩ "CTCJMNȩ ?Lȩ 'ȩ GLBCVȩ GLAJSBGLEȩ RFCȩ BGKCL viding new protection to users.
sions relevant for policymaking,
8) Make information available to the public
using an online tool.
ȩ GQȩ ?ȩ K?HMPȩ GQQSCȩ GLȩ $P?LACȩ ?LBȩ #SPMNCȩ 2FCȩ
mission is trying to favour wide access,
the creation of data ecosystems while pro-
3. To anticipate and monitor the impact of AI
on work and employment.
4. AI as a tool for a sustainable and ecological
economy, providing a vision for a ‘greener’
 41 4. AI in the EU

AI enabling an ecological transition. For with the suggestion to at least double the starting
instance, the EU could take a leading role salary for researches in the public sector in view of
ȩ GLȩ RFCȩ BCTCJMNKCLRȩ MDȩ QNCAGȏAȩ AFGNQȩ GLȩ RFC the high degree of competition with industry.
semi-conductor industry, preparing the EU
industry for the post-silicon era. 21 3LGRCB)GLEBMK

5. Setting a framework for ethics and trust 'Lȩ 

ȩ RFCȩ %MTCPLKCLRȩ -ȑACȩ DMPȩ 1AGCLACȩ 3)ȩ
to enable the growth of AI.
ȩ ?ȩ PCJC?QCBȩ ?ȩ LMRCȩ MLȩ ~0M@MRGAQ
ȩ ?SRM-
ȩ UFGAFȩ CVNJMPCBȩ
6. 0CBSAGLEȩ RFCȩ @P?GLȩ BP?GLȩ ?LBȩ ȏLBGLEȩ U?WQȩ
ȩ RMȩ MȎCPȩ ?RRP?ARGTCȩ UMPIGLEȩ AMLBGRGMLQȩ RM
both French and international researchers.
2FCȩȏL?JȩPCNMPRȩNJC?BQȩDMPȩ?LȩGLAJSQGTCȩ?LBȩBGTCPQCȩ
AI stressing, for instance, a role to improve gen-
der balance. To develop an inclusive policy for AI,
%MT
K?RGMLȩ ?LBȩ ?PRGȏAG?Jȩ GLRCJJGECLAC
the opportunities and recommended three actions
concerning challenge areas, facilities and skills. The
same entity released a report that same year ‘Ar-
RGȏAG?Jȩ GLRCJJGECLACȩ MNNMPRSLGRGCQȩ ?LBȩ GKNJGA?RGMLQȩ
for the future of decision making’ (2016b) that in-
cluded a presentation of what AI is, a review of the
use of AI by governments, and a discussion on the
RFCȩPCNMPRȩQSEECQRQȩGLAMPNMP?RGLEȩ?ȩBS?JȩM@HCARGTCȩ
ȩRFCȩJ?Lȩ2SPGLEȩ
‘'JSTU

UP
FOTVSF
UIBU
UIF
EFWFMPQNFOU
PG
"*
UFDIOPM
PHZ
EPFT
OPU
DBVTF
BO
JODSFBTF
JO
TPDJBM
BOE
FDP
OPNJD
JOFRVBMJUZ
4FDPOE
UP
DBMM
PO
"*
JO
PSEFS
UP

SFEVDF
UIJT’ (Mission Villani, 2018: 133).
The mission advocates building a data-focused
economic policy grounded in a European data eco-
ȩ
system. The report holds that the state has to be a
key driver in these various areas of transformation
and suggests setting up an inter-ministerial coordi-
nator to implement this strategy, with support from

GVOEJOH
BOE
JOOP
CȎCARQȩMLȩJ?@MSPȩK?PICRQȩ'Lȩ
Institute was created as the national institute for
data science, headquartered at the British Library
(Hall & Pesenti, 2017).
In 2016, the House of Commons Science and Tech-
nology Committee released a report on ‘Robotics
?LBȩ ?PRGȏAG?Jȩ GLRCJJGECLACȩ &MSQCȩ MDȩ !MKKMLQ
2016). The report deals mostly with the broad is-
sues (economic, social, ethical, and regulatory)
triggered by this disrupting technology, but the last
AF?NRCPȩCV?KGLCQȩ~UIF
SFTFBSDI
?ȩQF?PCBȩQNCAG?JGQRȩACLRPCȩ-LCȩMDȩRFCȩLMTCJȩ?QNCARQȩ WBUJPO
MBOETDBQF
GPS
SPCPUJDT
BOE
"*’. Its conclusions
GQȩ RMȩ QCRȩ SNȩ DMSPȩ RMȩ QGVȩ 'LRCPBGQAGNJGL?PWȩ 'LQRGRSRCQȩ were rather critical: ‘(PWFSONFOU
MFBEFSTIJQ
IBT

DMPȩPRGȏAG?Jȩ'LRCJJGECLACȩ'ȩGLQRGRSRCQȩL?RGMLUGBC
ȩ CFFO
OPUJDFBCMZ
MBDLJOH
5IFSF
JT
OP
(PWFSONFOU

organised into a network: the National Network of TUSBUFHZ
GPS
EFWFMPQJOH
UIF
TLJMMT

BOE
TFDVSJOH
UIF

'LRCPBGQAGNJGL?PWȩ'LQRGRSRCQȩDMPȩPRGȏAG?Jȩ'LRCJJGECLAC DSJUJDBM
JOWFTUNFOU

UIBU
JT
OFFEFE
UP
DSFBUF
GVUVSF

HSPXUI
JO
SPCPUJDT
BOE
"*’ (UK House of Commons,
These new entities are meant to improve the rela- 2016: 33). In its answer to the report, the govern-
tionships between public basic research and private ment acknowledged that more could be done to
ȏPKQȩ 'Rȩ GQȩ ?JQMȩ ?ȩ KC?LQȩ RMȩ ?TMGBȩ RFCȩ @P?GLȩ BP?GLȩ take a leading role, and in March 2017 it announced
of researchers. This network should be integrated an industry-led review on the conditions necessary
in a future European AI research area network that for the AI industry to continue to thrive and grow
could be modelled on the EMBL (European Molecular in the UK. In April 2018, the government published
Biology Laboratory), which has been operating suc- GRQȩ'ȩ1CARMPȩ"C?Jȩ3)ȩ%MTȩȩ?QȩN?PRȩMDȩ?ȩJ?PECȩ
ACQQDSJJWȩ QGLACȩ ȩ -RFCPȩ #SPMNC?Lȩ N?PRLCPQFGNQȩ
could be built with the instruments of the H2020
programme, like the current public-private partner-
ships in robotics and Big Data. Enhancing the appeal
of careers in public research is another concern,
industrial strategy to place the UK at the forefront
of AI development and use. The Deal invests £ 950
million, of which two thirds is new money and the
rest is budget already allocated.

2FCȧ"C?JȧF?QȧȏTCȧK?GLȧNGJJ?PQ
*EFBT: Support research in data science AI with
£300 million, including £49 million for the Alan Tu-
ȩGLAJSBGLEȩ`ȩ@GJJGMLȩRMȩBCTCJMNȩ%ȩ?LBȩCVRCLBȩ
ring Institute, stimulate the uptake of AI, including
within the public sector, and invest up to £100 mil-
lion to support AI innovation to raise productivity.
The Institute’s public policy programme, in particu-
lar, is where the UK government comes for inde-
pendent advice on data science, AI, and ethics.
1FPQMF:
Work with schools, universities and industry
to ensure a highly skilled workforce, including £ 400
million to support STEM teaching in schools, and up to
ȩ?BBGRGML?Jȩ.F"ȩ@SPQ?PGCQȩGLȩ'PCJ?RCBȩȏCJBQȩUGRFȩ
the target of reaching 1 000 government-supported
PhD places per year in 2025. Also enable access to
highly skilled global talent and promote diversity.
*OGSBTUSVDUVSF:
Enhance the UK’s data infrastructure
making more public-sector data available, includ-
ing geospatial data under the guidance of a newly
APC?RCBȩ%CMQN?RG?Jȩ!MKKGQQGML
ȩ?LBȩNPMKMRCȩ"?R?ȩ
4. AI in the EU 42
Most national
AI strategies focus
on strengthening
research, supporting
industry, sharing
data, and modernising
public administration.
Trusts as a way to share public- and private-sec-
tor data in a secure way. Promote the upgrading
and maintenance of the physical digital infrastruc-
RSPC
broadband.
#VTJOFTT
FOWJSPONFOU:
The government has set up
an AI council with industry and academia to over-
QCCȩRFCȩQRP?RCEWȩ?LBȩ?ȩLCUȩ-ȑACȩDMPȩ'ȩRMȩQSNNMPRȩ
its delivery. Support for business and innovation will
include a £ 2.5 billion investment fund, reaching
£ 7.5 billion in partnership with the private sector.
1MBDFT: Support the continued development of Lon-
BMLȩ?Qȩ#SPMNCQȩA?NGR?JȩDMPȩ'ȩ@SRȩ?JQMȩCVN?LBȩQSN-
port to regional clusters with the appropriate infra-
structure and skills. Further support for the Alan
Turing Institute to become the national academic
institute for AI and data science.
 43 4. AI in the EU

$GLJ?LB -RFCP#3AMSLRPGCQ

The government appointed a steering committee in Several other Member States already have AI strat-
+?WȩȩBSCȩRMȩPCNMPRȩGLȩNPGJȩȩȩȏPQRȩPCNMPRȩ egies in place or are in the process of adopting
U?QȩNS@JGQFCBȩGLȩ-ARM@CPȩȩ$GLLGQFȩ%MTȩȩ
setting out a vision of a country embracing AI in
every aspect of life in an open and ethical way. It
ȩ QSNNMPRGLEȩ
GBCLRGȏCBȩICWȩ?PC?QȩDMPȩ?ARGMLȩGLAJSBGLEȩ
• Enhancing competitiveness by developing
ȩ
industrial ecosystems to help in the applica-
tion of AI, and measures to facilitate the
access and use of AI and data for companies.
• Increasing the quantity and quality of data
shared in Finland, through legal frameworks
them. From those illustrated above, most strategies
F?TCȩQGKGJ?Pȩ?GKQȩ?LBȩM@HCARGTCQȩCK@P?AGLEȩ'ȩ?LBȩ
@CAMKCȩ ?ȩ JC?BGLEȩ AMSLRPWȩ GLȩ RFCȩ ȏCJB
innovation and business, promoting the use of AI
in the public sector, boosting R&D and skills, estab-
JGQFGLEȩMLCȩMPȩKMPCȩL?RGML?JȩACLRPCQȩMDȩCVACJJCLAC
and developing strategies to promote the sharing
and use of data more widely between the public
and private sectors. A European strategy needs to
@SGJBȩMLȩRFCQCȩL?RGML?JȩCȎMPRQȩ?LBȩGLTCQRKCLRQȩRMȩ
make the whole greater than the sum of its parts.
that are ‘based on the importance of the 1SKK?PW?LBAMLAJSQGMLQ
data to business operations (not on data
ȩ NPMRCARGMLȩ ȏPQR
ȩ PCESJ?RMPWȩ Q?LB@MVCQȩ DMP This chapter shows thatȧK?LWȧ#SPMNC?LȧAMSLRPGCQȧ
ȩ CVNCPGKCLR?RGML
ȩ GLAPC?QCBȩ ?AACQQȩ RMȩ RFC ?QȧUCJJȧ?QȧRFCȧ#!ȧ?PCȧBCTCJMNGLEȧQRP?RCEGCQȧ?LBȧ
public of the government’s Mydata network, NPMEP?KKCQȧ RMȧ ESGBCȧ RFCȧ BCTCJMNKCLRȧ MDȧ ',
and support to companies to turn their data with shared concerns over the need for an agreed
into products. ethical framework and applications that clearly
@CLCȏRȩ#SPMNC?LȩQMAGCRWȩ?LBȩSNFMJBȩRFCȩ#SPMNC?Lȩ
• Launching an AI accelerator pilot in 2018 values enshrined in the Treaties. The High-Level
based on open data, open code, open inter- #VNCPRȩ%PMSNȩCQR?@JGQFCBȩGLȩȩ@WȩRFCȩ#SPMNC?Lȩ
faces with companies providing anonymised Commission is elaborating ethical and investment
data for research, partnership with univer- guidelines that will provide a framework for subse-
sities to research on the data, and access to quent developments.
ȩ QK?JJȩAMKN?LGCQȩRMȩCVNCPGKCLRȩUGRFȩRFCȩSQCȩ
of AI-based solutions. In addition, a number +MQRȧ L?RGML?Jȧ QRP?RCEGCQȧ ?L?JWQCBȧ EGTCȧ QGKGJ?Pȧ
of open piloting and testing environments ?RRCLRGMLȧ RMȧ QRPCLERFCLGLEȧ RFCGPȧ PCQC?PAF base,
ȩ UGJJȩ@CȩQCRȩSNȩRMȩD?AGJGR?RCȩCVNCPGKCLR?RGMLȩ
• #QR?@JGQFGLEȩ ?ȩ LCUȩ !CLRPCȩ MDȩ #VACJJCLACȩ DMP
AI and basic research, including a virtual
university and the creation of massive open
ȩ MLJGLCȩ AMSPQCQȩ ++-!ȩ RMȩ QSNNMPRȩ RFCȩ
including the setting up of one or more national cen-
tres for AI, QSNNMPRȧ DMPȧ RFCGPȧ GLBSQRPW, and SMEs,
and awareness of the LCCBȧ RMȧ QF?PCȧ B?R?ȧ @CRRCP
between all the stakeholders: the public sector, in-
dustry, and the public. They also focus on applica-
tions aimed at KMBCPLGQGLEȧNS@JGAȧ?BKGLGQRP?RGMLQ,
upskilling of the population and workforce. ?QȩUCJJȩ?QȩQNCAGȏAȩQCARMPQȩQSAFȩ?QȩFC?JRF

• 1SNNMPRGLEȩPCQC?PAFȩ?LBȩGLLMT?RGML
ȩ?LBȩCV
tending the application of AI in public service,
ȩ DMPȩ CV?KNJCȩ UGRFȩ RFCȩ BCTCJMNKCLRȩ MDȩ NCP
sonalised AI assistants, and new partnership
models between the public and private sector
to facilitate the adoption and use of AI.
 5. The AI ecosystem in China 44

$IJOB
JT
JNQMFNFOUJOH
JUT
BNCJUJPVT
TUSBUFHZ
PO
"*

SUMMARY

This chapter highlights some of the main features of the AI developments in China,
UFGAFȩ F?Qȩ GBCLRGȏCBȩ 'ȩ ?LBȩ PCJ?RCBȩ ?NNJGA?RGMLQȩ ?Qȩ QRP?RCEGAȩ RMȩ GRQȩ DSRSPCȩ CAMLMKGAȩ
and social development, and to its role as a global superpower. As a result, it has put
in place a strongly coordinated approach to AI, including government policy, industrial
?NNJGA?RGMLQȩ?LBȩPCQC?PAFȩUGRFȩRFCȩM@HCARGTCȩMDȩ@CAMKGLEȩRFCȩUMPJBȩJC?BCPȩGLȩ'ȩ@Wȩ
2030. This is an ambitious but achievable target as the JRC analysis of the AI inno-
T?RGMLȩCAMQWQRCKȩGLȩ!FGL?ȩAMLȏPKQȩ2FGQȩMNGLGMLȩGQȩQF?PCBȩ@WȩK?GLȩK?PICRȩ?L?JWQRQȩ
In fact, AI deployments with a high impact on daily activities have already started,
including authentication processes, medical diagnosis, premium insurance, transpor-
tation and retail facilities, and security. Potential obstacles to the full development of
RFCȩQRP?RCEWȩPCJ?RCȩRMȩRFCȩBGȑASJRWȩMDȩF?TGLEȩ'ȩ?BMNRCBȩGLȩRP?BGRGML?JȩGLBSQRPGCQ
ȩ?LBȩ
prepare enough specialists in the country able to understand the domain of applica-
RGML
ȩGRQȩLCCBQȩ?LBȩCVNCAR?RGMLQ
ȩGBCLRGDWȩRFCȩ@CQRȩ?JEMPGRFKȩ?LBȩ?NNPM?AF
ȩ?LBȩRFCLȩ
customise it and develop it further.
 45 5. The AI ecosystem in China

THE AI
ECOSYSTEM
IN CHINA
!FGL?QCAMLMKGA?LBNMJGAWAMLRCVR China has put
The development of AI is a global phenomenon, but
the case of China is particularly interesting and not
in place a strongly
as well known as that of the USA. For this reason,
we dedicate a slightly longer section to this country,
coordinated approach
drawing on a comprehensive study carried out for
RFCȩ(0!ȩ@Wȩ$CGHMMȩȩ
to AI, including
China is already a leading global force in the digital
government policy,
CAMLMKWȩ $MPȩ CV?KNJC
ȩ GRȩ F?Qȩ  ȩ Ϥȩ MDȩ RFCȩ EJM@?Jȩ
share of e-commerce, it processes 11 times more
industrial applications
mobile payments than the USA, it is home to one-
third of the world’s unicorns22, and leads the im-
and research
NJCKCLR?RGMLȩMDȩ%ȩKM@GJCȩAMKKSLGA?RGMLQȩ5MCR-
zel et al., 2017). This digital success is fuelled by
with the objective
three main factors: (i) a large and young Chinese
market enabling rapid commercialisation of digital
of becoming
@SQGLCQQȩ KMBCJQȩ GGȩ ?ȩ PGAFȩ BGEGR?Jȩ CAMQWQRCKȩ CV-
panding quickly beyond a few large companies;
the world leader
and (iii) strong governmental support for compa-
nies through favourable economic and regulatory
FK'?V
conditions, and a multiple role as strategic investor,
consumer of digital technologies, and provider of driver to increase the quality of products and pro-
access to key data. BSARGTGRW
ȩ ?LBȩ CVN?LBȩ GLRCPL?Jȩ AMLQSKNRGML
ȩ RFSQȩ
At the macroeconomic level, China can no longer
depend solely on increases in capital and labour to
achieve the desired levels of sustainable economic
growth, nor on producing cheap goods to sell else-
where as has been the predominant model for the
last three decades. Against this background, there
has been a recent shift into innovation as a strategic
BCAPC?QGLEȩ RFCȩ BCNCLBCLAWȩ MLȩ CVRCPL?Jȩ K?PICRQȩ
Digital technologies in general, and AI in particular,
?PCȩCVNCARCBȩRMȩJC?BȩRMȩFGEFCPȩCȑAGCLAWȩGLȩNPMBSARQȩ
and services, achieving large productivity gains and
a more dynamic economy, with businesses able to
AMKNCRCȩEJM@?JJWȩ?LBȩCTCLȩCVNMPRȩ~+?BCȩGLȩ!FGL?ȩ
digital business models and solutions (Woetzel et
al., 2017).
 5. The AI ecosystem in China 46

But there are other interests beyond mere economics. AMKNSRGLEȩ GQȩ PCJCT?LRȩ QGLACȩ GRȩ GQȩ CVNCARCBȩ RF?Rȩ ?ȩ
China, as a global power, is using initiatives such as considerable part of AI services will be provided
~-LCȩ@CJR
ȩMLCȩPM?BȩMPȩGLQRGRSRGMLQȩQSAFȩ?QȩRFCȩQG?Lȩ from the cloud. Therefore, market developments
Infrastructures Investment Bank to shape a geopo- in both AI and cloud computing will be considera-
litical environment more favourable to its interests bly aligned.
(Delage, 2017). Digital technologies, and AI in par-
ticular, are essential constituents of this strategy,
showcasing a state-of-the-art new industry fully
BCTCJMNCBȩGLȩ!FGL?ȩ?LBȩ?ȩBGȎCPCLRȩ?NNPM?AFȩDPMKȩ
the West to economy and policy through technology.
Since 2014,
The strict application of national laws to cyberspace the government
and the use of AI-related technologies to maintain
QMAG?JȩF?PKMLWȩ?LBȩAMLRPMJȩ?PCȩRFCȩK?GLȩCV?KNJCQȩ has launched a series
MDȩRFGQȩBGȎCPCLRG?RCBȩ?NNPM?AFȩ
of key national
The potential of AI is clear to China’s policymak-
ers. As stated by President Xi, China ‘… should un- economic initiatives
swervingly follow an independent innovation path
featuring Chinese characteristics, stick to the guid-
that are relevant to AI
ing principles of independent innovation, leap-frog-
ging development in key sectors, with development
with the goal of creating
supported by science and technology and oriented
towards the future’ (Xi, 2014, p. 134). In fact, since
a EUR 13 billion
2014, the government has launched a series of key AI market in China
national economic initiatives that are relevant to AI
with the goal of creating a EUR 13 billion AI market by 2018, and make
in China by 2018, and helping China to lead the world
in AI by 2030. China lead the world
In the case of AI, and in contrast to what happened in AI by 2030.
in the recent past, development of the sector re-
lies on a plan that encompasses all relevant play-
ers within an ecosystem, from universities and re-
QC?PAFȩACLRPCQȩRMȩCVGQRGLEȩAMKN?LGCQȩ?LBȩLCUȩȏPKQȩ
in the entrepreneurial / innovation milieu, which is
also under development in China. It will be, there-
DMPC
ȩ?ȩȏPQRȩPC?Jȩ?QQCQQKCLRȩMDȩ!FGL?QȩAF?LACQȩRMȩ
APC?RCȩ ?Lȩ GLBCNCLBCLRȩ FMKCEPMUL
ȩ DSJJWȐCBECBȩ ment initiatives related to AI following a timeline
innovation and industrial ecosystem.
However, AI is not an isolated ecosystem as it inter-
?ARQȩ?LBȩ@CLCȏRQȩDPMKȩLMRȩMLJWȩRFCȩBCTCJMNKCLRȩ
of successful innovations such as ecommerce or
mobile payments but also from cloud computing,
industry 4.0, robotics, IoT, blockchain, and IT se-
curity ecosystems. The particular case of cloud
%MTCPLKCLRNMJGAGCQ?LBGLGRG?RGTCQ
In this section, we summarise the main govern-
up to 2018. The timeline is interesting because it
displays how NMJGAGCQȧ ?LBȧ GLGRG?RGTCQȧ QR?PRCBȧ ?Rȧ
GLBSQRPWȧ JCTCJȧ QJMUJWȧ EMȧ SNȧ RFCȧ T?JSCȧ AF?GLȧ RMȧ
GLAJSBCȧ ?NNJGA?RGMLQȧ ?LBȧ AMLQSKCPȧ NPMBSARQ. It
is particularly relevant to follow the developments
RF?Rȩ ?JJMUCBȩ AMKN?LGCQȩ GLȩ BGȎCPCLRȩ QCARMPQȩ RMȩ
share and access data to train newer and better
algorithms. Crucially then, the government issued
 47 5. The AI ecosystem in China

QNCAGȏAȩ'ȩESGBCJGLCQȩAMTCPGLEȩRFCȩUFMJCȩT?JSCȩLCR- The strategy includes the promotion of an open

work in a holistic perspective. industrial ecosystem based on the dynamic per-
ception of consumer demand and new industrial
internet applications, such as intelligent moni-
toring, remote diagnosis and management, and
whole industry chain tracing.

*OUFSOFU

"DUJPO
(VJEBODF
4UBUF
$PVODJM
+VMZ



The internet+ strategy focuses on intelligent man-
ufacturing and innovation, promoting the develop-
ment of cloud computing, IoT, intelligent industrial
robots, industrial control systems, IntelliSense com-
ponents, industrial operating systems and soft-
ware, and additive manufacturing technologies. It
also includes an upgrade of production equipment,
processes and a focus on (big) data sharing. The
QRP?RCEWȩM@HCARGTCQȩQNCAGȏAȩRMȩ'ȩ?PC

• To increase public support for the develop-

ment of AI and core technological break-
throughs;

• To promote the popularisation and applica-

ȩ RGMLȩ MDȩ 'ȩ GLȩ ȏCJBQȩ QSAFȩ ?Qȩ QK?PRȩ FMKC

smart terminal, smart car, and robots;

• To create a new computing cluster that

supports ultra-large-scale DL;

• To build a massive training resource base

including voice, images, videos, maps and
other data;

• To strengthen the construction of innovation

platforms of basic AI resources and public
services;
.BEF
JO
$IJOB

4UBUF
$PVODJM
.BZ


The goal of this strategy is to make China an ad- • To research, develop and industrialise key AI
vanced and prestigious manufacturing power by technologies such as computer vision, intel-
2025. In fact, intelligent manufacturing is positioned ligent speech processing, biometrics, natural
as the main development direction of ‘Made in Chi- language understanding, intelligent decision
na’. The main industries targeted are machinery, avi- control and new human-computer interaction;
?RGML
ȩQFGNNGLE
ȩ?SRMKM@GJCQ
ȩJGEFRȩGLBSQRPW
ȩRCVRGJCQ
ȩ
food and electronics, and the general aim is to im- • To support cooperation among security en-
prove the precision manufacturing and agile manu- terprises and internet companies, in order
facturing capabilities. to develop and promote big data analysis
 5. The AI ecosystem in China 48

technologies such as image recognition and •ȩ Wȩ

ȩK?HMPȩ@PC?IRFPMSEFQȩUGJJȩ@CȩK?BC
enhance the intelligent service level of secur- in the basic AI theory. Some of the techno-
ity products; logies and applications will reach the world
ȩ JC?BGLEȩJCTCJȩ'ȩUGJJȩ@CAMKCȩRFCȩK?HMPȩBPGTGLEȩ
• To promote the application of internet force for industrial upgrading and economic
technologies, intelligent sensing, pattern restructuring in China. It will have an impact
recognition, data analytics and robotics; on society.

• To nurture a number of key enterprises and • By 2030, AI theory, technologies and applica-
innovation teams to lead the global develop- tions will be world leaders. China will become
ment of AI. ȩ RFCȩUMPJBQȩK?HMPȩ'ȩGLLMT?RGMLȩACLRPCȩ

3PCPU
*OEVTUSZ
%FWFMPQNFOU
1MBO




2FCȩNJ?LȩKCLRGMLQȩQGVȩICWȩR?QIQȩ
.JOJTUSZ
PG
*OEVTUSZ
BOE
*OGPSNBUJPO
5FDIOPMPHZ

/BUJPOBM
%FWFMPQNFOU
BOE
3FGPSN
$PNNJTTJPO
• To build an open, collaborative AI technology
.JOJTUSZ
PG
'JOBODF
"QSJM

 innovation system, cutting edge in terms of
The main goal of this plan is the development of basic theory, key common technologies,
GLRCJJGECLRȩGLBSQRPG?Jȩ?LBȩQCPTGACȩPM@MRQȩ1GVȩ?PC?Qȩ innovation platforms, and high-end talent.
?PCȩ BCȏLCBȩ DMPȩ GLBSQRPWȩ UCJBGLEȩ PM@MRQ
ȩ T?ASSKȩ
robots, fully autonomous programming intelligent •ȩ 2Mȩ DMQRCPȩ ?ȩ FGEFCLBȩ ?LBȩ CȑAGCLRȩ QK?PR
robots, human-computer collaboration robots, economy, develop the emerging AI industry,
dual-arm robots and heavy-duty autonomous ve- promote industrial upgrade and create AI
FGAJCQ
ȩNJSQȩDMSPȩ?PC?QȩDMPȩRFCȩQCPTGACȩBMK?GLȩȏPCȩ innovation areas.
rescue robots, surgical robots, intelligent public ser-
vice robots, and intelligent nursing robots. The plan • To build a safe and supportive smart socI-
also proposes that the government establishes ȩ CRW
ȩBCTCJMNȩCȑAGCLRȩ?LBȩGLRCJJGECLRȩQCPTGACQ

robotic innovation centres and encourages banks including for social governance, enhance
and private investors to support the development public safety and security based on AI and
of robotics and robots. promote sharing and mutual trust in social
interactions.
/FX
(FOFSBUJPO
"*
%FWFMPQNFOU
1MBO

4UBUF
$PVODJM


+VMZ


 • To strengthen military and civilian integra-
This is the key strategy for AI where guidelines, ȩ RGMLȩGLȩRFCȩȏCJBȩMDȩ'
ȩNPMKMRCȩRFCȩRUMU?Wȩ
strategic goals, key tasks and supporting measures conversion of AI technology between military
for a new-generation AI development by 2030 are ȩ ?LBȩAGTGJG?LȩQCARMPQȩ?LBȩHMGLRJWȩ@SGJBȩSNȩ?LB
QCRȩSNȩ2FCȩNJ?LȩBCȏLCQȩTCPWȩAJC?PJWȩRFCȩQRP?RCEGAȩ
goals in AI development:
ȩQ?DCȩ?LBȩCȑAGCLR
• By 2020, most AI technologies and applica-
tions will be on a par with the most advanced
levels worldwide, and the AI industry will
become a new important economic growth
net contributor. AI will have an impact on
our daily lives.
share military and civilian innovation resources.
•ȩ 2MȩAMLQRPSARȩ?ȩS@GOSGRMSQ
intelligent infrastructure system and strength-
en the upgrading of infrastructures such as
networks, big data and high-performance
computing.
bȩ 2Mȩ NJ?Lȩ QGELGȏA?LRȩ QAGCLACȩ ?LBȩ RCAFLMJMEW
NPMHCARQ
ȩ?LBȩQRPCLERFCLȩMTCP?JJȩAMMPBGL?RGML
 49 5. The AI ecosystem in China

0CEGML?J JMA?JGLGRG?RGTCQ industrial organisation, innovation and entrepre-

There are many initiatives taking place in China
at the regional and local level and supported by
the respective governments. This seems to create
?ȩN?PRGASJ?PȩRWNCȩMDȩAMKNCRGRGMLȩ@CRUCCLȩBGȎCPCLRȩ
locations that is a main feature of the AI innovation
ecosystem. Below, we give a few illustrative cases.
#FJKJOHȩ 'Lȩ 
ȩ RFCȩ JMA?Jȩ EMTCPLKCLRȩ MDȩ CGHGLEȩ telligent Equipment Manufacturing Industrial Park,
issued a guideline to support technologies, appli-
cations and the AI industry. It sets the goal for the
city to become an AI innovation hub with global in-
neurship. In 2017, Shanghai had about 500 innova-
tive parks hosting 16 000 start-ups.
(VBOH[IPVȩ %S?LEXFMSȩ F?Qȩ J?SLAFCBȩ ?ȩ DSJJȩ QCRȩ
of industrial parks supported by the local gov-
CPLKCLRȩ RFCȩ %S?LBMLEȩ 'LRCJJGECLRȩ +?LSD?ARSPGLEȩ
Demonstration Base, the Intelligent Industrial Park
in Huangpu Machinery Valley, the Zengcheng In-
the Huadu New Energy Vehicles and Intelligent
Manufacturing Equipment Industrial Park and the
Liwan 3D Printing Industrial Park. All of them in-
ȐSCLACȩ@Wȩȩȩ?LRGAGN?RGLEȩL?RGML?JȩM@HCARGTCQȩ clude direct support to innovative companies, par-
AAMPBGLEȩRMȩRFCȩ CGHGLEȩEMTCPLKCLR
ȩRFCȩPCEGMLȩF?Qȩ ticularly those linked to the transformation of the
more than half of the key AI research institutes in manufacturing industry.
China, nearly 400 AI enterprises, and during the
ȏPQRȩLGLCȩKMLRFQȩMDȩȩRFCPCȩUCPCȩGLTCQRKCLRQȩ )BOH[IPV: A radically distinct approach is taking
GLȩȩ'ȩNPMHCARQȩRMR?JJGLEȩ31ȩȩ@GJJGMLȩ
;IPOHHVBODVO: often referred to as China’s Silicon
Valley - is the main high-tech business district of
ȩRF?Rȩ?PCȩȏJJCBȩUGRFȩCLRPCNPCLCSPQȩUFMȩNPCTG-
place in Hangzhou. Alibaba’s success has shaped
Hangzhou’s tech scene. The city is home to many
incubators, funded in part by government subsi-
BGCQ
CGHGLEȩ 'Rȩ GLAJSBCQȩ QMKCȩ MDȩ RFCȩ KMQRȩ NPCQRGEGMSQȩ ously worked at Alibaba.
tech universities and about 500 colleges, national
QAGCLRGȏAȩPCQC?PAFȩGLQRGRSRGMLQȩ?LBȩ0"ȩACLRPCQȩDMPȩ
multinational companies. There is a plan to locate
the largest industrial cluster for integrated circuits
in China there. The high-tech companies located in
the district generated about US$ 600 billion in rev-
ȧD?ACȧ?LBȧTMGACȧ
enues in 2017 and the area is home to 67 unicorns.
ȧ RMECRFCPȧ UGRFȧ RFCȧ RCAFLGOSCQȧ RF?Rȧ
It is estimated that 30 000 overseas returnee en-
trepreneurs have established over 8 000 compa-
nies in Zhongguancun. The district also published
a plan at the end of 2017 for cultivating AI talent
UGRFȩ RFCȩ CVNCAR?RGMLQȩ MDȩ ?AFGCTGLEȩ ?Lȩ MSRNSRȩ DMPȩ
the AI industry there of more than EUR 7 billion in
ȩ RP?LQNMPR?RGML
ȩ SP@?Lȩ RP?ȑAȩ K?L?EC-
2020.
4IBOHIBJ: Shanghai has its adaptation of the cen-
tral government plan called ‘13th Five-Year Plan
for Shanghai Manufacturing Transformation and
Upgrading’, issued by the Shanghai Municipal Peo-
ȩ AMK@GLGLEȩ AMLTCLGCLAC
ȩ CȑAGCLAW
ȩ NCPQML-
2FCQNCAGȍAGRGCQMDRFC'GLLMT?RGML
 CAMQWQRCKGL!FGL?
2FCȧ 'ȧ GLLMT?RGMLȧ CAMQWQRCKȧ GLȧ !FGL?ȧ GQȧ KMQRȧ
BCTCJMNCBȧGLȧRFCȧ?PC?QȧMDȧGK?EC
PCAMELGRGML
QSNNMPRȧQSAFȧȏCJBQ (basically machine and DL and
the algorithms that accompany both). There is also
recent and particular emphasis on the develop-
ment of AI chips. As a consequence, the more de-
veloped areas of application in China are security,
FC?JRFA?PC
ment, commerce and driverless cars. Adding all of
RFCKȩ RMECRFCPȩ AMLȏESPCQȩ ?ȩ QACL?PGMȩ MDȩ BCCNȩ CAM-
nomic and social changes that have already started
to take place in China. Applications of AI are now
widely available and impact the daily lives of China’s
AGRGXCLQ
NJCQȩ %MTCPLKCLRȩ GLȩ ȩ 'Rȩ ?GKQȩ RMȩ A?NRSPCȩ RFCȩ ?JGQ?RGMLȩ?LBȩQSPTCGJJ?LACȩ1CCȩ MVȩ
opportunities from emerging technologies such
as AI, quantum communication, VR and preci- China’s large companies work closely with the
sion medical treatment, as well as new trends in government, and take advantage of the positive
 5. The AI ecosystem in China 50

perceptionȩMDȩAMLQSKCPQȩRMȩCVNCPGKCLRȩUGRFȩ?NNJGA
ations unheard of outside China. This fact perme- ɎɛɤǮ1MKCCV?KNJCQMD'?NNJGA?RGMLQ
GMLQ
ates the AI innovation ecosystem, but arguably, the GL!FGL?
foremost instance of alignment between consumers,
companies and government is the implementation Alibaba’s core business is selling goods and providing ding a
of the so-called ‘social credit system’, tentatively platform for  trade. Just to give a sense of scale,, durin during
ng
scheduled for 2020. This AI-based system, which |0FKDIBP}">V} a creation of the company that takess place cee
would be controversial in other countries, will be LK  ,LSBJ?BO
 IF?>?> MOL@BPPBA >Q MB>H QFJB JLOB
OB QE>K
QE>K
>KK
able to collect user behavioural data in both phys-   LOABOP MBO PB@LKA QEOLRDE MLMRM PQLOBP PQLOBP
ical space and cyberspace and provide a rating of PBIIFKDMOLAR@QPȭQQBATFQE3/JFOOLOPCLOSFOQR>IQOVLRQPK'
PK'
citizen reputation. This rating could then be used to fashion consultant helped match items with the one selected.
elected.
give or deny access to a range of services provided In one day it sold JLOBQE>K20”?FIIFLKTLOQELC OQELC
by the state. goods V @LKQO>PQ
 LK  !V?BO +LKA>V ,LSBJ?BO ?BO 

QEB?FDDBPQLKIFKBPELMMFKDA>VFKQEB20
>IIOBQ>FIBOP@LJ?FKBA
J?FKBA
Military involvement is another salient feature of the brought in 20”?FIIFLK)KFDEQ

AI ecosystem in China: civilian AI breakthroughs are
applied for military use (under the national strat- 1BK@BKQ}P J>FK PLRO@B LC A>Q> FP 4B!E>Q TEF@E
 FK 

egy of ‘military-civil fusion’) and, vice versa, mili- reached more than ?FIIFLKJLKQEIV>@QFSBRPBOP OP. Its
tary-based developments are transferred to civilian Ȯ>DPEFMMOLGB@QFPFQPEB>IQE@>OBMI>QCLOJQL?LLH>MMLFKQJBKQP
QJBKQP
industry. According to the Central Military Commis- >KAJ>HBM>VJBKQPTEF@E
FK
was available in  
sion for Science and Technology, AI could accelerate JBAF@>I C>@FIFQFBP PBOSFKD  JFIIFLK @RPQLJBOP LJBOOP.
military transformation by means of intelligent and 'K
FQI>RK@EBAFQP'MLTBOBAAF>DKLPQF@JBAF@>IFJ>DFKD
J>DFKDD
autonomous unmanned systems, AI-enabled intel- D over
PBOSF@B FK  ELPMFQ>IP 1EB PVPQBJ FP QO>FKBA RPFKD
ligence analysis, war-gaming simulation and train-  ?FIIFLK FJ>DBP COLJ 1BK@BKQ PL@F>I KBQTLOH
BQTLOH
GLE
ȩBCDCLAC
ȩMȎCLAC
ȩ?LBȩAMKK?LBȩGLȩGLDMPK?RGMLȩ complemented by anonymised patient data.
warfare; and intelligent support to command deci-
sion-making (Kania, 2017). In the other direction, F$IVQB@E
>CLOJBOPQ>OQRM
FPKLT>@LKPLIFA>QBA@LJM>KV
LJM>KV
!FGL?ȩU?LRQȩRMȩQGELGȏA?LRJWȩGLAPC?QCȩRFCȩQNGJJMTCPȩMDȩ PMB@F>IFPBA FK RPBO FKQBOC>@BP >KA SLF@B OB@LDKFQFLK 'QP '
military developments into the civilian industries. enabled general practitioner medical robot passed the written
test of China’s national medical licensing examination ation in
Another feature of the AI innovation ecosystem ,LSBJ?BOThe robot can automatically capture apture
in China is the competition among cities and re- and analyse patient information and make
ke an
gions for talent, in particular from the new policies initial diagnosis.
launched by second-tier cities. As a result, there
?PCȩ AMLQGBCP?@JCȩ BGȎCPCLACQȩ GLȩ RFCȩ ?AASKSJ?RGMLȩ
of talent and opportunities in new tech industries
between East and West China.
ȩȏL?JȩDC?RSPCȩMDȩRFCȩCAMQWQRCKȩPCDCPQȩRMȩRFCȩJGKGR-
ed role awarded to critical views of AI in China. Such
views could be perceived as a lack of loyalty to a
policy that is highly supportive of AI. However, this
1EBOB>OBJ>KVPQ>OQRMPFKQEB>OB>LCC>@BOB@LDKFQFLKJLKD
JLKKD
fy 
them Megvii claims that their cameras are able to identify 

people per second in a crowded train station, while LLVision
FP ABSBILMFKD 'BK>?IBA DI>PPBP PQLOFKD RM QL 
 

images of suspects on the device to achieve faster ter face
recognition. They are part of an ecosystem of companies
mpanies
supporting the government’s objective to set up a
national facial recognition database that willl store
?@QCLACȩMDȩAPGRGAGQKȩAMSJBȩJC?BȩRMȩ?ȩJ?AIȩMDȩTCPGȏ FKCLOJ>QFLK>?LRQQEB@LRKQOV}P?FIIFLK@FQFWBKP
QFWBKP,
able algorithms, trustworthy systems, or a situation and be able to identify them within three seconds.
in which moral responsibilities are outsourced to
algorithms. The calls for policy support of algorithmic
 51 5. The AI ecosystem in China

transparency and accountability present in the EU 1MKCȩCVNCPRQȩ?JQMȩGLAJSBCȩRFCȩJ?AIȩMDȩDSLB?KCLR?Jȩ

?LBȩRFCȩ31ȩCEȩ%?PȏLICJ
ȩ+?RRFCUQ
ȩ1F?NGPM
ȩȩ research into AI. This gap implies the need to devel-
Smith, 2017) but notoriously absent in China.
The developments in China are very impressive. How-
CTCP
ȩ RFCȩ (0!ȩ ?L?JWQGQȩ GBCLRGȏCBȩ QMKCȩ E?NQȩ GLȩ RFCȩ
current ecosystem and the roadmap to 2030. These
QRPSARSP?JȩE?NQȩ?PCȩLMRȩQNCAGȏAȩRMȩ'ȩ@SRȩ?PCȩ?JQMȩT?JGBȩ
for other technically-based innovation ecosystems in
China, a view supported by the main market analysts
(Barton et al., 2017; Mubayi et al., 2017).
2FCȩȏPQRȩE?NȩPCDCPQȩRMȩRFCȩQJMUȩ?BMNRGMLȩMDȩ'ȩUGRFGLȩ
traditional industries, many of which are supported
ȩGRQȩLCCBQȩ?LBȩCVNCAR?-
by local governments and some have a high level of
BC@R
ȩUFCPCȩRFCȩGKN?ARQȩMLȩHM@QȩAMSJBȩ@CȩN?PRGASJ?P-
MNȩ?ȩQSȑAGCLRȩOS?LRGRWȩMDȩ'ȩR?JCLRȩ?LBȩCLQSPCȩRF?Rȩ
the local education system is able to supply it. The
BGȑASJRGCQȩMDȩGLAPC?QGLEȩRFCȩNMMJȩMDȩR?JCLRȩ?PCȩ?JQMȩ
related to the set of knowledge and skills needed
to succeed in the application of AI to a particular
domain and the shortcomings in the innovation-re-
sistant education system in general. The issue at
QR?ICȩGQȩRF?RȩGRȩUGJJȩLMRȩQSȑACȩRMȩAPC?RCȩ?Lȩ'ȩ?N-
plication - a programmer could do it almost me-
chanically with the appropriate directions - but to
become a ‘functional specialist’ able to understand
RFCȩBMK?GLȩMDȩ?NNJGA?RGML
tions, identify the best algorithm and approach,
and then customise it and develop it further (Bac-
JWȩBGȑASJRȩȩPCJ?RCBȩGQQSCȩPCDCPQȩRMȩRFCȩE?Nȩ@CRUCCLȩ cala et al., 2018).
CVGQRGLEȩR?JCLRȩ?LBȩGRQȩ?NNJGA?RGMLȩUGRFGLȩCVGQRGLEȩGL-
dustries, including the distance between research and
practical implementations in (traditional) industry.  1SKK?PW?LBAMLAJSQGMLQ
This chapter highlights some of the main features
of the AI developments in China, and in particular
the synergy between government policy, private-

$ÀQDOIHDWXUH
sector investment, and research. China has identi-
ȏCBȩ'ȩ?LBȩPCJ?RCBȩ?NNJGA?RGMLQȩ?QȩQRP?RCEGAȩRMȩGRQȩ
future economic and social development, and to
of the ecosystem its role as a global superpower. As a result, it has
put in place a strongly coordinated approach to AI,

refers to including government policy, industrial applications

?LBȩ PCQC?PAF
ȩ UGRFȩ RFCȩ M@HCARGTCȩ MDȩ @CAMKGLEȩ RFCȩ

the limited role world leader in AI by 2030. This is an ambitious

but achievable target as the the JRC analysis of
RFCȩ'ȩGLLMT?RGMLȩCAMQWQRCKȩGLȩ!FGL?ȩAMLȏPKQȩ2FGQȩ
attributed to critical opinion is shared by main market analysts (Barton
et al., 2017; Mark Purdy et al., 2017). In fact, AI
views of AI deployments with a high impact on daily activi-
ties have already started, including authentication
in China. processes, medical diagnosis, premium insurance,
transportation and retail facilities, and security.
Potential obstacles to the full development of the
QRP?RCEWȩPCJ?RCȩRMȩRFCȩBGȑASJRWȩMDȩF?TGLEȩ'ȩ?BMNR-
ed in traditional industries, and preparing enough
specialists in the country able not only to advance
technological innovation but also to adapt it to the
needs of potential users.
 .02
+3*2'̂"'+#,1'-,*
PERSPECTIVES
 6 Ethical and Social Perspective 54

"*
JT
TIBQJOH
PVS
XBZT
PG
UIJOLJOH
BDUJOH
TPDJBMJTJOH
BOE
XPSLJOH

SUMMARY

In this chapter, we highlight some of the key dimensions of the ethical and societal debates on
AI. As shown, RFCȧ BC@?RCȧ ?@MSRȧ CRFGA?Jȧ ?LBȧ QMAG?Jȧ GKNJGA?RGMLQȧ MDȧ 'ȧ DMPȧ GLBGTGBS?JQȧ ?LBȧ
QMAGCRGCQȧLCCBQȧRMȧKMTCȧDMPU?PBȧ?LBȧD?QR. Political and regulatory frameworks seem to be
lagging behind AI developments, which will soon pervade almost every aspect of our daily lives.

To counter this situation, ‘regulatory foresight’ and ‘ethics foresight analysis’ at the same
level of ‘technological foresight’, are needed to keep up with emerging implications of AI. This,
however, cannot happen without a strong commitment to the following issues: DMFBS
FUIJDBM
GSB
NFXPSLT
BOE
HVJEFMJOFT, including the right to choose between being cared for by a human or
?ȩPM@MR
ȩ?LBȩRFCȩPGEFRȩRMȩPCDSQCȩ@CGLEȩNPMȏJCB
ȩRP?AICB
ȩMPȩKC?QSPCB
ȩ?LBȩSFTQPOTJCMF
"*
EFTJHO

which is necessary to build trust in the applications.

We conclude that, to build and retain trust in AI, we need a multi-layered approach that includes
the critical engagement of civil society to discuss the values guiding and being embedded into
'
ȩNS@JGAȩBC@?RCQȩGLȩBGȎCPCLRȩDMP?ȩRMȩRP?LQJ?RCȩRFCQCȩT?JSCQȩGLRMȩQRP?RCEGCQȩ?LBȩESGBCJGLCQ
ȩ?LBȩ
responsible design practices that encode these values and guidelines into AI systems, so that
they are ethical by design.
 55 6 Ethical and Societal Perspective

ETHICAL AND
SOCIETAL
PERSPECTIVES
'LRPMBSARGML
To build and retain
When addressing ethical and social impacts of
RCAFLMJMEW
ȩGRȩGQȩLMRȩHSQRȩ?@MSRȩ?TMGBGLEȩBCRPGKCL- trust in AI we need
tal or unintended consequences, or guaranteeing
that humans have control over their technologies. critical engagement
1SAFȩ CVRCPL?JGQRȩ ?LBȩ GLQRPSKCLR?JGQRȩ NCPQNCARGTCQȩ
tend to separate technologies from humans. That of civil society
is, technologies as mere neutral tools devoid of
T?JSCQ
ȩ ?LBȩ FSK?LQȩ ?Qȩ QMJCȩ K?QRCPQȩ BCȏLGLEȩ RFCȩ to discuss the values
terms of engagement.
guiding AI and
In reality, the relationship between technologies and
FSK?LQȩ GQȩ KSAFȩ KMPCȩ AMKNJCVȩ -Lȩ RFCȩ MLCȩ F?LB
ȩ responsible design
RCAFLMJMEGCQȩ?PCȩ@SGJRȩ@WȩQNCAGȏAȩNCMNJCȩ?LBȩMPE?L-
GQ?RGMLQȩ  GHICPȩ CRȩ ?J
ȩ 
ȩ QMȩ RFCWȩ CK@MBWȩ ?LBȩ practices that encode
replicate social norms, values, and other economic,
ecological, political, or cultural forces at any given these values and
RGKCȩ-LȩRFCȩMRFCPȩF?LB
ȩRCAFLMJMEGCQȩQF?NCȩFMUȩUCȩ
work, move, communicate and ultimately live. There-
fore, CRFGA?Jȧ?LBȧQMAG?JȧGKNJGA?RGMLQȧMDȧRCAFLMJMEWȧ
LCCBȧRMȧAMLQGBCPȧRFGQȧDSLB?KCLR?JȧGLRCPRUGLGLEȧMDȧ
RFCȧ FSK?Lȧ ?LBȧ RFCȧ RCAFLMJMEGA?Jȧ BMK?GLQ. ‘Hu-
K?LQȩ?PCȩRCAFLMJMEGA?Jȩ@CGLEQ
ȩHSQRȩ?QȩRCAFLMJMEGCQȩ
are social entities’ (Verbeek, 2011, p.4).
Calls for the development and use of AI to be
placed ‘within limits of our fundamental norms,
values, freedoms and human rights’ (EESC, 2017)
are coming from a variety of viewpoints, including
academia, civil society organisations, practitioners,
think-tanks and businesses. There are encouraging
signs of a commitment, or at least an interest in
guidelines into AI
systems.
collaborative work, between governments, industry
and ethicists to design technologies for common
EMMBȩ ?LBȩ GBCLRGDWȩ QNCAGȏAȩ NPGLAGNJCQȩ DMPȩ ?AAMSLR-
ability, transparency, safety and social good (Pau-
wels, 2018). In this respect, we should also mention
Floridi’s (2018) plea for the EU to use ‘soft’ digital
ethics25 as a good approach for AI, because it can
support regulation and legislation and balance dif-
ferent values and interests.
 6 Ethical and Social Perspective 56

In a concise format, we are presenting an overview

of key ethical and societal issues in AI/ML, as cur-
rently discussed in policy, academia, civil society
and business. The aim is to inform a comprehensive
framing of ethical and societal challenges, based
not only on fundamental values and rights, but also
MLȩUGBCPȩBGQASQQGMLQȩMLȩUFMȩ@CLCȏRQȩDPMKȩ'ȩBCTCJ-
opments, to what end, and under what conditions.

Drawing on Stahl, Timmermans and Flick (2017),

we make a distinction between issues that could
have an impact at individual level (such as auton-
omy; identity; dignity; privacy and data protection)
and those that have impact at societal level (such
as fairness and equity; collective identity and good
life; responsibility, accountability and transparency;
NPGT?AWȩGLȩAMLLCARGMLȩRMȩQSPTCGJJ?LACȩ?LBȩB?R?ȏA?-
tion; democracy and trust).

-TCPTGCUMDGLBGTGBS?J?LBAMJJCARGTC
 GKNJGA?RGMLQMD'

!F?JJCLECQ?RGLBGTGBS?JJCTCJ It is crucial
'Rȩ GQȩ GKNMPR?LRȩ RMȩ RFGLIȩ ?@MSRȩ FMUȩ ?PRGȏAG?Jȩ GLRCJJG-
gence might bring new challenges in relation to in-
to think how
BGTGBS?JȩFSK?Lȩ@CGLEQȩ'LȩRFGQȩAMLRCVR
ȩGRȩGQȩAPSAG?JȩRMȩ
consider how the concepts of autonomy and identity
the concepts
of individuals as well as security, safety and privacy
GQQSCQȩKGEFRȩAF?LECȩSLBCPȩRFCȩGLȐSCLACȩMDȩ'
of autonomy
?SRMLMKW
and identity
Autonomy is one of the central values in West-
ern ethics (Juth, 2009). It represents the idea that
of individuals
everyone has the capacity to know what is good or
bad for them, to live according to their standards as well as security,
and that they have power over their thoughts and
actions, thus promoting individual choice, rights safety and privacy
and freedoms. AI might contribute to human devel-
MNKCLRȩ@WȩCVRCLBGLEȩFSK?LȩA?N?@GJGRGCQ
ȩMPȩAMSJBȩ issues might change
XQGHUWKHLQÁXHQFH
stimulate productivity and prosperity and lead to
active work until a later age. It could also have the
MNNMQGRCȩCȎCARQ
ȩ?QȩBGQASQQCBȩGLȩ!F?NRCPȩ
of AI.
 57 6 Ethical and Societal Perspective

When it comes to interaction, there is also a need
for people to know if they are interacting with an-
other person or a machine. It is often not clear for
us today if we are communicating via digital media
with a bot or a human being. Also, the content pre-
2017), which is still a controversial proposal when
considering, for instance, that at the present time
accountability is ultimately related to human re-
sponsibility (EECS, 2016).
sented to us and selected by algorithms interferes B.PGT?AW?LBB?R?NPMRCARGML
with our freedom to receive and impart informa- In the world of hyper-connected devices, crucial
tion (Rathenau Institute, 2017). questions that are posed relate to privacy, security
and data protection. Smart meters, toys, fridges
@'BCLRGRW
ȩ RMȩ KCLRGMLȩ HSQRȩ ?ȩ DCU
ȩ ?JJȩ F?TCȩ @SGJR
According to Fearon (1999), identity is seen either
as a social category with its membership rules or
‘a set of attributes, beliefs, desires or principles of
action that a person thinks distinguish herself in
socially relevant ways and that the person takes a
special pride in’. In this case, the meaning of iden-
tity becomes very close to dignity or pride. As we
interact and even develop close relationships with
AI systems or robots, we need to understand better
possible transformations of how we perceive them
?LBȩ NFMLCQ
in AI systems that are sending our data to the
manufacturers of these devices, often without our
knowledge. AI applications in healthcare are even
more sensitive, for instance when suggesting diag-
nosis or treatments (AI Now, 2017) or when data
is (mis)used by insurance companies or technology
companies (Powles and Hodson, 2017). The right
to protection of personal information and privacy
is therefore increasingly recognised as crucial in
#SPMNC?LȩQMAGCRGCQȩ#%#
ȩȩ
and how they shape our conception of the world.
$MPȩGLQR?LAC
ȩNPMȏJGLE
ȩR?PECRCBȩ?BTCPRGQGLEȩMPȩMRFCPȩ 2FCȩ LCUȩ %".0ȩ RF?Rȩ A?KCȩ GLRMȩ DMPACȩ GLȩ +?Wȩ ȩ
'NMUCPCBȩRCAFLGOSCQȩA?LȩBCCNJWȩ?ȎCARȩMSPȩGBCL- GQȩP?GQGLEȩCVNCAR?RGMLQȩGLȩPCJ?RGMLȩRMȩRFCȩNPMRCARGMLȩ
tities (Floridi, 2015). The way in which human intel- of personal data used by algorithms. The full im-
JGECLACȩ?LBȩAMELGRGTCȩA?N?AGRGCQȩ?PCȩ?ȎCARCBȩ@WȩRFCȩ pact of the Regulation will only emerge in the years
interaction with machines and AI is an important RMȩAMKC
ȩ@SRȩGRȩGQȩAJC?PJWȩ?ȩK?HMPȩGQQSCȩRF?RȩLCCBQȩ
area of research, as indicated in Section 2.3. monitoring.

A"GELGRW
Human dignity is the right of a person to be re-
spected and valued and treated in an ethical man-
ner, as mentioned in the Article 1 of the Charter of
Fundamental Rights of the EU. It should be one of
the basic notions used for the protection from harm
caused to humans, and vulnerable groups in par-
!F?JJCLECQ?RQMAGCR?JJCTCJ
AI and big data have broader consequences for soci-
ety and the well-being of communities, and nations
worldwide. These changes have direct impact on re-
shaping of power relations and the transformation
ticular, and it should be respected while developing MDȩ RFCȩ QMAG?Jȩ AMLRP?ARȩ 'Rȩ GQȩ GKNMPR?LRȩ RMȩ CV?KGLCȩ
new technologies (Rathenau Institute, 2017). these changes in order to mitigate risks and dangers.

Individuals’ rights and responsibilities could start
eroding as a result of the increasing interaction of
humans and machines (EDPS, 2018). At the mo-
ment, smart devices have no moral responsibility
and that is why it could be potentially harmful to
JCRȩRFCKȩK?L?ECȩFSK?Lȩ@CGLEQȩ#%#
?$?GPLCQQ?LBCOSGRW
5GJJȩ 'ȩ BCAPC?QCȩ MPȩ GLAPC?QCȩ QMAG?Jȩ GLCOS?JGRGCQȩ 2Mȩ
address this question, we can distinguish two levels:
social inequalities generated by the application of
AI in the workplace, and individual level inequal-
ȩȩ&MU- ities as a result of unfair AI-driven decision-making.
ever, the European Parliament called for the EC With respect to the former, the evidence is as yet
RMȩ AMLQGBCPȩ ?ȩ QNCAGȏAȩ JCE?Jȩ QR?RSQȩ DMPȩ PM@MRQȩ #.
ȩ unclear, as discussed in Chapter 9. With respect to
 6 Ethical and Social Perspective 58

the latter, research has shown that the use of algo- CAMJMEGA?JȩAF?LECQȩRMȩMSPȩCVNCPGCLAC
ȩAPGRGA?JȩA?N?AG-
rithms could discriminate against particular groups ty and perceived reality (Borgmann, 1987; Feenberg,
MPȩ GLBGTGBS?JQ
ȩ DMPȩ GLQR?LACȩ GLȩ RFCȩ APGKGL?Jȩ HSQRGACȩ 1991, 1999; Mitcham, 1980; Coeckelbergh, 2018;
system (Chouldechova, 2017) and in recruitment Higgs et al., 2000). In this sense, it confronts us
NPMACQQCQȩ-,CGJJ
ȩȩ G?QȩAMSJBȩAMKCȩDPMKȩRFCȩ with who we are and how we want to live in our
training data or the algorithms, or from the biases society pursuing goals of well-being (Montréal
of the people who build them (The Future Society, Declaration 2017; IEEE, 2018).
2017; CNIL, 2017; see also Section 2.4). Besides
technical research (such as de-biasing or fairness
?AAMSLR?@GJGRW
evaluation tools), and education of developers and
society as a whole to be alert about potential un-
fairness, there is a need for the active involvement
of government in fostering the culture of equity,
free from stereotypes, and equal access ‘across
?LBȩ UGRFGLȩ QMAGCRGCQȩ #%#
ȩ ȩ RMȩ @CLCȏRȩ DPMKȩ ?@GJGRWȩ GQȩ AMLLCARCBȩ RMȩ PCQNMLQG@GJGRWȩ ?LBȩ GRȩ GQȩ CV-
the opportunities provided by AI.
A0CQNMLQG@GJGRW
?LBRP?LQN?PCLAW
3FTQPOTJCJMJUZ
BOE
BDDPVOUBCJMJUZ
We should think of responsibility in congruence
with human rights and values (CoE, 2017). Account-
tremely important because it deals with biases and
BGQAPGKGL?RGMLQȩA?SQCBȩ@WȩB?R?ȩKGLGLEȩ?LBȩNPMȏJGLEȩ
@!MJJCARGTCFSK?LGBCLRGRW?LBRFC~EMMBJGDC
AI impacts what we can consider the ‘good life’,
that is, it can deeply transform how we live today
ȩ
and in the future. Similar to other technologies,
AI can bring about social, economic, political and
Accountability is seen as a necessary condition for
the social acceptability of AI (Mission Villani, 2018).
CQGBCQȩRFCȩAMKNJCVGRWȩMDȩRFCȩMLJGLCȩCLTGPMLKCLR
locating responsibility and accountability becomes
KMPCȩBGȑASJRȩUGRFȩRFCȩ@JSPPCBȩPCJ?RGMLQFGNȩ@CRUCCLȩ
 59 6 Ethical and Social Perspective

people and data, action and consequences, which man beings, their behaviour, needs and aspirations
could be considered as a wicked problem without a (Markham et al., 2018, p.4), as well as control over
AJC?Pȩ?LQUCPȩ+?PIF?KȩCRȩ?J
ȩȩ'LȩRFCȩAMLRCVRȩ the users’ social behaviour and consumer prefer-
of AI, liability often follows responsibility. Current CLACQȩ ?I?PBHGCT?ȩ?LBȩ%CFJ
ȩȩ
laws and regulations struggle with the emerging
issues as discussed in Chapter 7.
5SBOTQBSFODZ
Transparency means not only transparency of algo-
PGRFKQȩMDRCLȩPCDCPPCBȩRMȩ?Qȩ~@J?AIȩ@MVCQ
ȩ@SRȩ?JQMȩ the use of tracking devices and corresponding plat-
of data and Automated Decision Making (ADM). The
importance of the latter is particularly salient when
Connected to mass surveillance is the appearance
of a phenomenon of voluntary self-surveillance that
E?GLCBȩNMNSJ?PGRWȩRFPMSEFȩRFCȩQMA?JJCBȩ/S?LRGȏCBȩ
Self movement. Self-surveillance is seen through
forms, with the goal of controlling one’s indicators
for the allegedly better management of life and
ADM is used by public agencies, such as predictive FC?JRFȩ 4CQLGAJSHCTGAȩ CRȩ ?J
ȩ ȩ 5FGJCȩ RFGLI-
NMJGAGLEȩ MPȩ PGQIȩ ?QQCQQKCLRȩ GLȩ RFCȩ APGKGL?Jȩ HSQRGACȩ
system (AI Now, 2018). The decisions made by al-
EMPGRFKQȩK?WȩLMRȩ@CȩSLBCPQRMMBȩMPȩCVNJ?GLCBȩ?LBȩ
it is not clear who is responsible for them. There-
fore, transparency is seen as key for accountabili-
ty, together with a right to meaningful information
GLȩ ?SRMK?RCBȩ BCAGQGMLQȩ ?Qȩ DMPCQCCLȩ @Wȩ RFCȩ %".0
ȩ
ing about controlling their lives, users are often
not aware of the fact that there are more parties
involved in the technology and that the data they
leave are not only available to themselves. A right
‘not to be measured, analysed or coached’ by AI is
needed (Rathenau Institute, 2017).
MPȩ CTCLȩ ?ȩ ~PGEFRȩ RMȩ CVNJ?L?RGMLȩ 1CJ@QRȩ ?LBȩ .MU-
les, 2017; Edwards and Veale, 2017; see Chapter 7).
Although it is not likely that the code will ever be
completely transparent because of IP rights, other
information such as variables used, with their val-
ues and deviations, and the amount and type of
RP?GLGLEȩB?R?ȩSQCBȩAMSJBȩ@CȩK?BCȩCVNJGAGRȩ?LBȩAML-
RPG@SRCȩRMȩ?Lȩ~CȎCARGTCȩRP?LQN?PCLAWȩ!M#
D"CKMAP?AW?LBRPSQR
While dealing with ethical and societal implications
of using algorithms, it is crucial to tackle their im-
pacts on democracy and the trust of citizens. The
CȎCARQȩ MDȩ LCUȩ NCPT?QGTCȩ RCAFLMJMEGCQȩ MLȩ NMUCPȩ
relationships between governments, citizens and
businesses are not straightforward. As we have
ȩ@ȩ discussed above, the lack of autonomy, privacy, re-
Besides transparency, there is a need to audit al-
gorithms, i.e. checking that they conform to certain
properties (Kroll et al., 2016). Auditing should be
done in collaboration with government and civil so-
ciety. However, lack of access to the training data is
the main problem for public auditing of algorithms
(Villani, 2018).
ȩ R?PECRGLEȩ
sponsibility, accountability, transparency and pos-
sibility for mass surveillance can impact deeply our
democratic systems. In addition to issues discussed
above, we should consider:
1PMJUJDBM
QSPҨMJOH
Big data may be used not only for commercial ad-
TCPRGQGLEȩ @SRȩ ?JQMȩ DMPȩ NMJGRGA?Jȩ NPMȏJGLE
?LBȩK?LGNSJ?RGMLȩȩPCACLRȩCV?KNJCȩGQȩ!?K@PGBECȩ
B1SPTCGJJ?LAC
B?R?ȍA?RGML Analytica and the way it used algorithms to target
Digital politics and media researchers argue that USA voters in the electoral campaign for the presi-
we live in a mass-surveillance culture, character- dential election in 2016. The accumulation of large
GQCBȩ@WȩB?R?ȏA?RGMLȩMDȩGLBGTGBS?JQȩJGTCQȩRF?RȩGQȩBGȑ- amounts of data in the hands of one company or
ASJRȩRMȩCQA?NCȩ4?Lȩ"GHAI
ȩȩ0?RFCL?Sȩ'LQRGRSRC
ȩ political party may have serious consequences for
ȩ)MPȎȩCRȩ?J
ȩȩ2FCȩUMPJBQȩ@GEECQRȩBGEGR?Jȩ the erosion of democracy (Brundage et al., 2018).
companies control a large part of the online public Social scoring systems based on AI optimisations
sphere and hold most personal data. Through the of social processes might be dangerous for privacy,
analysis of big data, they obtain ‘truths’ about hu- equality and freedom (Botsman, 2017).
 6 Ethical and Social Perspective 60

#PUT
EJTJOGPSNBUJPO
BOE
EFFQ
GBLFT and implementation (see Chapter 12). A societal

AI has been used as a political advertising tool to debate about the opportunities and challenges in
GLȐSCLACȩ TMRCPȩ @CF?TGMSPȩ @SRȩ ?JQMȩ RMȩ K?LGNSJ?RCȩ the use of algorithms in government is needed to
RFCȩNS@JGAȩ'LȐSCLAGLEȩNS@JGAȩMNGLGMLȩRFPMSEFȩ?JEM- direct future developments.
rithm-enabled persuasion and manipulation, the
use of political bots to spread false information
has become easier than ever before. Bots have de-
creased trust in online environments and have cre-
ated an advantage for populist politics that reduces
trust (Brundage et al., 2018). The so-called ‘deep
fakes’, false audio-visual content generated by AI,
are a novel challenge that we are likely to face in
the very near future (EPSC, 2018).
'SFFEPN
PG
FYQSFTTJPO
$PCCȩ ȐMUȩ MDȩ GLDMPK?RGMLȩ ?LBȩ RFCȩ PGEFRȩ RMȩ PCACGTCȩ
and impart information without interference are
guaranteed by Article 10 of the European Court of
Human Rights. These rights are being questioned
ȩ?LBȩAMLQGBCPȩUFCRFCPȩRFCȩQR?RCȩQFMSJBȩCVCPAGQCȩ
@CA?SQCȩMDȩRFCȩGKN?ARȩMDȩ?JEMPGRFKQȩ?LBȩNPMȏJGLEȩ
E!SKSJ?RGTCILMUJCBEC?Q?AMKKMLEMMB
In the longer term, there is also a need to start to
PCȐCARȩ?Rȩ?JJȩJCTCJQȩGLȩQMAGCRWȩMLȩRFCȩCVRCLRȩRMȩUFGAFȩ
the digital data that constitutes the cumulative
knowledge about society should also be consid-
ered a national asset, in a similar way to cultural
assets and landscapes. The Mission Villani alluded
to this issue in arguing for an AI strategy struc-
tured around the goals of sovereignty and strategic
autonomy (Mission Villani, 2018). In this respect, we
LCCBȩRMȩPCȐCARȩMLȩRFCȩLCCBȩRMȩUGBCLȩRFCȩTGCUȩMDȩBGEG-
tal data as a public good to also include a dimension
of preserving data about society for future genera-
RGMLQ
some degree of sovereignty over national data.
GLȩJGKGRGLEȩSQCPQȩAFMGACȩ'LȩRFGQȩAMLRCVR
ȩNJ?RDMPKQȩ
are becoming new information gatekeepers (Co-
lombo et al., 2017) and their responsibility in main-
R?GLGLEȩDPCCBMKȩMDȩCVNPCQQGMLȩLCCBQȩRMȩ@CȩAJ?PGȏCBȩ
(Rathenau Institute, 2017).
"MHPSJUINT
JO
HPWFSOBODF
The use of algorithms in governance has become
AI is of the utmost importance to help build co-
herence among the actions and policies to address
RFCȩ 1SQR?GL?@JCȩ "CTCJMNKCLRȩ %M?JQȩ 1"%Qȩ ?LBȩ
achieve the EU Agenda 2030. The use of big data
and AI is essential to identify synergies and trade-
MȎQȩ?KMLEȩRFCȩT?PGMSQȩQRP?RCEGCQȩDMPȩ?AFGCTGLEȩRFCȩ
individual goals, and develop a systemic approach
very popular among researchers, and policymakers.
Researchers are mostly interested in the perva-
QGTCLCQQȩ MDȩ ?JEMPGRFKGAȩ EMTCPL?LACȩ ?LBȩ GRQȩ CȎCARQȩ
on power, freedom and human rights. According to
Danaher (2017), ‘algocracy’ means using data min-
RMȩ1"%Qȩ2FCȩJCTCJȩMDȩBCKMAP?AWȩMDȩRFCȩBGEGR?JȩUMPJBȩ
?LBȩRFCȩU?WȩUCȩ?BBPCQQȩ'ȩ?ȎCARQȩRFCȩQSQR?GL?@GJ-
ity of our future.
ing and analytics to predict and control human be-
haviour. This is done through hyper-nudges: dynam-
ic and personalised algorithmic nudging that can
change and update based on the behaviour of their
intended target (Yeung, 2016). This could threaten
democratic societies if citizens lost agency and the
possibility to participate freely in public life.
At the same time, there may be opportunities to
leverage some of the characteristics of the current
digital transformation, including algorithms in gov-
ȩ~PCESJ?RMPWȩDMPCQGEFRȩ#%#
ȩȩ?LBȩ~CRF-
ernance to increase democratic accountability and
participation, and drastically change policy design
1SKK?PW?LBAMLAJSQGMLQ
In this chapter, we highlight some of the key di-
mensions of the ethical and societal debates on
AI. As shown, RFCȧBC@?RCȧ?@MSRȧCRFGA?Jȧ?LBȧQMAG?Jȧ
GKNJGA?RGMLQȧ MDȧ 'ȧ DMPȧ GLBGTGBS?JQȧ ?LBȧ QMAGCRGCQȧ
LCCBQȧ RMȧ KMTCȧ DMPU?PBȧ ?LBȧ D?QR. Political and
regulatory frameworks seem to be lagging behind
AI developments, which will soon pervade almost
every aspect of our daily lives. To counter this sit-
S?RGML
ics foresight analysis’ (Floridi (2018) at the same
level of ‘technological foresight’ (EDPS, 2018a), are
 61 6 Ethical and Social Perspective
needed to keep up with the emerging implications
of AI. This however cannot happen without a strong
commitment to the following issues:
$MFBS
FUIJDBM
GSBNFXPSLT
BOE
HVJEFMJOFT
There is a recognition in Europe at both EC, and na-
tional level that a clear ethical framework is need-
ed. This is a rapidly evolving area with guidelines
DPMKȩRFCȩ&GEF*CTCJȩ#VNCPRȩ%PMSNȩCVNCARCBȩ@WȩRFCȩ
end of 2018, as well as panels on ethics set up by
This ?NNPM?AFȧ RMȧ PCQNMLQG@JCȧ BCQGELȧ GQȧ LCACQ-
Q?PWȧ RMȧ @SGJBȧ RPSQRȧ GLȧ RFCȧ ?NNJGA?RGMLQ, but it is
not by itself enough. As introduced in Section 6.1,
technology is a social product which embodies and
replicates social norms, values, and other economic,
ecological, political, or cultural forces at any given
time. Therefore, it is crucial to develop a multi-
layered approach to AI design, including:
• 2FCȧAPGRGA?JȧCLE?ECKCLRȧMDȧAGTGJȧQMAGCRW in
L?RGML?JȩEMTCPLKCLRQȩ5Cȩ?PCȩ?JQMȩCVNCARGLEȩUMPIȩ the development and use of algorithms in
DPMKȩ RFCȩ -#!"
ȩ RFCȩ 3,
ȩ ?LBȩ NPMDCQQGML?Jȩ MPE?LG- ȩ BGȎCPCLRȩ ?NNJGA?RGMLȩ ?PC?Qȩ UGRFȩ BGQASQQGML
sations such as the IEEE to contribute to the de- ȩ ?@MSRȩUFMȩ@CLCȏRQ
ȩ?LBȩUF?RȩT?JSCQȩQFMSJB
bate. Then we will need to see how these proposed ȩ @CȩCLAMBCBȩGLRMȩRFCȩQWQRCKQȩ(?Q?LMȎ
ȩ
frameworks align and what gaps or inconsistencies, Boucher et al., 2014);
GDȩ?LW
ȩRFCWȩJC?TCȩRȩJC?QRȩGLȩ#SPMNC
ȩUCȩA?LȩCVNCARȩ
that RFCȧCRFGA?JȧDP?KCUMPIȧ?LBȧESGBCJGLCQȧQFMSJBȧ • 2FCȧ CQR?@JGQFKCLRȧ MDȧ KSJRGQR?ICFMJBCPȧ
@CȧAMKN?RG@JCȧUGRFȧRFCȧ#3ȧNPGLAGNJCQȧ?LBȧPCESJ?- fora to promote such public debate and trans-
RMPWȧDP?KCUMPIQȧ late their outcomes into socio-political strate-
gies for AI enforcing ethical and social values
In reviewing the issues, two new rights are suggested: (Cath et al., 2018; Stilgoe, 2017);

1) The right to a meaningful human contact. • 2FCȧBCQGELȧNP?ARGACȧMDȧAMLQGBCPGLEȧCRFGA?Jȧ

Every person should have the right to choose ȧ ?LBȧ QMAG?Jȧ T?JSCQ (Aldewereld et al., 2015;
human contact and refuse to be cared for by Van den Hoven et al., 2017; Nascimento et
a robot if he or she does not feel comforta- al., 2016), and potential sources of bias in the
ble with it (EP, 2016). At the same time, robots entire development process of the AI system:
should respect the autonomy of humans in de- from the selection of multidisciplinary and
cision-making. diverse design teams, to the selection and
labelling of training data, evaluation of
ȩ 2FCȩ PGEFRȩ RMȩ PCDSQCȩ @CGLEȩ NPMȏJCB
ȩ RP?AICB
ȩ outputs, and assessment of outcomes.
measured, analysed, coached or manipulated
(CoE, 2017; Rathenau Institute, 2017).

3FTQPOTJCMF
"*
EFTJHO
2FCȧ RP?LQJ?RGMLȧ MDȧ T?JSCQȧ ?LBȧ LMPKQȧ GLRMȧ RFCȧ
BCQGELȧ ?LBȧ MNCP?RGMLȧ MDȧ 'ȧ QWQRCKQȧ QFMSJBȧ @Cȧ
N?PRȧ MDȧ PCESJ?RMPWȧ DP?KCUMPIQȧ ?LBȧ ESGBCJGLCQ
For instance, ‘privacy by design’ or ‘data protec-
tion by design and by default’ are now part of the
%".0ȩ 1GKGJ?PJW
ȩ UCȩ QFMSJBȩ RFGLIȩ MDȩ BCQGELGLEȩ 'ȩ
applications that are ‘transparent, comprehensible,
monitorable and accountable by design’, backed
up by frameworks for auditing and evaluating with
agreed international standards.
 7 Legal perspective 62

5IF
(%13
IBT
CFFO
B
TPVSDF
PG
JOTQJSBUJPO
GPS
MFHJTMBUPST
PVUTJEF
&VSPQF

SUMMARY

In this chapter, we highlight the main legal issues raised by AI that concern, among others,
DSLB?KCLR?JȩPGEFRQ
ȩGLAJSBGLEȩNCPQML?JȩB?R?
ȩRFCȩRP?LQN?PCLAWȩ?LBȩCVNJ?GL?@GJGRWȩMDȩ?JEMPGRFKQ

?LBȩ RFCȩ QSGR?@GJGRWȩ MDȩ CVGQRGLEȩ JG?@GJGRWȩ PSJCQȩ 5Cȩ ?JQMȩ PCTGCUȩ RFCȩ #SPMNC?Lȩ JCE?Jȩ DP?KC
work with respect to data ownership, access and sharing regimes and intellectual property
PGEFRQ
ȩ?LBȩAMLQGBCPȩRFCȩBGȎCPCLRȩ?NNPM?AFCQȩRMȩPCESJ?RCȩBGQPSNRGTCȩRCAFLMJMEGCQȩ5CȩLMRCȩRFCȩ
RCLQGMLQȩ@CRUCCLȩNPMRCARGLEȩPGEFRQȩMDȩGLBGTGBS?JQȩ?LBȩȏPKQȩ?LBȩCLAMSP?EGLEȩGLLMT?RGMLȩUGRFȩ
RPWGLEȩRMȩK?VGKGQCȩMNCLLCQQȩ?LBȩRP?LQN?PCLAWȩ2FCȩR?QIȩMDȩNMJGAWK?IGLEȩGQȩRMȩK?L?ECȩAF?LECȩ
and regulate the adoption of new technologies to ensure that they are accepted to society
and respect our fundamental values. We conclude that Europe is well placed to establish a
BGQRGLARGTCȩDMPKȩMDȩ'ȩRF?RȩGQȩCRFGA?JJWȩPM@SQRȩ?LBȩNPMRCARQȩRFCȩPGEFRQȩMDȩGLBGTGBS?JQ
ȩȏPKQ
ȩ?LBȩ
QMAGCRWȩ?RȩJ?PECȩ$MPȩCV?KNJC
ȩRFCȩ%CLCP?Jȩ"?R?ȩ.PMRCARGMLȩ0CESJ?RGML
ȩMNNMQCBȩ@WȩK?LWȩUFCLȩ
in preparation, is now perceived as a European asset and is inspiring similar approaches out-
QGBCȩ#SPMNCȩ#VRCLBGLEȩRFGQȩLMRGML
ȩUCȩQFMSJBȩAMLQGBCPȩRFCȩFGEFȩQR?LB?PBQȩMDȩRFCȩ#SPMNC?Lȩ
legal and regulatory landscape in a similar way as those in environmental quality which are
an asset for Europeans and their future generations to build upon, not a barrier.
 63 7 Legal perspective

LEGAL PERSPECTIVE
+?GLJCE?JAF?JJCLECQGBCLRGȍCB
 GL#SPMNC?L'QRP?RCEGCQ The task of
As indicated in Chapter 4, several European Mem- policymaking
ber States, as well as the European Commission,
are considering the challenges and opportunities is to manage
?ȎMPBCBȩ@Wȩ'ȩ2FGQȩQCARGMLȩQSKK?PGQCQȩRFCȩK?GLȩ
legal challenges raised by widespread use of AI as change and regulate
GBCLRGȏCBȩGLȩ'ȩQRP?RCEGCQȩ?LBȩNMJGAWȩBMASKCLRQ26.
Section 7.2 and 7.3 develop more in detail data and the adoption of new
Intellectual Property (IP) related issues.
technology so that
-LCȩMDȩRFCȩKMQRȩAPGRGA?JȩGQQSCQȩGLȩRFCȩQRP?RCEGCQȩCV-
amined is how to ensure that AI algorithms operate they are acceptable
in full respect of the European values and funda-
mental rights27 as well as of its ethical principles to society and respect
(see Chapter 6). Key elements of concern are po-
tential forms of discrimination by algorithms, and our values.
the need to develop not only ethical guidelines but
also tools and design principles such as ‘ethics by into force, there is a need to monitor its applica-
BCQGELȩ-RFCPȩJGLCQȩMDȩ?ARGMLȩGLAJSBCȩK?LB?RGLEȩRFCȩ RGMLȩGLȩRFCȩ'ȩAMLRCVRȩ1MKCȩ?SRFMPQȩOSCQRGMLȩRFCȩ
assessment of impact on human rights or discrim- QSGR?@GJGRWȩMDȩRFCȩ%".0ȩRMȩBC?JȩUGRFȩ'ȩQCCȩ.MSJJCRȩ
GL?RGMLȩ UFCLȩ ?JEMPGRFKQȩ ?ȎCARȩ GLBGTGBS?JQȩ MPȩ ?PCȩ ?ȩ ?LBȩ @ȩ "CJDMPECȩ CRȩ %CP?PBȩ ȩ 5?RAFCPȩ
used in public administration, or require algorithm ȩMLȩRFCȩ@?QGQȩMDȩRFCȩBGȑASJRWȩRMȩAMKNJWȩUGRFȩ
?SBGRGLEȩKCAF?LGQKQȩ5GRFGLȩRFGQȩAMLRCVR
ȩRFCPCȩGQȩ the principles of purpose limitation (since AI sys-
?ȩ LCCBȩ RMȩ PCȐCARȩ MLȩ GDȩ ?LBȩ UFCLȩ ?SRMK?RCBȩ BC- tems are by nature general-purpose systems) and
cision-making should be allowed. Naturally, some proportionality, as well as the requirement to ob-
SQCPQȩ ?LBȩ QCARMPQȩ CEȩ HSQRGAC
ȩ FC?JRFȩ ?PCȩ KMPCȩ tain legitimate consent. Criticism also concerns the
sensitive than others in this regard. perceived limitations as regards the so-called right
RMȩCVNJ?L?RGMLȩ
In terms of QFSTPOBM
EBUB
QSPUFDUJPO
 certain pro-
TGQGMLQȩ GLȩ RFCȩ %CLCP?Jȩ "?R?ȩ .PMRCARGMLȩ 0CESJ?RGMLȩ 5GRFGLȩRFGQȩAMLRCVR
ȩUSBOTQBSFODZ and FYQMBJOBCJMJUZ
%".0ȩ #!
ȩ ?ȩ F?TCȩ ?BBPCQQCBȩ QMKCȩ MDȩ RFCȩ of algorithms are regarded as essential to under-
concerns mentioned above. These include: i) data standing how decisions are made and give citizens,
protection by design and by default, ii) privacy im- users and operators the possibility to challenge
N?ARȩ ?QQCQQKCLRQ
ȩ GGGȩ ?ȩ PCJ?RGTCȩ PGEFRȩ RMȩ CVNJ?L them. However, it is not yet clear how these re-
ation as concerns the logic involved in automated quirements should be addressed. Some referenc-
BCAGQGMLK?IGLE
ȩ?LBȩGTȩRFCȩPGEFRȩLMRȩRMȩ@CȩQS@HCARȩ es may be found in public documents concerning:
to a decision based solely on automated process- i) the processes to be put in place, (ii) the inter-
ing28. These provisions could serve as inspiration for ference with IP protection, iii) the responsible ac-
MRFCPȩ ?PC?Qȩ Qȩ RFCȩ %".0ȩ F?Qȩ MLJWȩ PCACLRJWȩ AMKCȩ tors to scrutinise algorithm decision making, and

GTȩ RFCȩ BCTCJMNKCLRȩ MDȩ ESGB?LACȩ RMȩ ?QQGQRȩ GLȩ CV-
plaining AI. The EC is analysing the issue and has
commissioned a study, AlgoAware29, to shed light
on these (and related) aspects.
-JBCJMJUZ is another issue frequently raised in the
legal literature and in policy debates. This is an es-
sential aspect to ensure that possible victims have
CȑAGCLRȩPCBPCQQȩQWQRCKQ
ȩ@SRȩ?JQMȩRF?RȩGLBSQRPWȩF?Qȩ criteria. This requires assigning quality levels for
7 Legal perspective 64
UIF
MJBCJMJUZ
BOE
TBGFUZ
GSBNFXPSLT Dor AI, the IoT
and robotics.33
AI products or services, like everything else, can-
not be 100% guaranteed to be safe. A possible ap-
NPM?AFȩ RMȩ K?L?ECȩ Q?DCRWȩ GQȩ RMȩ BCȏLCȩ ?AACNR?@JCȩ
JCTCJQȩMDȩQ?DCRWȩDMPȩGBCLRGȏCBȩPGQIQȩRMȩ@CȩAF?P?ARCP-
ised by the probability of occurrence and severity
JCE?Jȩ ACPR?GLRWȩ 2FCȩ AMKNJCVGRWȩ MDȩ 'ȩ QWQRCKQȩ ?LBȩ software and processes applicable from design to
the variety of actors intervening in the value chain, validation, according to the criticality of the soft-
coupled with its self-learning and autonomous abil- ware modules and corresponding datasets contrib-
GRGCQ
ȩA?LȩK?ICȩGRȩTCPWȩBGȑASJRȩRMȩ?JJMA?RCȩPCQNMLQG- SRGLEȩ RMȩ RFCȩ PGQIQȩ GBCLRGȏCBȩ BBGRGML?Jȩ KGRGE?RGMLȩ
bilities. Moreover, since the liability regime has been KC?QSPCQȩ MDȩ ?ȩ RCAFLGA?Jȩ CEȩ BGTCPQGȏCBȩ N?P?JJCJȩ
MLJWȩN?PRG?JJWȩF?PKMLGQCB
ȩRFGQȩQS@HCARȩGQȩ?JQMȩ?ȩK?R- developments) or operational nature (e.g. a cer-
ter of national law. tain level of human control) may need to be im-
plemented when residual risks are considered un-
The main question is therefore: does the current acceptable.
liability and safety framework provide adequate
mechanisms to deal with AI products and services In addition to the areas mentioned above, nation-
JG?@GJGRWȩ 'Dȩ LMR
ȩ UF?Rȩ ?BHSQRKCLRQȩ ?PCȩ LCACQQ?PWȩ al AI strategies also refer to the challenges posed
2FGQȩGQQSCȩGQȩ@CGLEȩCVNJMPCBȩ@WȩRFCȩ#VNCPRȩ%PMSNȩMLȩ by cybersecurity (see Chapter 10), and the need to
Liability and New Technologies, recently created by facilitate testing platforms, including regulatory
RFCȩ#!ȩQCCȩ!F?NRCPȩȩUFGAFȩGQȩCVNCARCBȩRMȩPCNMPRȩ Q?LB@MVCQ
ȩ RMȩ QSNNMPRȩ GLLMT?RGMLȩ ?Qȩ UCJJȩ ?Qȩ MRF-
in mid-2019. The group is split into two assemblies. er legal areas such as consumer protection and
2FCȩ ȏPQRȩ ?QQCK@JWȩ UGJJȩ DMASQȩ MLȩ RFCȩ .PMBSARȩ *G?- competition law. Both legal commentators and the
bility Directive30ȩ ?LBȩ UGJJȩ CV?KGLCȩ RMȩ UF?Rȩ CVRCLRȩ general press have recently warned of possible an-
the provisions of the Directive are adequate to ti-competitive behaviour that could take place in
solve the questions of liability in relation to tra-
ditional products, new technologies and new so-
ȩDMPȩCV?KNJC
ȩR?AGRȩAMJJSQGMLȩGLȩNPGACȩ
cietal challenges31. The second assembly, dealing
ȩNPGACȩBGQAPGKGL?RGML
ȩB?R?ȩAMLACLRP?RGMLȩ?LBȩ
with new technologies, ‘will assess whether and to
UF?Rȩ CVRCLRȩ CVGQRGLEȩ JG?@GJGRWȩ QAFCKCQȩ ?PCȩ ?B?NR-
ed to the emerging market realities following the
development of the new technologies such as Arti-
ȏAG?Jȩ'LRCJJGECLACȩ9x;ȩ?LBȩ?QQGQRȩRFCȩ!MKKGQQGMLȩGLȩ
developing EU-wide principles which can serve as
guidelines for possible adaptations of applicable
laws at EU and national level as regards the new
RFGQȩAMLRCVRȩ1SAFȩAMLACPLQȩPCJ?RCȩRMȩAMJJSQGMLȩNP?A-
RGACQȩGLAJSBGLE
ȏVGLE
to the role of privacy in competition assessment.
National public authorities and stakeholders have
recalled the need to remain vigilant in this regard
(UK House of Lords 2018). The increasing power of
certain platforms also raises concerns as regards
their accountability as quasi-public actors in the
digital society (EPSC 2018).
RCAFLMJMEGCQȩ9x;32 The regulatory regime for non-personal machine-
generated data (or sensor data) has been a recur-
WȩKGB
ȩRFCȩEPMSNȩGQȩCVNCARCBȩRMȩGQQSCȩ?ȩESGB- ring question in the policy debate concerning AI or,
ance document on the interpretation of the Product more generally, the industry 4.0 revolution. This
Liability Directive in light of technological develop- QNCAGȏAȩGQQSCȩGQȩCVNJMPCBȩDSPRFCPȩ@CJMU
ments and to publish a report on the broader im-
plications for potential gaps in and orientations for
 65 7 Legal perspective

. -ULCPQFGN
?AACQQ?LBB?R?QF?PGLE essential facilities doctrine as well as the criteria
developed to assess possible abuse of a dominant
The discussion on the ownership of data has been position in refusals to licence make traditional com-
CVRCLQGTCJWȩ BC@?RCBȩ GLȩ RFCȩ FGQRMPWȩ MDȩ GLDMPK?RGMLȩ petition law tests of limited use to promote access
technology law. The advent of the big data era RMȩB?R?ȩ"PCVJȩCRȩ?J
ȩ"PCVJȩ
ȩ.MSJJCRȩ@ȩ0C-
brought a new impetus to this discussion both for
@MRFȩ NCPQML?Jȩ ?LBȩ LMLNCPQML?Jȩ B?R?ȩ -Q@MPLCȩ
Clarke, 2016). This section will focus mainly on
the current discussion on the legal regime for non-
personal machine-generated data although in
ACPR?GLȩA?QCQȩRFCȩBGQRGLARGMLȩK?Wȩ@Cȩ?PRGȏAG?JȩQCCȩ
also Section 12.1).
The issue of the lack of a comprehensive legal
ACLRȩQRSBGCQȩNPCBGARȩRF?RȩAMLȐGARQȩ?PCȩJGICJWȩRMȩ?PGQCȩ
in the near future, but recognise that legal inter-
TCLRGMLȩ K?Wȩ @Cȩ NPCK?RSPCȩ ?Rȩ RFGQȩ QR?ECȩ -Q@MPLCȩ
Clarke, 2016 and in a more nuanced sense also
Deloitte, 2018).
The need to create a new intellectual property right
to incentivise data production and collection, com-
mercialisation, and access is something that has
framework for machine-generated data was raised
in the Communication on Building a European Data
ȩ 8CAFȩ ȩ ?LBȩ 4?Lȩ Q@PMCAIȩ CRȩ
Economy (EC 2017a). The EC, worried about the
SLBCPCVNJMGR?RGMLȩ ?LBȩ J?AIȩ MDȩ RP?BCȩ GLȩ K?AFGLCȩ
ȩUFGJCȩ"PCVJȩ?
ȩ
generated data, put forward a series of legislative
and non-legislative options for discussion, among
them the creation of a new data producer’s right.
However, in the later Communication Towards a
common European data space (EC 2018c), data
ownership is hardly mentioned. Instead, the EC
focuses on access to data and on fundamental
measures such as a proposal to review the PSI
@CCLȩ CVNJMPCBȩ @Wȩ QAFMJ?PQȩ UGRFȩ BGTCPEGLEȩ TGCUQȩ
$MPȩ CV?KNJC
al. (2017) favour the introduction of new owner-
QFGNȩPGEFRQȩDMPȩGLBSQRPG?JȩB?R?
Hugenholtz (2018) and Ramalho (2017a) are quite
critical of such an idea. In short, recent economic
research falls short of making a clear case for cre-
ating new IP rights for data (Kerber, 2016; Kerber &
Frank, 2017; Duch-Brown et al., 2017).
While discussions on ownership are still inconclu-
sive, the debate has shifted to the issue of data
Directive (EC, 2018d), a recommendation on ac-
ȩ BGȎCPCLRȩ MNRGMLQȩ ?PCȩ @CGLEȩ
ACQQȩ RMȩ ?LBȩ NPCQCPT?RGMLȩ MDȩ QAGCLRGȏAȩ GLDMPK?RGMLȩ
ȩGLAJSBGLEȩQNCAGȏAȩPCESJ?RGML
ȩAPC?RGLEȩ?ȩ
(EC, 2018i) and a guidance document on the busi-
LCQQRM@SQGLCQQȩ?LBȩ@SQGLCQQRMEMTCPLKCLRȩCV-
change (EC, 2018f). Recently, the EC has set up an
ȩ
CVNCPRȩ EPMSNȩ RMȩ ?QQGQRȩ GLȩ ?QQCQQGLEȩ GQQSCQȩ ?PMSLBȩ
business-to-government data sharing.
ȩ?ȩD?GPȩSQCȩMPȩ?LȩMNCLȩLMPKȩRF?RȩAMTCPȩ'ȩ
There is no legal or statutory title providing for own-
ership of data either by traditional property rights
ȩ%CGECPȩCRȩ
or by intellectual property since data per se is not
?AACQQȩ 'Lȩ RFGQȩ PCE?PB
AMLQGBCPCB
new RVBMJҨFE
right to access data, or a system of
compulsory licences to grant access to data, either
on a horizontal or sectoral basis34ȩ "PCVJȩ @
JIIP 2018; Mezzanotte, 2017; Weber, 2017). Some
QAFMJ?PQȩF?TCȩCTCLȩQSEECQRCBȩ?ȩQNCAGȏAȩAMNWPGEFRȩ
CVACNRGML
and other related uses without the authorisation
MDȩRFCȩAMNWPGEFRȩFMJBCPQȩ?KMLEȩMRFCPQ
al., 2018, and Schafer et al., 2015). Here we should
?ȩAMNWPGEFRȩMPȩN?RCLR?@JCȩQS@HCARȩK?RRCPȩRȩKMQR
ȩ recall that in 2016 the EC tabled a legislative pro-
under certain circumstances data could become
protected by the database’s TVJ
HFOFSJT right or by
the Trade Secrets Directive (EC, 2016c), provided
that certain conditions are met. In practice, provi-
sions on data access, use and sharing are regulated
at the contractual level on the basis of a de facto
ȩRFCȩNPMNMQ?JȩF?QȩLMRȩ@CCLȩCVCKNRCB
ownership. Focusing on competition law, some au-
thors point out that the strict requirements of the
posal that, among other things, proposed a man-
B?RMPWȩ CVACNRGMLȩ DMPȩ RCVRȩ ?LBȩ B?R?ȩ KGLGLEȩ 2"+ȩ
for reQC?PAFȩNSPNMQCQȩQS@HCARȩRMȩ?ȩQCPGCQȩMDȩAMLBG-
tions (Article 3 in EC 2016b).35 Although generally
welcomed by the research community and by copy-
PGEFRȩQAFMJ?PQ
from criticism in relation to the conditions under
which it should operate, being more restrictive than
 7 Legal perspective 66

RFCȩPCEGKCȩ?NNJGA?@JCȩGLȩMRFCPȩAMSLRPGCQȩ%CGECPȩCRȩ 2FCNPMRCARGMLMD'GLTCLRGMLQ

?J
ȩȩ1?KSCJQML
ȩȩJJȩGLȩ?JJ
ȩAMNWPGEFRȩCV-  APC?RGMLQ@WGLRCJJCARS?J
ceptions would only authorise the re-use but would
not, strictly speaking, facilitate access to datasets.
As indicated at the beginning of this section, some
scholars have noted that the distinction between
personal and non-personal data is becoming more
NPMNCPRWPGEFRQ
National AI strategies have so far paid little at-
tention to the challenges AI poses to the IPR legal
framework. However, the increasing importance of
'ȩRCAFLMJMEGCQȩ?LBȩRFCȩE?NQȩGBCLRGȏCBȩ@WȩQAFMJ?PQȩ
and more blurred. It is not only that non-personal
B?R?ȩ MPȩ ?LMLWKGQCBȩ B?R?ȩ A?Lȩ @Cȩ QS@HCARȩ RMȩ RCAF-
ȩ0?K?JFMȩȩ@ȩ
niques linking them back to individuals, but also
that recent reforms have made clear that the no-
tion of personal data includes data that allow in-
dividualisation, even if they do not reveal the indi-
vidual’s civil or legal identity36 (Poullet, 2018b). This
must be taken into account when considering spe-
GLȩ@MRFȩRFCȩAMNWPGEFRȩ?LBȩN?RCLRȩQWQRCKȩCEȩ%S?B-
?KSXȩȩ+GAF?SVȩ
Schönberger 2018 or WEF 2018) reveal the need
for further investigation.
As regards the patentability of AI, the announce-
KCLRȩ MDȩ RFCȩ #SPMNC?Lȩ .?RCLRȩ -ȑACȩ #.-ȩ MDȩ QNC-
AGȏAȩ ESGB?LACȩ MLȩ N?RCLR?@GJGRWȩ APGRCPG?ȩ DMPȩ 'ȩ GQ
AGȏAȩPCEGKCQȩDMPȩQCLQMPȩMPȩGLBSQRPG?JȩB?R?ȩRF?RȩK?Wȩ
also include data susceptible to being considered
as personal data.
The EU can
N?PRGASJ?PJWȩUCJAMKCBȩ2FCȩLCUȩQCARGMLȩMLȩPRGȏAG?Jȩ
Intelligence and Machine Learning provides some
AJ?PGȏA?RGMLȩMLȩUFCLȩ'PCJ?RCBȩQS@HCARȩK?RRCPȩF?Qȩ?ȩ
technical character.37 The protection of AI-generated
works or inventions seem to be more problematic. In
light of the humanist approach of copyright law, it is
questionable that AI-generated works deserve cop-
yright protection. As regards patent laws, although
a priori nothing prevents AI-generated inventions
from being patented, the assessment of the non-

provide guidance inventive step or the allocation of ownership may

also raise issues that are not entirely solved.
to the European While some copyright scholars clearly advocate for
countries AI-generated works to be placed in the public do-
main, others have put forward a series of propos-
on regulatory als aimed at ensuring a certain level of protection.
5GRFȩ LMR?@JCȩ CVACNRGMLQ
ȩ RFCQCȩ NPMNMQ?JQȩ ?PCȩ QRGJJȩ
issues of AI RMMȩT?ESCȩ2FCWȩBMȩLMRȩ?JU?WQȩQSȑAGCLRJWȩBCR?GJȩRFCȩ
possible elements underpinning such protection.
while fostering
There is no doubt that certain AI-generated cre-
the development ations/inventions may share the characteristics of
GLDMPK?RGMLȩ EMMBQȩ zȩ LMLCVAJSB?@JCȩ ?LBȩ LMLPG-
of a single market T?JPMSQȩL?RSPCȩzȩRF?RȩHSQRGDWȩRFCȩAPC?RGMLȩMDȩOS?QGȩ
monopolistic rights to foster innovation and com-
for new technologies. mercialisation. However, there are concerns whether
incentives are needed, especially in cases where
the investment cost is low, and what consequences
such rights might have on the market, including on
67 7 Legal perspective
 7 Legal perspective 68

creations or inventions made by humans. Would

more property rights encourage or rather deter in-
LMT?RGMLȩ5CȩAJC?PJWȩLCCBȩRMȩGLTCQRGE?RCȩRFCQCȩGQ-
sues further from a law and economics approach
before favouring one solution or another. Every
0CESJ?RMPW?NNPM?AF disruptive
Every disruptive technology induces changes with
an impact on society. Regulation is often consid-
technology
ered necessary to address some of these impacts.
2FCȩ ȏPQRȩ BGȑASJRWȩ UFCLȩ BCAGBGLEȩ ?@MSRȩ RFCȩ KMQRȩ
induces changes
appropriate regulatory framework for AI is to an-
ticipate the impact it may have on society. The sec- with an impact
MLBȩGQȩRMȩAMNCȩUGRFȩRFCȩN?ACȩMDȩ'ȩ?BT?LACQȩ-RFCPȩ
AF?JJCLECQȩGLAJSBCȩRFCȩJCTCJȩMDȩAMKNJCVGRWȩMDȩ'
ȩGRQȩ
autonomous and self-learning features and its per-
on society.
vasiveness across sectors. These features require
@MRFȩ OS?JGȏCBȩ GLRCPBGQAGNJGL?PWȩ A?N?AGRWȩ RMȩ @CRRCPȩ
understand technologies and predict impact) and
ȐCVG@GJGRWȩ RMȩ CLTGQ?ECȩ DP?KCUMPIQȩ RF?R
ȩ UFGJCȩ CL- PCESJ?RMPWȩP?ACȩRMȩRFCȩ@MRRMKȩ%MJBD?P@ȩ?LBȩ2PCȐCP
ȩ
forceable, can adapt with time) in policymaking. 2018a). Several actors have called for a global dia-
2FCȩGKNJCKCLR?RGMLȩMDȩPCESJ?RMPWȩQ?LB@MVCQȩK?Wȩ logue on the establishment of norms and ethical
be particularly useful in this respects. rules or principles (e.g. UK House of Lords, 2017, or
Mission Villani, 2018).
While a hard-law approach may be needed in cer-
tain instances, co-regulation and soft-law schemes Considering the various issues at stake, and spe-
?PCȩMDRCLȩNPCDCPPCBȩGLȩMPBCPȩLMRȩRMȩQRGȐCȩGLLMT?RGML AGȏA?JJWȩ RFCGPȩ CRFGA?Jȩ BGKCLQGML
ȩ GRȩ @CAMKCQȩ APSAG?J
ȩ
zȩCTCLȩGDȩRFCPCȩK?Wȩ@CȩA?QCQȩUFCPCȩPCESJ?RGMLȩ?JQMȩ whichever the approach adopted, to ensure stake-
acts as an enabler for innovation. The role of cer- holder involvement and, in particular, the engage-
RGȏA?RGMLȩ ?LBȩ QR?LB?PBGQ?RGMLȩ @MBGCQ
ȩ ?Qȩ UCJJȩ ?Qȩ ment of civil society in the law-making process.
that of other organisations representing stakehold- Inclusive and/or innovative platforms are being cre-
er interests (notably citizens and consumers), are ated in an attempt to respond to these challenges.
considered important. There is also a discussion on Thus, as already noted, the EC has recently estab-
the approach of ‘regulation by design’, the need to lished the European AI Alliance to engage in a broad
act in an anticipatory vs. reactive way, innovation or and open discussion of all aspects of AI development
precautionary principles, or to have either sectoral and its impact on the economy and society. The UK
or horizontal regulation. While some institutions GQȩ ~CQR?@JGQFGLEȩ ?ȩ +GLGQRCPG?Jȩ 5MPIGLEȩ %PMSNȩ MLȩ $S-
have proposed the adoption of regulation address- ture Regulation to scan the horizon and identify the
ing part of the challenges highlighted in this section areas where regulation needs to adapt to support
(EP 2017), most of them remain rather cautious. emerging technologies such as AI’, and a ‘Regula-
tors’ Pioneer Fund to support regulators to develop
'Lȩ?ȩEJM@?JȩAMLRCVR
ȩEMTCPL?LACȩKMBCJQȩQFMSJBȩ@Cȩ new approaches which enable emerging technolo-
CVNJMPCBȩGLȩMPBCPȩRMȩCLQSPCȩRF?RȩPCESJ?RMPWȩCȎMPRQȩ gies’.38 Also in the UK, a recent report has proposed
are not isolated and to avoid regulatory patchworks RFCȩAPC?RGMLȩMDȩ?ȩLCUȩ@MBW
ȩRFCȩ-ȑACȩDMPȩ0CQNMLQG-
JC?BGLEȩ RMȩ DMPSKȩ QFMNNGLEȩ #%#ȩ ȩ MPȩ RMȩ ?ȩ ble Technology, with three main functions: empower
 69 7 Legal perspective

regulators, inform the public and policymakers and world and informing good practices for the use of
QSNNMPRȩNCMNJCȩRMȩȏLBȩPCBPCQQȩ+GJJCPȩCRȩ?J
ȩ personal data even beyond EU data markets. The
European policymaker can have a role in the setting
1SKK?PW?LBAMLAJSQGML of standards for human-centred AI, as has already
been done in relation to privacy.
'ȩ GQȩ KMQRȩ JGICJWȩ EMGLEȩ RMȩ F?TCȩ QGELGȏA?LRȩ GKN?ARQȩ
on our lives, even more than past waves of digit- Access and use of quality data are fundamental for
?JGQ?RGMLȩ Qȩ UGRFȩ ?LWȩ K?HMPȩ RCAFLMJMEGA?Jȩ AF?LEC
ȩ many AI applications. However, empirical evidence
RFGQȩ UGJJȩ @PGLEȩ QGELGȏA?LRȩ QMAGCR?Jȩ @CLCȏRQ
ȩ @SRȩ UGJJȩ shows that an optimal framework for trading and
also pose new challenges across many economic sharing of data has still to be realised (see Sec-
segments, including transport, energy, healthcare, RGMLȩ ȩ 5GRFGLȩ RFGQȩ AMLRCVR
ȩ QAFMJ?PQȩ F?TCȩ NSRȩ
BCDCLAC
ȩ CBSA?RGML
ȩ JCGQSPC
ȩ D?PKGLEȩ ?LBȩ ȏL?LACȩ DMPU?PBȩ?ȩQCPGCQȩMDȩNPMNMQ?JQȩRMȩ?BHSQRȩRFCȩPCESJ?-
The task of policymaking is to manage change and
regulate the adoption of new technology in order to
ensure that they are acceptable to society and re-
spect our fundamental values. The EU can provide
guidance to the European countries on regulatory
issues of AI while fostering the development of a
single market for new technologies.
To foster the development and uptake of AI, citizens
and industry operators need to be reassured that
it complies with ethical and regulatory frameworks.
tory framework for non-personal machine-gener-
ated data. These go from enacting new ownership
rights on data, to establishing access regimes or
incentivising data sharing through soft-law instru-
ments. Sectoral approaches and solutions focused
on regulating access (rather than on privileging the
creation of new rights) are gaining prominence. In a
AMLRCVRȩUFCPCȩRFCȩ#SPMNC?LȩB?R?ȩJ?LBQA?NCȩGQȩBC-
centralised and fragmented (in comparison with that
of other regions), the discussion on data governance
becomes even more relevant and opens opportuni-
!MLQSKCPQȩ ?LBȩ SQCPQȩ MDȩ 'ȩ BCTGACQȩ QFMSJBȩ @CLCȏRȩ
from mechanisms allowing redress in case of dam-
ages and from tools that allow them to supervise
decisions taken by AI systems. Economic actors also
LCCBȩ QSȑAGCLRȩ JCE?Jȩ ACPR?GLRWȩ RMȩ GLTCQRȩ GLȩ 'ȩ RCAF-
nology. The initiatives adopted both at the European
and national levels provide some hints on how these
aspects may be addressed.
The future EC guidelines on the ethical develop-
KCLRȩ MDȩ 'ȩ ?LBȩ RFCȩ ȏLBGLEQȩ MDȩ MLEMGLEȩ QRSBGCQȩ MLȩ
RP?LQN?PCLAWȩ ?LBȩ CVNJ?GL?@GJGRW
ȩ ?JMLEȩ UGRFȩ N?P?JJCJȩ *?QRȩ@SRȩLMRȩJC?QR
ȩUCȩF?TCȩ@PGCȐWȩPCDCPPCBȩRMȩRFCȩ
initiatives undertaken by others stakeholders, may
lay the foundations for a fairer AI. The work of the
#VNCPRȩ %PMSNȩ MLȩ *G?@GJGRWȩ ?LBȩ ,CUȩ 2CAFLMJMEGCQȩ GQȩ
?JQMȩCVNCARCBȩRMȩ@PGLEȩAJ?PGRWȩMLȩRFCȩKMQRȩQSGR?@JCȩ
rules to cope with unanticipated damages caused by
AI systems.
RGCQȩDMPȩBGȎCPCLRȩKMBCJQȩRMȩCKCPECȩ1CCȩȩ
In addition, the potential capacity of AI systems to
generate inventions or creations has attracted the
attention of scholars, some of them sceptical about
RFCȩA?N?AGRWȩMDȩRFCȩCVGQRGLEȩ'.ȩDP?KCUMPIȩRMȩ?AAMK-
modate this phenomenon. Further economic and
JCE?Jȩ PCQC?PAFȩ GQȩ LCCBCBȩ RMȩ ?QQCQQȩ RMȩ UF?Rȩ CVRCLRȩ
?BHSQRKCLRQȩRMȩRFCȩJCE?JȩDP?KCUMPIȩMPȩRFCȩAPC?RGMLȩ
of new rights are needed at all.
needs and tensions in the general debate on how to
better regulate AI. In the discussion around regulat-
ing AI, RFCȧOSCQRGMLȧGQȧLMRȧMLJWȧGDȧRFCPCȧGQȧ?ȧLCCBȧ
DMPȧPCESJ?RMPWȧGLRCPTCLRGMLȧ@SRȧK?GLJWȧUFGAFȧ?N-
NPM?AFȧ QFMSJBȧ @Cȧ ?BMNRCBȧ RMȧ ȏLBȧ ?ȧ @?J?LACȧ RMȧ
QSNNMPRȧGLLMT?RGMLȧ?LBȧ?RȧRFCȧQ?KCȧRGKCȧNPMRCARȧ
UGBCPȧQMAGCR?JȧGLRCPCQRQȧThe question is also how
Compared with the other AI champions, such as the to ensure that policymakers are well equipped to
USA and particularly China, Europe is noted for its make adequate and timely decisions in a rapidly
high regulatory standards for personal data pro- changing, technology- and market-driven environ-
RCARGMLȩ2FCȩ%".0ȩGQȩGLQNGPGLEȩPCESJ?RGMLȩ?APMQQȩRFCȩ ment with an intrinsic global dimension.
 8 Educational perspective 70

"*
XJMM
DIBOHF
XIBU
BOE
IPX
XF
MFBSO

SUMMARY

From an educational perspective, we observe in this chapter that AI has potential positive
impacts on shortages of skills, learning, and teaching. Three crucial points stand out from
RFCȩPCTGCUȩȏPQRJW
ȩRFCȩLCCBȩRMȩSLBCPQR?LBȩ@CRRCPȩFMUȩRFCȩGLRCP?ARGMLȩUGRFȩ'ȩGKN?ARQȩFSK?L
intelligence in cognitive capacities in both adults, and even more importantly, children.
Secondly, we need to think beyond current needs and practices, and consider how AI is likely
to change the relationship between education and work, and human development. Thirdly, we
highlight possible risks related to AI in education, particularly privacy and ethical ones. In this
AF?NRCP
ȩUCȩ?JQMȩNPMTGBCȩ?LȩGLGRG?JȩMTCPTGCUȩ?LBȩECMEP?NFGAȩBGQRPG@SRGMLȩMDȩRFCȩ?A?BCKGAȩMȎCPȩ
of study topics related to AI. This is only the beginning of a monitoring and analysis process
MDȩ?A?BCKGAȩQSNNJWȩ@SRȩGRȩ?JPC?BWȩFGEFJGEFRQȩNMRCLRG?JȩNPGMPGRWȩ?PC?QȩUFCPCȩEPC?RCPȩCȎMPRȩGQȩ
LCCBCBȩRMȩNPCN?PCȩQMAGCRWȩRMȩDSJJWȩF?PLCQQȩRFCȩ@CLCȏRQȩMDȩ'ȩ
 71 8 Educational perspective

EDUCATIONAL
PERSPECTIVE
'LRPMBSARGML
Education and
Since the beginning of the 1980s, and until re-
cently, educational applications of AI have mainly training are crucial
focused on the knowledge-based approach (see
Sleeman and Brown, 1982). The most prominent to harness AI, but
line of research has been concerned with Intelli-
gent Tutoring Systems, or ITS (Woolf 2009). These AI can also help
systems have been successful mainly in relatively
limited and unambiguous domains, such as mathe- us rethink what
K?RGAQȩ?LBȩNFWQGAQȩ0GRRCPȩCRȩ?J
ȩȩ%P?CQQCPȩCRȩ
al., 2005). competences
As student testing plays an important role in many and skills will be
CBSA?RGML?Jȩ QWQRCKQ
ȩ LSKCPMSQȩ NPMHCARQȩ RPWȩ RMȩ
use AI for automatic test generation and assess- needed in the future
ment. AI is also used to diagnose student attention,
emotion, and conversation dynamics in comput- QLIFSB>CRIȭIIFKD
CPQSNNMPRCBȩ JC?PLGLEȩ CLTGPMLKCLRQ
ȩ DMPȩ CV?KNJCȩ
for course development and management, in an life, and master
attempt to generate optimal groups for collabora-
tive learning tasks, and to recognise patterns that the digital
predict student drop-out (Nkambou et al., 2018;
0MQĸȩCRȩ?J
ȩȩ2MȩBMȩRFGQȩCȎCARGTCJW
ȩJ?PECȩB?R?- transformation
sets are needed for training the systems, and the
student behaviour needs to be actively monitored of our society.
to provide feedback for personalised learning. This
creates technical needs to monitor students unob- QA?LQȩDMPȩQRSBCLRQȩ?RȩPGQI
ȩ?LBȩBCRCARQȩBWQJCVG?ȩ@Wȩ
RPSQGTCJW
ȩ DMPȩ CV?KNJCȩ SQGLEȩ TGBCMȩ NPMACQQGLEȩ ?LBȩ tracking reader eye movements (Jakobbson, 2017).
remote eye-tracking, with associated ethical and AI-based systems have also been successfully de-
regulatory challenges. veloped for the diagnosis of autism spectrum dis-
MPBCPȩ?LBȩ?RRCLRGMLBCȏAGRȩBGQMPBCPQ
In special needs education, AI-based approaches
have shown potential, for instance, in the early de- This section will take a closer look at some of these
RCARGMLȩMDȩBWQJCVG?ȩ"PGE?Qȩ?LBȩ'M?LLGBMS
ȩȩȩ challenges related the impact of AI on skills de-
UCJJNS@JGQFCBȩ CV?KNJCȩ GQȩ RFCȩ 1UCBGQFȩ AMKN?LWȩ mand, and on learning and teaching.
*CVNJMPCȩRF?RȩF?QȩBCTCJMNCBȩ?ȩQWQRCKȩRF?RȩOSGAIJWȩ
 8 Educational perspective 72

'GKN?ARMLQIGJJQBCK?LB
JC?PLGLE to be massive needs for high-level AI competences.
 ?LBRC?AFGLE Another issue is the likely impact of AI on changing
QIGJJQȩBCK?LBQȩDMPȩRFCȩCVGQRGLEȩUMPIDMPACȩ?LBȩRFCȩ
"GPCAR'GKN?ARML?BT?LACBQIGJJQ population at large. Some of these are mentioned
 BCK?LB GLȩRFCȩLCVRȩQCARGMLȩMLȩRFCȩGKN?ARȩMDȩ'ȩMLȩJC?PLGLEȩ
'Lȩ1CARGMLȩ
ȩUCȩPCNMPRȩMLȩ?ȩȏPQRȩQA?LȩMDȩRFCȩRWNCȩ
The development of new AI and ML models requires and distribution of AI-related study topics in Euro-
very high levels of competence in several areas. This pean countries.
GQȩMLCȩMDȩRFCȩPC?QMLQȩUFWȩ'ȩCVNCPRQȩ?PCȩLMUȩ@CGLEȩ
N?GBȩTCPWȩFGEFȩQ?J?PGCQȩ2FCȩLSK@CPȩMDȩ'ȩLCSP?JȩCV-  'KN?ARMD'MLJC?PLGLE
perts may be doubling annually, but the basic knowl-
edge needed for state-of-the-art work in this area The way in which human intelligence and cogni-
PCOSGPCQȩ?BT?LACBȩJCTCJQȩMDȩQAGCLRGȏA
ȩK?RFCK?RGA?Jȩ RGTCȩA?N?AGRGCQȩ?PCȩ?ȎCARCBȩ@WȩRFCȩGLRCP?ARGMLȩUGRFȩ
and technical skills that are not easy to acquire. De- machines and AI is an important area of research.
velopment of new AI methods requires a good un- 5FGJQRȩ PCACLRȩ QAGCLRGȏAȩ JGRCP?RSPCȩ F?Qȩ DMASQCBȩ MLȩ
BCPQR?LBGLEȩMDȩQR?RGQRGAQ
ȩJGLC?Pȩ?JEC@P?
ȩBGȎCPCLRG?Jȩ the interactions between AI systems and adults,
equations, as well as computer architectures and RFCPCȩ ?PCȩ GKNMPR?LRȩ BGȎCPCLACQȩ GLȩ RFCȩ U?Wȩ AFGJ-
esoteric39 programming approaches and tools. The BPCLȩ BC?Jȩ UGRFȩ ?PRGȏAG?Jȩ QWQRCKQȩ RF?Rȩ LCCBȩ RMȩ @Cȩ
required skill set is scarce, and estimates of the ac- researched further.
RS?JȩLSK@CPȩMDȩNCMNJCȩUGRFȩ'ȩQIGJJQȩT?PWȩQGELGȏA?LRJW
ȩ
ranging from tens of thousands to a hundred thou- 'LȩECLCP?JȩRCPKQ
ȩ'ȩA?Lȩ@CȩSQCBȩGLȩRFPCCȩBGȎCPCLRȩ
sand. It is estimated that there are around 22 000 U?WQȩRF?RȩK?WȩF?TCȩBGȎCPCLRȩGKNJGA?RGMLQȩDMPȩRFCȩ
PhD researchers in AI and some 5 000 people who development of human cognitive capabilities both
have written academic articles or presented at AI in children and adults.
conferences in recent years.40
$GPQR
ȧ 'ȧ A?Lȧ QSNNMPRȧ CVGQRGLEȧ A?N?@GJGRGCQ. When
'RȩGQȩCVNCARCBȩRF?RȩRFCȩFGEFȩTGQG@GJGRWȩMDȩ'ȩ?LBȩRFCȩ competences are understood as combinations of
current demand will relatively rapidly direct talent BMK?GLQNCAGȏAȩ CVNCPRGQCȩ ?LBȩ @CF?TGMSP?Jȩ PCNCP-
RMȩRFGQȩ?PC?ȩQȩ?LȩCV?KNJC
ȩQGLACȩGRQȩJ?SLAFȩGLȩ+?Wȩ toires42, AI can now reduce the need for domain spe-
2018, about 90 000 students from over 80 coun- AGȏAȩILMUJCBECȩ?LBȩRFCPCDMPCȩK?ICȩRP?LQTCPQ?Jȩ?LBȩ
RPGCQȩF?TCȩCLPMJJCBȩGLȩRFCȩQGVUCCIȩ#JCKCLRQȩMDȩ'ȩ domain-independent generic skills more important.
zAMSPQCȩMPE?LGQCBȩ?QȩN?PRȩMDȩRFCȩ'ȩ#BSA?RGMLȩNPM-
gramme at the Finnish Centre of AI.41 Due to the 1CAMLB
ȧ 'ȧ A?Lȧ QNCCBȧ SNȧ AMELGRGTCȧ BCTCJMNKCLRȧ
FGEFȩU?ECȩBGȎCPCLRG?JQ
ȩASPPCLRȩQRSBCLRQȩGLȩQR?RGQ- ?LBȧAPC?RCȧAMELGRGTCȧA?N?@GJGRGCQȧRF?RȧUMSJBȧLMRȧ@Cȧ
tics, mathematics, mathematical physics, computer NMQQG@JCȧUGRFMSRȧRCAFLMJMEW. The mechanisation of
and chip design, and perhaps neurophysiology may human work has made possible things that would be
PCAMLQGBCPȩRFCGPȩA?PCCPȩN?RFQȩ?LBȩȏLBȩLCUȩMNNMPRS- impossible without technology; similarly, the mech-
LGRGCQȩ?QȩCVNCPRQȩGLȩ'ȩ+MPCMTCP
ȩFGEFJCTCJȩ'ȩAMK- anisation of cognitive work makes possible new ac-
NCRCLACQȩK?Wȩ?JQMȩCKCPECȩDPMKȩSLCVNCARCBȩNJ?ACQ
tivities that have not been possible before. This has
DMPȩCV?KNJC
ȩRFPMSEFȩMNCLȩQMDRU?PCȩ?LBȩMNCLȩF?PB already happened. It would be impossible to design
ware communities. a modern microprocessor or a neural chip without
AMKNSRCP?GBCBȩBCQGELȩRMMJQȩRF?RȩSQCȩCVRCLQGTCȩ@MB-
-LCȩP?RFCPȩGKKCBG?RCȩPCQSJRȩMDȩRFGQȩQGRS?RGMLȩGQȩRF?Rȩ ies of design knowledge.
high-level AI talent could probably be provided as a
service, similarly to the Infrastructure of software 2FGPB
ȧ'ȧK?WȧPCBSACȧRFCȧGKNMPR?LACȧMDȧQMKCȧFS-
as a service. This may mean that there is not going K?LȧAMELGRGTCȧA?N?@GJGRGCQ
ȧMPȧK?ICȧRFCKȧM@QMJCRC.
 73 8 Educational perspective

$MPȩCV?KNJC
ȩ?Qȩ'ȩA?LȩAMLTCPRȩQNCCAFȩRMȩRCVRȩ?LBȩ CRFGA?Jȩ GKNJGA?RGMLQȩ 5FCL
ȩ DMPȩ CV?KNJC
ȩ QRSBCLRQȩ
vice versa, and do mathematical calculations, dys- and their achievements are assessed using such AI
JCVG?ȩ MPȩ BWQA?JASJG?ȩ K?Wȩ @CAMKCȩ QMAG?JJWȩ JCQQȩ GK- systems, the assessment is necessarily based on
portant than it has been in the past. This has clear APGRCPG?ȩRF?RȩPCȐCARȩASJRSP?Jȩ@G?QCQȩ?LBȩFGQRMPGA?JJWȩ
@CLCȏRQȩ DMPȩ GLBGTGBS?JQ
ȩ @SRȩ RFCȩ MTCP?JJȩ GKN?ARȩ MDȩ salient measures of success. Supervised learning
making skills redundant through AI is not easy to algorithms create unavoidable biases, and these
predict. From a pedagogic point of view, it may be ?PCȩASPPCLRJWȩ@CGLEȩBC@?RCBȩCVRCLQGTCJWȩ
KMPCȩ@CLCȏAG?JȩRMȩSQCȩ'ȩRMȩFCJNȩNCMNJCȩRMȩBCTCJMNȩ
AMKNCRCLACQȩ RF?Rȩ ?JJMUȩ RFCKȩ RMȩ MTCPAMKCȩ BGȑ- The rapid advances in NLP and AI-based hu-
culties in reading and counting, instead of using AI man-machine interfaces will also generate new
to make redundant skills that underpin important
ȩJC?PLGLEȩ@Wȩ
cognitive capabilities.
It is also often assumed that AI systems enable
new levels of personalisation and diversity for in-
formation systems. Much of this personalisation,
FMUCTCP
ȩ PCQSJRQȩ DPMKȩ ȏLCEP?GLCBȩ A?RCEMPGQ?RGMLȩ
NCB?EMEGA?JȩNMQQG@GJGRGCQȩ$MPȩCV?KNJC
teaching machines shows clear potential, while
real-time machine translation also opens up new
possibilities in language learning.
5FGJQRȩGRȩGQȩNMQQG@JCȩRMȩGK?EGLCȩK?LWȩCVAGRGLEȩNMQ-
sibilities for AI in teaching, UGRFMSRȧ AJC?Pȧ NMJGAGCQ
RF?RȩNSRQȩSQCPQȩGLRMȩNPCBCȏLCBȩAJ?QQCQȩJRFMSEFȩ that put emerging technical possibilities in the
RFCQCȩQWQRCKQȩK?Wȩ@Cȩ?@JCȩRMȩCȑAGCLRJWȩQGKSJ?RCȩ @PM?BCPȩ AMLRCVRȩ MDȩ RFCȩ RP?LQDMPK?RGMLȩ MDȩ CBSA?-
personalisation, they do not necessarily support RGMLȧ ?LBȧ RFCȧ DSRSPCȧ MDȧ JC?PLGLE
ȧ CBSA?RGML?Jȧ 'ȧ
BCCNCPȩJCTCJQȩMDȩBGTCPQGRWȩ'ȩQWQRCKQȩA?Lȩ@CȩCVACJ-
lent predictive machines, but this strength may be
an important weakness in domains where learning
and development are important.43
UGJJȧNPM@?@JWȧK?GLJWȧ@CȧNPMTGBCBȧ?QȧQMJSRGMLQȧRMȧ
CVGQRGLEȧNPM@JCKQ. Instead of renewing the system
and orienting it towards the needs of a post-indus-
trial economy and knowledge society, AI may there-
fore mechanise and reinvent outdated teaching
 'KN?ARMD'MLRC?AFGLE

There are some clear opportunities for AI in teach-

ing, such as student assessment in its various
forms, and personalised tutoring systems. These
systems use a knowledge-based architecture with
Educational
a EPNBJO
NPEFM
that describes the area to be
learned and a TUVEFOU
NPEFM that describes the cur-
AI will
rent state of the student’s knowledge and learning.
Lȩ CVNCPRȩ QWQRCKȩ MPȩ NCB?EMEGA?Jȩ KMBCJȩ K?L?ECQȩ probably
the introduction of learning materials to the student
through an adaptive and interactive user inter- mainly
face. As student behaviour and learning can also
be monitored in ITS environments in great detail,
intelligent tutoring environments have also been an
be provided
important source of data for research into learning
(Porayska-Pomsta 2015).
as solutions
As these supervised AI learning algorithms are
to existing
based on historical data, they can only see the
world as a repetition of the past. This has deep
problems.

NP?ARGACQȩ ?LBȩ K?ICȩ RFCKȩ GLAPC?QGLEJWȩ BGȑASJRȩ RMȩ
change. It is therefore crucial to look beyond the
current practices and ask fundamental questions
ȩUCȩGBCLRGȏCBȩȩQRSBWȩ
about what competences and skills are needed in
a digitally transformed society, and how we should
teach them.
8 Educational perspective 74
numbers from undergraduate courses in national
languages. Notwithstanding this limitation, to be
?BBPCQQCBȩGLȩDSRSPCȩUMPI
topics related to AI.
$GESPCȩȩQFMUQȩRFCȩAMSLRPWȩNPMȏJCQȩDMPȩ?A?BCK-
GAȩMȎCPȩ?Rȩ?JJȩJCTCJQȩ@WȩRFCK?RGAȩȏCJBȩMDȩRFCȩNPM-
'QIGJJQ?LB?A?BCKGAQSNNJW
2MȩF?TCȩ?ȩȏPQRȩSLBCPQR?LBGLEȩMDȩRFCȩASPPCLRȩ?A?BCK-
ic supply on AI, we analysed all the 780 universities
in the EU with a website containing both study pro-
grammes, which are the result of the combination
MDȩ RCAFLGOSCQȩ ?LBȩ ?NNJGA?RGMLȩ ȏCJBQȩ R?SEFRȩ GLȩ RFCȩ
academic programmes in the AI domain: comput-
er vision and natural language processing, ML and
"*
ȩ?LBȩPM@MRGAQȩ2FCȩȏPQRȩMLCȩGQȩ?ȩAMK@GL?RGMLȩMDȩ
EP?KKCQȩ?LBȩICWUMPBQȩPCJ?RCBȩRMȩ'ȩȩȏPQRȩQAPCCL- RUMȩQS@ȏCJBQȩKCRFMBQȩ?LBȩBCTCJMNKCLRQȩGLȩGK?ECȩ
ing showed that a higher number of bachelor studies NPMACQQGLE
ȩ M@HCARȩ BCRCARGMLȩ ?LBȩ AMKNSRCPȩ TGQGML
ȩ
have pages in local languages compared to English, ?LBȩRFCȩȏCJBȩRF?RȩRPC?RQȩL?RSP?JȩJ?LES?EC
ȩGLAJSBGLEȩ
while master studies are mostly advertised in En- speech recognition, machine translation and com-
glish and doctoral studies are almost only in Eng-
lish. This implies that our analysis, based on Eng-
lish only, has been able to capture almost all doc-
toral programmes in the EU countries, and most
K?QRCPQ
ȩ @SRȩ AJC?PJWȩ SLBCPCQRGK?RCQȩ RFCȩ MȎCPGLEȩ
NSR?RGML?JȩJGLESGQRGAQȩ2FCȩȏESPCQȩQFMUȩRF?RȩGLȩKMQRȩ
countries a high proportion of programmes focus
on robotics (81% in all EU), followed by computer
vision and NLP (12% in all EU) and computer vision
and NLP (7%).

Figure 10.$,WKHPDWLFSURÀOHVLQDFDGHPLFSURJUDPPHVDVRIWRWDOSURJUDPPHV

 75 8 Educational perspective

In Figure 11, we show a geographical distribu-
RGMLȩMDȩSLGTCPQGRGCQȩMȎCPGLEȩQRSBWȩRMNGAQȩPCJ?RCBȩRMȩ
AI. The locations are superimposed on the distri-
bution of science and technology skills at NUTS2
JCTCJ
ȩRMȩQFMUȩUFCPCȩRFCPCȩ?PCȩQGELGȏA?LRȩE?NQȩ?LBȩ
opportunities to upskill the local workforce (see
also Chapter 13). It is worth underlying that this is
only a preliminary analysis and that the academic
MȎCPȩ GQȩ SLBCPPCNPCQCLRCBȩ UFCPCȩ QRSBWȩ RMNGAQȩ ?PCȩ
MȎCPCBȩ GLȩ RFCȩ L?RGML?Jȩ J?LES?EC
ȩ ?Qȩ ?Rȩ RFGQȩ QR?ECȩ We have also provided an initial overview and geo-
we only analysed material available in English.
impacts human intelligence in cognitive capacities
in both adults, and even more importantly, children.
Secondly, we need to think beyond current needs and
practices, and consider how AI is likely to change the
relationship between education and work, and human
development. Thirdly, we have highlighted possible
risks related to AI in education, particularly privacy
and ethical ones (for further reading see Tuomi, 2018).
EP?NFGAȩBGQRPG@SRGMLȩMDȩRFCȩ?A?BCKGAȩMȎCPȩMDȩQRSBWȩ
topics related to AI. The academic supply of study
1SKK?PW?LBAMLAJSQGMLQ topics related to AI shows a strong focus on robot-
GAQ
ȩPCȐCARGLEȩMLCȩMDȩRFCȩASPPCLRȩ?PC?QȩMDȩQRPCLERFȩ
'Lȩ RFGQȩ AF?NRCP
ȩ UCȩ F?TCȩ GBCLRGȏCBȩ RF?Rȩ 'ȩ ?JPC?BWȩ in Europe. Further research is needed to improve
contributes to education particularly through intel- on this analysis, which nevertheless provides some
ligent tutoring systems, student assessment, and K?RCPG?JȩDMPȩPCȐCARGMLȩMLȩ?PC?QȩMDȩRFCȩASPPCLRȩJMUȩ
helping to identify potential learning challenges. level of science and technology skills where rein-
forcing the provision of courses may be consid-
5Cȩ F?TCȩ CVNJMPCBȩ NMRCLRG?Jȩ GKN?ARQȩ MLȩ QFMPR?ECQȩ ered important. This is relevant to the discussion in
of skills, on learning and on teaching. Three crucial Chapters 12 and 13 on possible strategies to start
NMGLRQȩ QR?LBȩ MSRȩ DPMKȩ RFCȩ PCTGCUȩ ȏPQRJW
ȩ RFCȩ LCCBȩ preparing society, and the most vulnerable regions,
to understand better how the interaction with AI to the challenges that AI will bring.

Figure 11. Distribution of universities offering AI study topics over human resources in science and technology 2017.
Source: Eurostat Regional Yearbook 2018, and JRC
 9 Economic perspective 76

"*
IBT
B
XJEF
SBOHF
PG
QPUFOUJBM
FDPOPNJD
JNQMJDBUJPOT
GPS
FNQMPZNFOU
BOE
JODPNF
EJTUSJCVUJPO

SUMMARY

In this chapter, we analyse the literature and evidence available to date on the potential
impacts, both positive and negative, with respect to work, growth and inequality. In relation to
UMPI
ȩUCȩȏLBȩRF?RȩLCGRFCPȩRFCMPWȩLMPȩCTGBCLACȩ?PCȩTCPWȩAMLAJSQGTCȩ?RȩRFCȩNPCQCLRȩRGKCȩ'ȩAMSJBȩ
complement and enhance human activity, replace an increasing number of routine tasks, or
@MRFȩLMRFCPȩ?PC?ȩMDȩSLACPR?GLRWȩGQȩRFCȩCVRCLRȩRMȩUFGAFȩ'ȩF?QȩRFCȩNMRCLRG?JȩRMȩQNSPȩCAMLMKGAȩ
EPMURFȩ5FCLȩAMLQGBCPCBȩ?Qȩ?ȩECLCP?JNSPNMQCȩRCAFLMJMEW
ȩ'ȩAMSJBȩQNPC?Bȩ?APMQQȩK?LWȩHM@Qȩ
?LBȩGLBSQRPG?JȩQCARMPQ
ȩ@MMQRGLEȩNPMBSARGTGRW
ȩ?LBȩWGCJBGLEȩQRPMLEȩNMQGRGTCȩEPMURFȩ2MȩRFCȩCVRCLRȩ
that ML generates new insights from data, it may also contribute to the production of new
GBC?Qȩ?LBȩGLLMT?RGMLȩGRQCJDȩ#AMLMKGAȩEPMURFȩKMBCJQȩ?PCȩQR?PRGLEȩRMȩCVNJMPCȩT?PGMSQȩQACL?PGMQȩ
but there is no empirical evidence yet that favours one or the other.

5FCLȩGRȩAMKCQȩRMȩGLCOS?JGRW
ȩUCȩȏLBȩRF?Rȩ'ȩA?Lȩ?ȎCARȩSLD?TMSP?@JWȩRFCȩBGQRPG@SRGMLȩMDȩGLAMKCȩ
RFPMSEFȩK?LWȩAF?LLCJQȩ2FCȩKMQRȩBGQASQQCBȩAMLACPLȩHM@ȩNMJ?PGQ?RGML
ȩPCBSARGMLȩGLȩHM@ȩOS?JGRWȩ
?RȩRFCȩJMUCPȩCLB
ȩ?LBȩ?JQMȩEPC?RCPȩBGȑASJRWȩDMPȩJMUCPQIGJJCBȩUMPICPQȩRMȩ?BHSQRȩRMȩAF?LECȩ?LBȩ
ȏLBȩLCUȩHM@Q
ȩUGRFȩJMLECPȩNCPGMBQȩMDȩSLCKNJMWKCLRȩRF?LȩRFMQCȩUGRFȩFGEFCPȩOS?JGȏA?RGMLQȩ?LBȩ
skills. Some regions of Europe are more vulnerable to these challenges if no action is taken
to prepare society. Therefore, a very proactive strategy needs to be put in place to build the
resilience of regions across Europe.
 77 9 Economic perspective

ECONOMIC
PERSPECTIVE
In this chapter, we discuss the potential impact of
'ȩ MLȩ HM@Q
ȩ EPMURFȩ ?LBȩ GLCOS?JGRGCQ
ȩ @?QCBȩ MLȩ RFCȩ History suggests
QAGCLRGȏAȩJGRCP?RSPCȩ?LBȩCTGBCLACȩ?T?GJ?@JCȩRMȩB?RCȩȩ
There are concerns over the potential impact of AI that technological
on the labour market and income inequalities. Some
?J?PKGQRȩQRSBGCQȩNPCBGARȩK?QQGTCȩHM@ȩJMQQCQȩ?Qȩ'ȩPC- innovation may
NJ?ACQȩFSK?LȩJ?@MSPȩ-LȩRFCȩMRFCPȩF?LB
ȩFGQRMPGAȩCT-
idence shows that previous waves of innovation re- be a substitute for
placing workers with machines ultimately generated
KMPCȩHM@Qȩ?LBȩFGEFCPȩGLAMKCQȩ5CȩCV?KGLCȩGDȩRFCPCȩ human work but
?PCȩPC?QMLQȩUFWȩRFGQȩRGKCȩRFGLEQȩAMSJBȩ@CȩBGȎCPCLRȩ
in the case of AI. AI is likely to improve productivity ultimately increases
and economic growth over a wide range of sectors
in the economy. This could give a welcome boost employment. The
to the current productivity slow-down. At the same
time, the non-rival and general-purpose character- nature of AI may
istics of AI could speed up change across multiple
sectors and lead to greater impacts than past waves accelerate and boost
of innovation, creating more friction in labour mar-
kets. We also focus attention on the distribution- economic growth and
?JȩGKN?ARȩMDȩ'ȩ?LBȩRFCȩPGQIȩRF?RȩGRȩK?WȩCV?ACP@?RCȩ
?KMLEȩCVGQRGLEȩPCEGML?Jȩ?LBȩGLAMKCȩBGQN?PGRGCQ create new jobs. We
.MRCLRG?JGKN?ARMD'MLHM@Q need to ensure that
Debates on the impact of technological change on QEFP?BKBȭQPRP>II
employment go back at least to the beginning of
the Industrial Revolution. Back then, the Luddites Do we have any reasons to assume that this time
attacked spinning machines because they displaced UGJJȩ@CȩBGȎCPCLR
ȩMPȩA?LȩUCȩAMLRGLSCȩRMȩBP?UȩAMK-
FSK?LȩJ?@MSPȩGLȩRFCȩRCVRGJCȩGLBSQRPWȩ&MUCTCP
ȩRFCȩ DMPRȩDPMKȩRFCȩFGQRMPGA?JȩCTGBCLACȩRMȩB?RCȩȩ
long-run historical evidence eases our fears about
being made redundant by machines. Ever since the
Industrial Revolution, and despite many subsequent
waves of innovation and mechanisation, human
employment and incomes have strongly increased,
Traditional economic theory revolves around the
K?ELGRSBCȩ MDȩ RFCȩ QS@QRGRSRGMLȩ CȎCARȩ @CRUCCLȩ
workers and machines. It suggests that machines
(including AI algorithms in computers) and human
labour are not perfect substitutes. Machines may
LMRȩ BCAPC?QCBȩ 1RGJJ
ȩ UGRFȩ CTCPWȩ K?HMPȩ LCUȩ GLTCL- replace workers but cannot do entirely without
RGML
ȩRFCQCȩDC?PQȩPCQSPECȩzȩ?QȩUGRFȩ'ȩ RFCKȩ RFCWȩ ?PCȩ RMȩ QMKCȩ CVRCLRȩ AMKNJCKCLR?PW
 9 Economic perspective 78

When the demand for machines increases be- for workers and machines but without a realloca-
cause they are more productive and cheaper than
workers, this will also increase demand for keep-
ing workers who are necessary to operate the ma-
chines as the goods produced with the machines
become cheaper. Moreover, machines will increase
workers’ productivity and wages for the remaining
human tasks. In this view, technological improve-
ment boosts the demand for labour. The long-run
historical evidence seems to support this.
More recently, however, economists have started to
ȩ Aȩ @?QCBȩ MLȩ 31ȩ B?R?ȩ MȎCPQȩ QMKCȩ QSN-
look at this from a new perspective. First, recent
RGMLȩ MDȩ RFCQCȩ NPMBSARGMLȩ D?ARMPQȩ RMȩ QNCAGȏAȩ R?QIQȩ
-RFCPȩ CȎCARQȩ LCCBȩ RMȩ @Cȩ ?BBCBȩ RMȩ AMKNJCRCȩ RFCȩ
picture. Machines may increase the productivity of
workers for their remaining tasks and thereby push
up wages and/or reduce product prices. Changes in
PCJ?RGTCȩNPGACQȩ?APMQQȩNPMBSARQȩUGJJȩ?ȎCARȩBCK?LBȩ
DMPȩNPMBSARQȩDPMKȩBGȎCPCLRȩQCARMPQȩ?LBȩAF?LECȩRFCȩ
composition of the product basket in the economy.
Recent empirical evidence (Acemoglu and Restre-
NM
port for this task-based model of labour markets
evidence shows that while employment may be CȎCARQȩMDȩ?SRMK?RGMLȩ2FCȩQF?PCȩMDȩU?ECQȩGLȩ%".ȩ
growing in advanced industrial economies, a rapid has been declining strongly in manufacturing, but
increase in the capital share is causing a relative not in services. The authors show that the tradi-
decline in the labour share in value added (or in total tional price-driven substitution and productivity ef-
income generated by the economy) (Karabarbou-
nis and Neiman, 2013). #TGBCLACȧMDȧPGQGLEȧGLAMKCȧ
GLCOS?JGRWȧJCLBQȧDSPRFCPȧQSNNMPRȧRMȧRFCȧQACNRGA?Jȧ
ȩRMȩ?ȩJCQQCPȩCVRCLR
ȩRFCȩQCARMPȩAMKNMQG-
DCARQȩ?PCȩRMMȩUC?IȩRMȩCVNJ?GLȩRFGQȩBCAJGLCȩ2FCȩK?GLȩ
drivers of the decline have been the change in task
AMLRCLRȩ?LB
TGCUȧ MLȧ RFCȧ GKN?ARȧ MDȧ PCACLRȧ RCAFLMJMEGA?Jȧ ?B- RGMLȩCȎCARȩ2FCWȩBCAMKNMQCȩRFCȩR?QIȩAMLRCLRȩCȎCARȩ
T?LACKCLRȧ MLȧ J?@MSPȧ K?PICRQȧ That triggered the
QC?PAFȩ DMPȩ LCUȩ CAMLMKGAȩ KMBCJQȩ RMȩ CVNJ?GLȩ RFCQCȩ
observations.
A recent task-based model of employment (Ace-
moglu and Restrepo, 2016, 2017, 2018a, 2018b)
GLȩ@MRFȩ?ȩBGQNJ?ACKCLRȩ?LBȩPCGLQR?RCKCLRȩCȎCARȩ'Lȩ
the USA at least, displacement due to automation
MDȩCVGQRGLEȩR?QIQȩF?Qȩ@CCLȩQRPMLECPȩRF?LȩEPMURFȩGLȩ
new tasks.
#TGBCLACȩ DMPȩ LCUJWȩ APC?RCBȩ HM@Qȩ GQȩ @?QCBȩ MLȩ LCUȩ
NPMTGBCQȩ ?ȩ @CRRCPȩ CVNJ?L?RGMLȩ DMPȩ RFCQCȩ NFCLMKC- MAASN?RGML?Jȩ RGRJCQȩ GLȩ HM@ȩ QR?RGQRGAQȩ 'Rȩ GQȩ LMRȩ AJC?P
ȩ
na. This model starts from the observation that the however, whether new titles represent new tasks
NPGK?PWȩCȎCARȩMDȩ?SRMK?RGMLȩGQȩRMȩPCNJ?ACȩUMPICPQȩ MPȩMLJWȩ?ȩPC@SLBJGLEȩMDȩCVGQRGLEȩMLCQȩ2FCPCȧGQȧLMȧ
with machines when the latter are more productive AMLQCLQSQȧ MLȧ RFCȧ BCȏLGRGMLȧ MDȧ R?QIQȧ RF?Rȧ K?Wȧ
?Rȩ ?ȩ QNCAGȏAȩ R?QIȩ 2F?Rȩ BGQNJ?ACKCLRȩ CȎCARȩ GLCT T?PWȧ ?APMQQȧ BGȎCPCLRȧ JCTCJQȧ MDȧ EP?LSJ?PGRWȧ It is
itably reduces the wage share of value added for not clear how distinct tasks are from other related
that task and increases the share of capital or prof- R?QIQ
ȩFMUȩQIGJJQȩRF?RȩA?Lȩ@CȩBMK?GLQNCAGȏAȩMPȩBM-
its in value added. However, technological innov- K?GLECLCP?Jȩ ?ȎCARȩ RFCȩ ?@GJGRWȩ RMȩ NCPDMPKȩ ?ȩ R?QI
ȩ
ation may also lead to the creation of new tasks for and how the automation/replacement of particular
UMPICPQȩRF?RȩBGBȩLMRȩCVGQRȩNPCTGMSQJWȩ?LBȩA?LLMRȩ@Cȩ R?QIQȩ?ȎCARQȩRFCȩQRPSARSPCȩMDȩGLBGTGBS?JȩMAASN?RGMLQ
BMLCȩ WCRȩ @Wȩ K?AFGLCQȩ 2FGQȩ PCGLQR?RCKCLRȩ CȎCARȩ
may increase human employment. The combination %PCEMPWȩ CRȩ ?Jȩ ȩ ?NNJWȩ ?ȩ AMKN?P?@JCȩ R?QI
MDȩ BGQNJ?ACKCLRȩ ?LBȩ PCGLQR?RCKCLRȩ CȎCARQȩ PC?JJM- based framework to EU data for the period 1999-
cates tasks between workers and machines. As such, ȩ2FCWȩAMLȏPKȩRFCȩCVGQRCLACȩMDȩ?ȩQRPMLEȩCK-
displacement and reinstatement are in addition to NJMWKCLRPCBSAGLEȩQS@QRGRSRGMLȩCȎCARȩ@SRȩ?JQMȩȏLBȩ
RFCȩRP?BGRGML?JȩCȎCARQȩGLȩRFCȩQS@QRGRSRGMLȩKMBCJȩ RF?RȩAMKNJCKCLR?PWȩBCK?LBȩ?LBȩQNGJJMTCPȩCȎCARQȩ
more than compensate for this, so that the net em-
2FCȩAPSAG?JȩBGȎCPCLACȩUGRFȩRFCȩRP?BGRGML?JȩQS@QRGRS- NJMWKCLRȩCȎCARȩMDȩRCAFLMJMEWȩGQȩNMQGRGTCȩ&MUCTCP
ȩ
RGMLȩ CȎCARȩ GQȩ RF?Rȩ RFCȩ J?RRCPȩ AF?LECQȩ RFCȩ BCK?LBȩ RFGQȩȏLBGLEȩBCNCLBQȩMLȩA?NGR?JȩGLAMKCȩE?GLQȩDPMKȩ
 79 9 Economic perspective

technological progress feeding back into product A second research direction was to focus on the
demand. If only wage income gains feedback into impact of industrial robots on employment and
BCK?LB
ȩRFCȩRMR?JȩJ?@MSPȩCȎCARȩGQȩMLJWȩF?JDȩ?QȩJ?PECȩ
This underlines the importance of income redistri-
bution policies (see Section 9.3).
Note that the historical period covered by the USA
and EU studies dates back long before AI emerged
wages. Robots constitute a mechanically embod-
ied form of AI that is more easily observable and
measurable than algorithms. However, they are an
GKNCPDCARȩNPMVWȩKC?QSPCȩDMPȩ'ȩ@CA?SQCȩPM@MRQȩ?PCȩ
rival products that can only be used for one task at
the time while AI algorithms are non-rival and can
?Qȩ?Lȩ?SRMK?RGMLȩRCAFLMJMEWȩQȩQSAF
ȩGRȩMȎCPQȩMLJWȩ be used for many tasks at the same time. In fact,
a general view on the impact of technological innov- neither the task-based model nor the pure labour
?RGMLȩMLȩCKNJMWKCLR
ȩLMRȩ?ȩQNCAGȏAȩ'ȩNCPQNCARGTC displacement studies consider the implications of
the non-rival nature of AI. Also, robots are mainly
#AMLMKGQRQȩF?TCȩRPGCBȩRMȩȏLBȩKMPCȩPCACLRȩCKNGP- employed in manufacturing, especially car manu-
ical evidence that is more closely related to the facturing, not in services. This results in selection
GLRPMBSARGMLȩMDȩ'ȩRCAFLMJMEW
ȩGLȩRUMȩU?WQȩȩȏPQRȩ bias in the evidence. Some evidence from the USA
avenue was to try to be forward-looking and es- suggests a negative impact of robots on employ-
timate the risk of automation of occupations.
Some alarmist studies produced estimates of
the percentage of human occupations at risk of
being automated in the foreseeable future, vary-
GLEȩ DPMKȩ ?JKMQRȩ ȩ Ϥȩ $PCWȩ ?LBȩ -Q@MPLC
ȩ ȩ Dauth et al., 2018).
down to 13 or 9 % (Arntz et al., 2016; Nedelkoska
and Quintini, 2018).44ȩȩK?HMPȩNPM@JCKȩUGRFȩRFCQCȩ
studies is that they focus only on the displace-
ȩ RFCPCȩ GQȩ
KCLRȩ CȎCARȩ ?LBȩ GELMPCȩ ?JJȩ RFCȩ NMQGRGTCȩ CKNJMW-
KCLRȩ CȎCARQȩ RF?Rȩ 'ȩ AMSJBȩ GLBSACȩ zȩ ?Qȩ CVNJ?GLCBȩ
in the task-based model by Acemoglu and Re-
strepo (2018c). Another problem is that the level
ȩ
MDȩ EP?LSJ?PGRWȩ MDȩ RFCȩ BCȏLGRGMLȩ MDȩ R?QIQȩ APC?RCQȩ
ȩ BGQASQQȩ RFCȩ NMQQG@GJGRWȩ MDȩ RCAFLMJME-
KCLRȩ?LBȩU?ECQȩ2FCPCȩGQȩAMLȐGARGLEȩCTGBCLACȩDMPȩ
the EU where one study suggests a negative im-
N?ARȩ UFGJCȩ ?ȩ QCAMLBȩ QRSBWȩ DMPȩ %CPK?LWȩ AMKCQȩ RMȩ
the opposite conclusion (Chiacchio et al., 2018;
Although the literature is not conclusive on the
LCRȩ CKNJMWKCLRȩ CȎCARQȩ MDȩ PM@MRGQ?RGML
?KNJCȩ CTGBCLACȩ NMGLRGLEȩ RMȩ K?HMPȩ PC?JJMA?RGMLQȩ
and broad structural changes. The speed and ef-
ȏAGCLAWȩ MDȩ RFCQCȩ AF?LECQȩ A?Lȩ ?JQMȩ GLȐSCLACȩ RFCȩ
LCRȩCKNJMWKCLRȩCȎCARȩ)MPGLCIȩ?LBȩ1RGEJGRXȩ
DMPȩ CV?KNJC
FGEFȩT?PG?@GJGRWȩGLȩRFCȩPCQSJRQȩ-AASN?RGML?JȩMPȩCTCLȩ ical unemployment as a transition phenomenon.
sub-occupational level approaches to skills are too Such a case gets strong support from the recent-
coarse-grained to be reliable. Both problems lead ly emerged view among scholars of globalisation
RMȩ ?Lȩ MTCPCQRGK?RGMLȩ MDȩ HM@ȩ BGQNJ?ACKCLRȩ CȎCARQȩ that the reallocation of workers among sectors is
?LBȩCVACQQGTCJWȩ?J?PKGQRȩPCQSJRQ not frictionless, it can take many years, and the
implied costs may be equivalent to multiple years
2FCȧ BCEPCCȧ MDȧ PMSRGLCȧ GLȧ R?QIQȧ GQȧ ?Lȧ GKNMPR?LRȧ of lost income for workers (see Hornok and Koren,
APGRCPGMLȧRMȧ?QQCQQȧUFCRFCPȧRFCWȧA?Lȧ@CȧPCNJ?ACBȧ 
ȩDMPȩ?LȩMTCPTGCUȩ'RȩGQȩPC?QML?@JCȩRMȩCVNCARȩ
@Wȧ'ȧ PWLHMJDQQMLȩCRȩ?J
ȩ
ȩ?JRFMSEFȩAPC?RGTCȩ that AI-induced reallocations would face similar
and social interaction tasks are not necessarily be- challenges.
yond the scope of today’s ML algorithms provided
that large quantities of data are available to train 2FCȧ QNCCBȧ MDȧ 'ȧ SNR?ICȧ GQȧ RFSQȧ GKNMPR?LRȧ LMRȧ
the algorithm. The organisation of work in a par- MLJWȧDMPȧRFCȧCVNCARCBȧNPMBSARGTGRWȧE?GLQ
ȧbut for
RGASJ?Pȩ QMAG?Jȩ AMLRCVR
ȩ P?RFCPȩ RF?Lȩ RFCȩ R?QIȩ GRQCJD
ȩ RFCȧGKNJGCBȧPC?JJMA?RGMLȧNPMACQQ, too. However, at-
may also be an obstacle to automation (Fernan- RCKNRQȩRMȩNPCBGARȩRFGQȩQNCCBȩ?PCȩQS@HCARȩRMȩUGBCȩSL-
dez-Macias and Bisello, 2017). certainty +A)GLQCWȩ%JM@?Jȩ'LQRGRSRC
ȩ
 9 Economic perspective 80

ȩȏL?J
ȩKMPCȩNFGJMQMNFGA?JȩNMGLRȩGQȩRMȩPCAMELGQCȩRF?Rȩ at a steady pace, although an initial increase in the
CKNJMWKCLRȩGQȩLMRȩHSQRȩN?GBȩK?PICRȩUMPIȩ'Rȩ?JQMȩGL- share of AI machines (capital) in the value of output
cludes work at home or in communities, and may may slow down growth.
have an intrinsic value. As stated by Korinek and
1RGEJGRXȩ
ȩ|HM@QȩNPMTGBCȩLMRȩMLJWȩGLAMKCȩ@SRȩ?JQMȩ The second type of scenario focuses on the impact
other mental services such as meaning, dignity and of AI on the production of innovative ideas which,
DSJȏJKCLRȩ RMȩ FSK?LQ}ȩ 2Mȩ AMSLRCP?ARȩ RFCȩ NMRCLRG?Jȩ so far, has been the unique property of human la-
JMQQȩMDȩN?GBȩHM@Q
ȩNMJGAGCQȩK?WȩLCCBȩRMȩNPMKMRCȩMRFCPȩ bour. The invention of new ideas can be seen as
DSJȏJJGLEȩU?WQȩRMȩQNCLBȩRGKCȩ1RCTCLQML
ȩ. a production process with a series of tasks, some
of which can be automated and taken over by AI.
.MRCLRG?JGKN?ARMD'MLEPMURF
ȩUCȩA?LȩQGKSJ?RCȩ?LȩCVRPCKCȩQACL?PGMȩUFCPC-
5Cȧ RCLBȧ RMȧ JMMIȧ ?Rȧ RFCȧ CKNJMWKCLRȧ CȎCARQȧ MDȧ
'ȧDPMKȧ?ȧUMPPGQMKCȧ?LEJCȧzȩRFCȩPGQIȩMDȩUMPICPQȩ
E?GL
by all tasks for the production of new ideas are tak-
en over by AI. This would lead to a ‘singularity’46
scenario whereby economic growth would become
@CGLEȩ BGQNJ?ACBȩ @Wȩ K?AFGLCQȩ zȩ ?LBȧ DMPECRȧ ?@MSRȧ GLȏLGRCJWȩFGEFȩGLȩȏLGRCȩRGKCȩ&MUCTCP
ȩCTCLȩRF?RȩQAC-
RFCȧ NPMKGQGLEȧ CȎCARQȧ MDȧ 'ȧ 'ȧ K?AFGLCQȧ A?Lȧ nario will run into constraints, including physical
@MMQRȧ NPMBSARGTGRWȧ AMKN?PCBȧ RMȧ FSK?Lȧ J?@MSPȧ AMLQRP?GLRQȩRMȩA?PPWȩMSRȩ?LȩGLȏLGRCȩLSK@CPȩMDȩK?-
?LBȧRFCPCDMPCȧECLCP?RCȧKMPCȧCAMLMKGAȧ@CLCȏRQȧ AFGLCȩR?QIQȩ*CQQȩCVRPCKCȩQACL?PGMQȩJC?BȩRMȩ?ȩQRCN-
DMPȧ?JJ This would be very welcome in the current wise increase in the level of the economic growth
NCPGMBȩUFCPCȩCAMLMKGCQȩ?PCȩQSȎCPGLEȩDPMKȩQJMUGLEȩ P?RCȩ@SRȩUGRFMSRȩ@CAMKGLEȩCVNJMQGTC
productivity growth.
Economists see human labour, (AI) machines, and
innovative ideas as production factors in the growth
equation (Aghion et al., 2017). These factors pro-
duce an output that is sold, the value of which is
shared between labour and machines (or capital).
%PMURFȩ CAMLMKGQRQȩ JMMIȩ ?Rȩ RFCȩ P?RCȩ MDȩ EPMURFȩ MDȩ
that production as well as the distribution of the
value of production between capital and labour.
2FCWȩ SQCȩ CVGQRGLEȩ KMBCJQȩ MDȩ CAMLMKGAȩ EPMURFȩ RMȩ
simulate two types of future AI-driven scenarios.
2FCQCȧEPMURFȧQACL?PGMQȧPCK?GLȧP?RFCPȧRFCMPCRGA?J
In terms of empirical evidence, RFCWȧ PSLȧ GLRMȧ RFCȧ
Q?KCȧNPM@JCKQȧ?QȧCKNJMWKCLRȧDMPCA?QRQȧRFCPCȧ
?PCȧLMȧB?R?ȧWCRȧRMȧE?SECȧRFCȧGKN?AR. In addition,
available evidence for previous episodes of mas-
QGTCȩKCAF?LGQ?RGMLȩKGEFRȩLMRȩȏRȩRFCȩNCASJG?PȩDC?-
tures of the AI revolution. Still, growth economists
are trying to push forward the ‘time frontier’ in the
empirical evidence, either by looking at past waves
MDȩGLLMT?RGMLȩGLȩMRFCPȩRCAFLMJMEGCQȩMPȩ@WȩCV?KGLGLEȩ
the economic nature of AI as a new technology.

2FCȩȏPQRȩRWNCȩMDȩQACL?PGMȩDMASQCQȩMLȩRFCȩNPMBSARGMLȩ !MAI@SPLȩ CRȩ ?Jȩ ȩ ȏLBȩ NPCJGKGL?PWȩ CTGBCLACȩ

of goods and services, and on the substitution or that 'ȧGQȧ?ȧECLCP?JNSPNMQCȧRCAFLMJMEWȧRF?RȧȏRQȧ
complementarity between humans and machines ?ȧUGBCȧP?LECȧMDȧ?NNJGA?RGMLQȧ?LBȧQNPC?BQȧP?NGBJWȧ
GLȩRF?RȩNPMBSARGMLȩNPMACQQȩ#TCLȩ?LȩCVRPCKCȩQACL ?APMQQȧK?LWȧQCARMPQȧ?LBȧGLBSQRPGCQ This could ac-
ario whereby nearly all human tasks would be car-
ried out by AI yields a reassuring outcome: the la-
ȩSLJGICJWȩ?LWȩNPCTGMSQȩRCAFLMJ-
bour share in the value of output will remain high
because remaining human tasks become very pro-
ductive and highly remunerated. The ‘last’ human
ACJCP?RCȩRFCȩNPMBSARGTGRWȩ?LBȩEPMURFȩCȎCARQȩ2FCWȩ
?JQMȩȏLBȩRF?Rȩ'ȩA?L
ogy, automate and accelerate the invention process
itself, and thus the speed of innovation. This would
ultimately translate into accelerated productivity
R?QIȩUGJJȩDCRAFȩ?LȩCVRPCKCJWȩFGEFȩU?ECȩ2F?RȩU?ECȩ EPMURFȩQȩ?LȩCV?KNJC
ȩE?PU?JȩCRȩ?JȩȩQFMUȩ
will have to be redistributed among unemployed how the use of AI in the pharmaceutical and chem-
workers in order to ensure some income for the lat- ical industries speeds up the process of discovery
ter.45 Economic growth will continue in that scenario of new molecules for useful applications. These
 81 9 Economic perspective

‘recombinant growth’ scenarios show how AI can

be better than humans at detecting promising new
AMK@GL?RGMLQȩMDȩCVGQRGLEȩRCAFLMJMEGCQȩMPȩILMUJCBECȩ

-RFCPȩ ?SRFMPQȩ A?SRGMLȩ ?@MSRȩ RFCȩ RGKCȩ GRȩ R?ICQȩ RMȩ

PMJJȩMSRȩRFCQCȩMNRGKGQRGAȩDMPCA?QRQȩ PWLHMJDQQMLȩCRȩ
al. (2017) argue that the Solow ‘productivity par-
The realisation
?BMVȩ ȩ P?NGBȩ EPMURFȩ GLȩ 'ȩ GLTCQRKCLRQȩ AMK@GLCBȩ
UGRFȩ QJMUȩ GLAPC?QCQȩ GLȩ NPMBSARGTGRWȩ @CLCȏRQȩ zȩ ?N-
of the full
plies to AI as well. 2FCȧPC?JGQ?RGMLȧMDȧRFCȧDSJJȧCD-
DCARQȧMDȧ'ȧMLȧNPMBSARGTGRWȧK?WȧR?ICȧK?LWȧWC?PQ
effects of AI
because it requires the development of comple-
mentary innovations, organisational changes and
on productivity
LCUȩFSK?LȩJ?@MSPȩQIGJJQȩ PWLHMJDQQMLȩCRȩ?Jȩȩ
ȏLBȩ QMKCȩ CTGBCLACȩ RF?Rȩ RFCȩ SNR?ICȩ MDȩ ?ȩ ECLCP?J
may take
NSPNMQCȩRCAFLMJMEWȩJGICȩ'ȩDMJJMUQȩ?ȩ(ASPTCȩCȎCARȩ
The initial impact may be negative because heavy many years.
intangible investments in reorganisation and re-
training drag down growth that only rebounds in
later years. Historical evidence suggests that ear-
lier general-purpose technologies followed a similar
pattern. The fact that AI is a general-purpose tech-
LMJMEWȩUMSJBȩLMRȩ@Cȩ?ȩPC?QMLȩRMȩCVNCARȩRFGLEQȩRMȩ@Cȩ
BGȎCPCLRȩRFGQȩRGKC
'ȩKGEFRȩ?JQMȩ?ȎCARȩȏPKȩMPE?LGQ?RGMLȩ?JMLEȩQCTCP?Jȩ
CQGBCQȩGRQȩBGPCARȩCȎCARQȩMLȩRFCȩNPMBSARGMLȩMDȩEMMBQȩ dimensions. First, it might change endogenously the
or ideas, 'ȧ A?Lȧ ?JQMȧ ?ȎCARȧ EPMURFȧ @Wȧ AF?LEGLEȧ degree of complementarity/substitutability between
ȏPKȧ@CF?TGMSPȧ?LBȧK?PICRȧAMKNCRGRGMLȩ'DȩCVGQRGLE UMPICPQȩ UGRFȩ BGȎCPCLRȩ QIGJJQȩ zȩ QMKCRFGLEȩ RF?Rȩ GQȩ
leading products or technologies are easily imitated AMLQGBCPCBȩ ȏVCBȩ GLȩ RFCȩ QR?LB?PBȩ EPMURFȩ KMBCJQȩ
by competitors in AI, it can either discourage poten- By facilitating the outsourcing of some low-skilled
RG?Jȩ GLLMT?RMPQ
ȩ MPȩ JC?Bȩ ȏPKQȩ ?Rȩ RFCȩ RCAFLMJMEWȩ tasks, AI would at the same time strengthen the
DPMLRGCPȩRMȩGLLMT?RCȩKMPCȩGLȩMPBCPȩRMȩCQA?NCȩȏCPACȩ complementarity of those who are retained within
competition by the imitators. At any rate, easier RFCȩȏPKȩUGRFȩRFCȩFGEFQIGJJȩR?QIQȩRF?RȩA?LLMRȩ@CȩMSR-
imitation implies that AI-intensive production sec- sourced. By the same token, if AI reduces informa-
tors might become congested in relatively little time. tion asymmetries and facilitates monitoring, it can
AI-driven platforms that handle massive amounts promote downstream delegation of power and more
MDȩB?R?ȩK?Wȩ?BTCPQCJWȩ?ȎCARȩAMKNCRGRGMLȩ?LBȩJC?Bȩ decentralisation of authority. The end result for AI
RMȩ RFCȩ CKCPECLACȩ MDȩ ?ȩ BMKGL?LRȩ NJ?RDMPKȩ zȩ N?P- intensive sectors would be larger and more horizon-
RGASJ?PJWȩGDȩRFCPCȩ?PCȩCAMLMKGCQȩMDȩQA?JC
ȩJMA?JȩCVRCP- R?JJWȩGLRCEP?RCBȩȏPKQ
ȩUGRFȩKMPCȩNPMȏRȩACLRPCQ
ȩUFGAFȩ
nalities or rents. Algorithmic collusion may distort outsource an increasing number of tasks to inde-
competition in some markets. This situation of mo- pendent self-employed agents (Aghion et al., 2017).
nopoly, or at best oligopoly, may restrict access to
the data collected by the platforms with non-trivial Finally, similarly to what has happened with the IT
implications for trade and government regulations revolution, RFCȧ CȎCARȧ MDȧ RFCȧ BGȎSQGMLȧ MDȧ ILMU
%MJBD?P@ȩ?LBȩ2PCȐCP
ȩ
ȩ?LBȩK?ICQȩ?ȩA?QCȩDMPȩ JCBECȧ DPMKȧ 'ȧ KGEFRȧ BGQNPMNMPRGML?RCJWȧ @CLCȏRȧ
?LRGRPSQRȩAMLQGBCP?RGMLQȩ%GJCQ
ȩ FGEFRCAFȧQCARMPQȧ(Baslandze, 2016).
 9 Economic perspective 82

 .MRCLRG?JGKN?ARMD'MLGLCOS?JGRW 2FCPCȧGQȧCTGBCLACȧDMPȧ#SPMNC
ȧRMM
ȧRF?RȧJ?@MSPȧK?P-
ICRȧNMJ?PGQ?RGMLȧJC?BQȧRMȧ?ȧUGBCLGLEȧMDȧU?ECȧE?NQ
2FCȧ K?HMPGRWȧ MDȧ RFCȧ JGRCP?RSPCȧ NPCBGARQȧ ?Lȧ SL- %MMQȩCRȩ?JȩȩȏLBȩRF?RȩGKNPMTCBȩRCAFLMJMEWȩF?Qȩ
D?TMSP?@JCȧ GKN?ARȧ MDȧ 'ȧ MLȧ RFCȧ BGQRPG@SRGMLȧ MDȧ led to increased demand for both well-paid, high-
GLAMKCQȧThe most widely discussed and debated
aspect is the increase in labour income inequality.
ȩ RFSQȩ
Evidence suggests that labour market polarisation
plays an important role in this. We observe polar-
ising labour markets because tasks that are not
QIGJJCBȩ?QȩUCJJȩ?QȩJMUN?GBȩJMUQIGJJCBȩHM@QȩUFGJCȩRFCȩ
BCK?LBȩ DMPȩ KGBBJCGLAMKCȩ HM@Qȩ BCAPC?QCB
supporting the hypothesis that technological pro-
gress can lead to income inequality. Yet, there are
QS@QR?LRG?JJWȩ BGȎCPGLEȩ RPCLBQȩ GLȩ J?@MSPȩ K?PICRȩ NM-
easily performed by AI tend to be found at opposite
ȩ
ends of the skills spectrum while AI tends to replace
ȩ 
ȩ UFGAFȩ A?Lȩ @Cȩ CV-
humans in tasks that correspond to the ‘mid-skill’
category (Autor et al., 2003). Acemoglu and Autor
(2011) and Autor and Salomons (2017) show sug-
ECQRGTCȩ CTGBCLACȩ DPMKȩ RFCȩ 31ȩ MLȩ FMUȩ HM@ȩ NMJ?P
isation translates into wage polarisation or even a
polarisation in working conditions.
J?PGQ?RGMLȩ@CRUCCLȩ#SPMNC?LȩAMSLRPGCQȩ%MMQȩCRȩ?J
ȩ "?PT?Qȩ ?LBȩ 5MJȎ
NJ?GLCBȩ@WȩAMSLRPWQNCAGȏAȩGLQRGRSRGMLQȩ?LBȩNMJGAGCQȩ
(Fernández-Macías, 2012; Fernández-Macías and
Hurley, 2016).
ȩPCNMPRȩ@WȩRFCȩ-#!"ȩȩQFMUQȩ?ȩBGȎCPCLRȩRWNCȩ
of polarisation that is regional and occurs within
AMSLRPGCQȩ $MPȩ GLQR?LAC
ȩ RFCȩ QF?PCȩ MDȩ HM@Qȩ ?Rȩ FGEFȩ
risk of automation varies by 12 % between regions
Ɏɛɤǯ2FCGKN?ARMD'MLUMPIGLE in Spain but only by 1 % between regions in Canada.
 AMLBGRGMLQ Technological progress tends to perpetuate the de-
velopmental divide within countries as regions that
Because its use in actual work processes is still very limited,
?PCȩ CVNCARCBȩ RMȩ @Cȩ KMPCȩ LCE?RGTCJWȩ ?ȎCARCBȩ @Wȩ
there is very little hard evidence yet on the impact of AI on
RCAFLMJMEGA?JȩNPMEPCQQȩ?JQMȩCVFG@GRȩJMUȩNPMBSARGTGRWȩ
working conditions. A recent Eurofound report based on the
growth and high unemployment rates.
NR>IFQ>QFSB>K>IVPFPLCȭSBBJBODFKDQB@EKLILDFBPFK#ROLMB>K
manufacturing (including advanced robotics and the Internet
Despite similar positive trends in labour demand
of Things, which make intensive use of AI) concludes that there
DMPȩ@MRFȩFGEFȩ?LBȩJMUN?WGLEȩHM@Q
ȩUCȩM@QCPTCȩBG-
are some concerns in this sense, in particular with respect to the
TCPEGLEȩRPCLBQȩGLȩPCQNCARGTCȩHM@ȩOS?JGRWȩ-LȩRFCȩMLCȩ
degree of autonomy, privacy and control of workers (Eurofound
hand, technological progress leads to increasing
 "FDFQ>I C>@QLOFBP TEBOB >II L?GB@QP >OB BNRFMMBA TFQE
U?ECQȩGLȩFGEFN?WGLEȩHM@QȩRF?RȩPCOSGPCȩQIGJJQȩUFGAFȩ
connected sensors and where workers collaborate with
complement AI (Deming, 2017). In contrast, tech-
advanced robots can easily become digital panopticons, where
nological progress causes even further reductions
human operators feel permanently monitored and controlled. It
in wages at the lower end of the wage distribution
PBBJPAFȯ@RIQQLJ>FKQ>FK>KVPBKPBLCMOFS>@V down to a level that does not support a reasonable
at work if every object is a monitoring device. And standard of living (Autor and Salomons, 2017). The
while management through algorithms and big data analytics
accompanying digitalisation of the economy and
@>KPFDKFȭ@>KQIVFJMOLSBQEBBȯ@FBK@VLCQEBI>?LROMOL@BPP
FQ
the emergence of platforms causes an increase in
@>K>IPLIB>AQLOLRQFKFP>QFLKLCQ>PHP
TLOHFKQBKPFȭ@>QFLK>KA
precarious forms of self-employment that are char-
QEB>PMEVUF>QFLKLC>KVPBKPBLC>RQLKLJV-KQEBLQEBOE>KA

the introduction of advanced robots, connected
devices and big data in the workplace can also
acterised by a limited duration, such as seasonal or
on-call work, as well as the absence of social secu-
PGRWȩAMTCP?ECȩ-#!"
ȩȩ2FGQ
ȧRMM
ȧAMLRPG@SRCQȧ
contribute to a decline of repetitive and routine RMȧGLAPC?QGLEȧGLCOS?JGRWȧCTCLȧDSPRFCP
work, as well as a reduction in the number of hazardous tasks.
'BK>?IBATLOHMI>@BP>OBIFHBIVQL?BP>CBO
TFQEJLOBPHFIIBA
As discussed in a previous subsection, the speed
workers carrying out less repetitive work.
of adoption of AI in the labour market could cause
 83 9 Economic perspective

transitional unemployment if AI makes workers

redundant at a faster pace than they can move on to
LCUȩHM@Qȩ)MPGLCIȩ?LBȩ1RGEJGRX
ȩȩ"GȎCPCLACQȧGLȧ
RFCȧN?ACȧMDȧ'ȧ?BMNRGMLȧ?APMQQȧBGȎCPCLRȧPCEGMLQȧ
?LBȧQCARMPQ
ȧ?QȧUCJJȧ?QȧBGȎCPCLACQȧ@CRUCCLȧUMPI-
CPQȧGLȧRFCGPȧ?@GJGRWȧ?LBȧQNCCBȧRMȧ?BHSQRȧRMȧMAASN
?RGML?JȧAF?LECQȧK?WȧDSPRFCPȧGLAPC?QCȧGLCOS?JGRW
For instance, historical data suggests that low-skill
UMPICPQȩ?PCȩQJMUCPȩRF?LȩFGEFQIGJJȩMLCQȩGLȩ?BHSQRGLEȩ
to sudden structural labour market changes and
transitioning to new sectors and occupations. This
results in longer period of transitional unemploy-
KCLRȩ DMPȩ JMUQIGJJȩ UMPICPQȩ %MMJQ@CC
ȩ ȩ 2FCȩ
Q?KCȩF?Qȩ@CCLȩCQR?@JGQFCBȩGLȩRFCȩAMLRCVRȩMDȩEJM@
?JGQ?RGMLȩ "GV!?PLCGPMȩ ȩ ȏLBQȩ RF?Rȩ DCK?JC
ȩ
less-educated and older workers face substantially
higher costs of reallocation.
New evidence from the impact of robots on the
Ɏɛɤǰ '?LBBGEGR?JJ?@MSPNJ?RDMPKQ
"FDFQ>I I>?LRO MI>QCLOJP @>K ?B ABȭKBA >P AFDFQ>I KBQTLOHP
QTLOHP
that coordinate labour service transactions in an algorithmic
rithmic
K?LSD?ARSPGLEȩ QCARMPȩ GLȩ %CPK?LWȩ QSEECQRQȩ RF?Rȩ T>VÄBPLIBBQ>I
%LJBWBQ>I
IQELRDEQEBV
EQEBV
RP?LQGRGML?Jȩ SLCKNJMWKCLRȩ CȎCARQȩ K?Wȩ LMRȩ @Cȩ are still marginal in terms of employment, they represent sent a
that strong as a large part of the workers manage radical departure from traditional forms of work and thus hus do
RFCȩRP?LQGRGMLȩUGRFGLȩRFCGPȩȏPKQȩ?LBȩ?APMQQȩMAASN KLQȭQTBIIFKQLBUFPQFKD@>QBDLOFBPLOOBDRI>QFLKP1EBFORPBLC
RPBLC
?RGMLQȩ 7CR
ȩ RFGQȩ HM@ȩ QCASPGRWȩ AMKCQȩ ?Rȩ RFCȩ AMQRȩ digital technologies, such as geolocation, big data analyticscs andd
MDȩ PCBSACBȩ U?ECȩ EPMURFȩ DMPȩ ?BHSQRGLEȩ UMPICPQȩ algorithmic management, allow them to break the traditional itionall
(Dauth et al., 2018). ?LRKA>OFBP?BQTBBKJ>OHBQP>KAȭOJP
MBOCLOJFKDCRK@QFLKP K@QFLKP
of both mediation and management for the transactionss they
1SAFȩQGELGȏA?LRȩPC?JJMA?RGMLQȩA?LȩC?QGJWȩECLCP?RCȩ?ȩ coordinate. They also imply a radical change in the nature ture of
sense of unfairness. Rodrik (2018) argues that so TLOH
?B@>RPBQEBFO@LLOAFK>QFLKBȯ@FBK@VC>@FIFQ>QBPQEBAFOB@Q
AFOB@Q
far this has been mainly associated with globalisa- provision of labour services in small discrete tasks, breaking
eaking
tion. Nevertheless, such sentiments may be evoked up the traditional bundling of tasks into jobs and occupations
pations
in the future with respect to automation and digital
technologies, leading to a political backlash against
technological progress, and AI in particular.
(M@ȩNMJ?PGQ?RGMLȩCVCPRQȩGRQȩCȎCARȩKMQRJWȩMLȩRFCȩBGQ-
persion of labour income. There are, however, add-
itional channels through which the overall distribu-
tion of income (and hence inequality) may change.
First, AI increases the share of capital income rel-
ative to that of human labour (Korinek and Stiglitz,
2017; Sachs, 2017). Together with the high de-
gree of concentration of AI industries, it may lead
to an increase in the inequality of capital income
$BOKĮKABW+>@Ĺ>P
What is the role of AI in digital labour platforms? Sincee they
rely on big data analytics and complex algorithmic rules,
digital labour platforms already make intensive nsive
use of AI techniques. But the potential for advanced
ced AAII
QB@EKFNRBPQL@LKQFKRBBUM>KAFKDQEB@LLOAFK>QFLKBȯ@FBK@VLCK@VLC
digital labour platforms is big, and will be crucial for its further
e
development.
"FDFQ>I I>?LRO MI>QCLOJP FIIRPQO>QB QEB MLQBKQF>I BȬB@Q LC '
'
on employment which goes beyond the direct replacement ment or
augmentation of existing labour input. The use of AI for the
management and coordination of labour has the potential ential
QL@E>KDBQEBLOD>KFP>QFLKLCTLOHPFDKFȭ@>KQIV >KQIV,
?LBȩ ?JQMȩ RMR?Jȩ GLAMKCȩ 1CAMLB
ȩ RFCȩ CVRP?ȩ UC?JRFȩ alter the very nature of employment contracts and further
created by AI is likely be shared unequally across develop the division of labour and task specialisation.
 9 Economic perspective 84

countries. Winner countries will have higher income or the reading of medical images. Even creative and
levels, and more room for domestic redistributive social tasks are not out of the realms of AI abil-
policies (Lee, 2017). Third, as time evolves, AI may GRGCQȩ PWLHMJDQQMLȩ?LBȩ+GRAFCJJ
ȩȩ'LȩRFCȩCLB
ȩ
enable improved health and skills, leading to forms 'ȩK?WȩNPMBSACȩBCNMJ?PGQGLEȩCȎCARQ
of human enhancement. If not limited by policies,
the ability to ‘purchase’ skills and health will fur-
RFCPȩ UGBCLȩ NPCCVGQRGLEȩ GLCOS?JGRGCQȩ &?P?PG
ȩ ȩ
Korinek and Stiglitz, 2017).
ȧRFCȧRFCMPCRGA?Jȧ?LBȧCKNGPGA?JȧCAMLMKGAȧ
ȧ
It is commonly agreed that policy measures are
LCCBCBȩRMȩAMSLRCP?ARȩRFCȩLCE?RGTCȩCȎCARQȩMDȩRCAF-
nological progress on equality. Besides increasing
?AACQQȩRMȩFGEFN?WGLEȩHM@QȩP?GQGLEȩRFCȩMTCP?JJȩQIGJJȩ
JCTCJȩ RFPMSEFȩ GLAPC?QCBȩ CBSA?RGMLȩ CVNCLBGRSPCQ
ȩ studies use historical data that go back long be-
policies should further ensure a reasonable stand-
ard of living. The literature discusses several types
 1SKK?PW?LBAMLAJSQGMLQ
2MȧB?RC
PCQC?PAFȧJGRCP?RSPCȧGQȧ?K@GESMSQȧMLȧRFCȧEPMURF
CKNJMWKCLRȧ?LBȧU?ECȧCȎCARQȧMDȧ'. It can be nega-
tive if machines only substitute human labour, but
positive if machines instead complement human
workers and increase overall productivity. Empirical
fore the rise of AI. They can tell us something about
technological change in general but little about AI
of policies to achieve this goal, such as universal QNCAGȏA?JJWȩ ȩ PCACLRȩ 31ȩ QRSBWȩ ȏLBQȩ ?ȩ LCE?RGTCȩ
basic income or guaranteed employment (Furman
and Seamans, 2018), policies that aim at a redis-
ȩ?JQMȩ@CA?SQCȩA?NGR?Jȩ
tribution from ‘winners’ to ‘losers’ and policies that
QFGDRȩ RFCȩ R?V?RGMLȩ MDȩ FSK?Lȩ J?@MSPȩ RMU?PBQȩ RFCȩ
R?V?RGMLȩ MDȩ A?NGR?Jȩ )MPGLCIȩ ?LBȩ 1RGEJGRX
GKN?ARȩ MDȩ RCAFLMJMEWȩ MLȩ HM@Qȩ ?LBȩ U?ECQȩ Lȩ #3ȩ
QRSBWȩȏLBQȩ?ȩNMQGRGTCȩGKN?AR
gains feedback into overall demand. Studies that
SQCȩPM@MRGAQȩB?R?ȩ?Qȩ?ȩPGT?JȩNPMVWȩDMPȩLMLPGT?Jȩ'ȩ
ȩ ȩ 'Lȩ investments generate inconclusive evidence about
any case, the within-country regional variation of the impact of AI. Some forward-looking studies es-
the impact of technological progress suggests that RGK?RCȩRFCȩLSK@CPȩMDȩHM@Qȩ?RȩFGEFȩPGQIȩMDȩ'BPGTCLȩ
NMJGAGCQȩQFMSJBȩ@Cȩ?BHSQR?@JCȩRMȩJMA?JȩLCCBQ
ȩUFCPCȩ ?SRMK?RGMLȩ 2FCWȩ NPMBSACȩ CVACQQGTCJWȩ ?J?PKGQRȩ PC-
JMA?JȩMȑACQȩA?LȩFCJNȩGLȩRFCȩBCQGELȩMDȩR?PECRCBȩNMJ- sults because they only look at substitution and omit
icies. It is important to be aware that such policies complementarity.
K?WȩJC?BȩRMȩGLCȑAGCLAGCQȩCOSGRWCȑAGCLAWȩRP?BC
MȎ
ȩPCBSAGLEȩRFCȩQGXCȩMDȩRFCȩNGCȩRMȩ@CȩBGQRPG@SRCBȩ $MASQGLEȩ MLJWȩ MLȩ CKNJMWKCLRȩ CȎCARQȩ GELMPCQȩ RFCȩ
Further empirical analysis is necessary to assess potential of AI for economic growth which could
RFCȩCȎCARGTCLCQQȩMDȩRFCQCȩNMJGAGCQ ECLCP?RCȩ @CLCȏRQȩ DMPȩ ?JJȩ 'ȩ ?Qȩ ?ȩ ECLCP?JNSPNMQCȩ
technology can rapidly spread across industry sec-
Despite these negative prospects of technological RMPQȩ?LBȩWGCJBȩQRPMLEȩNMQGRGTCȩEPMURFȩCȎCARQȩ2MȩRFCȩ
progress on equality, there might be some bene- CVRCLRȩRF?Rȩ+*ȩECLCP?RCQȩLCUȩGLQGEFRQȩDPMKȩB?R?
ȩ
ȏRQȩDPMKȩRFCȩNCASJG?PȩDC?RSPCQȩMDȩ'ȩQȩ?ȩECLCP?J it may contribute to the automation of production
purpose technology AI could yield equality en- of new ideas and innovation itself. The ‘productiv-
F?LAGLEȩCȎCARQȩ$MPȩGLQR?LAC
ȩ?AAMPBGLEȩRMȩACKM- GRWȩN?P?BMVȩ@CAMKCQȩKMPCȩ?NN?PCLRȩ?QȩRFCȩP?NGBȩ
glu and Restrepo (2016), AI can create new tasks growth in ML applications does not seem to be re-
which can be performed by high-skilled workers in ȐCARCBȩWCRȩGLȩRFCȩNPMBSARGTGRWȩEPMURFȩQR?RGQRGAQȩ
the short run. Yet, as these tasks become stand-
ardised in the long run, low-skilled workers can also The macroeconomic outcomes of AI are likely to be
@CLCȏRȩDPMKȩRFCKȩ'Lȩ?BBGRGML
ȩ'ȩQFMUQȩNMRCLRG?Jȩ QF?NCBȩ@WȩȏPKȩ@CF?TGMSPȩ'LȩRFGQȩPCQNCAR
ȩRFCȩGQQSCQȩ
to disrupt the spiral of labour market polarisation. of market structure, organisation and sectoral re-
It may be able to perform high-skill tasks that were allocation seem particularly relevant. In perspect-
previously beyond the abilities of technology, such ive, policy interventions might be needed precisely
?QȩRFCȩAJ?QQGȏA?RGMLȩMDȩA?QCȩBMASKCLRQȩDMPȩJ?UWCPQȩ in these areas in order to: i) ensure a competitive
 85 9 Economic perspective

AI as a
general-purpose
technology can
rapidly spread
across industry
sectors and yield
strong positive
growth effects.

stance of the market that incentivises innovation; ii)
avoid that economies of scale in data handling lead
RMȩ CVACQQGTCȩ K?PICRȩ AMLACLRP?RGML
ȩ ?LBȩ GLȩ ECLCP?Jȩ
other competitive advantages stemming from dif-
ferences in national regulations; and iii) ensure that
sectoral reallocation takes place with limited social
AMQRQ
QSEECQRQȩ RF?Rȩ NMJGAGCQȩ QFMSJBȩ ?JQMȩ @Cȩ ?BHSQR?@JCȩ
to local needs.
5FCRFCPȩ MPȩ LMRȩ NMJGAWȩ UGJJȩ @Cȩ CȎCARGTCȩ GLȩ KGRGE?R-
GLEȩRFCȩNMRCLRG?Jȩ?BTCPQCȩCȎCARQȩMDȩ?SRMK?RGMLȩGQȩ
an empirical question. 2FCȧL?RSPCȧMDȧ'ȧ?Qȧ?ȧECL-
ȩCEȩDPMKȩHM@ȩK?PICRȩBGQNJ?ACKCLR CP?JNSPNMQCȧ RCAFLMJMEWȧ MȎCPQȧ QMKCȧ PMMKȧ DMPȧ
FMNCȧRF?RȧRFCȧCȎCARȧMDȧ'ȧMLȧCOS?JGRWȧK?WȧLMRȧ@Cȧ
'ȩ A?Lȩ SLD?TMSP?@JWȩ ?ȎCARȩ RFCȩ BGQRPG@SRGMLȩ MDȩ GL- ?JJȧLCE?RGTC. It can be positive if AI boosts the cre-
come through many channels. The most discussed ?RGMLȩ MDȩ LCUȩ R?QIQȩ DMPȩ JMUQIGJJȩ UMPICPQȩ ?LBȩ CV-
?PCȩ HM@ȩ NMJ?PGQ?RGML
ȩ PCEGML?Jȩ BGȎCPCLACQȩ GLȩ 'ȩ pands its abilities to perform high-skill tasks.
adoption, or regional, sectoral or individual dif-
DCPCLACQȩ GLȩ RFCȩ QNCCBȩ MDȩ ?BHSQRKCLRȩ RMȩ QRPSARSP?J %GTCLȧRFCQCȧSLACPR?GLRGCQ
ȧRFCPCȧGQȧ?ȧLCCBȧDMPȧ?ȧ
labour market changes. Preliminary empirical ev- QRPMLEȧ KMLGRMPGLEȧ DP?KCUMPIȧ RMȧ ?JQMȧ BCRCARȧ ?Rȧ
GBCLACȩQCCKQȩRMȩAMLȏPKȩRFCQCȩAMLACPLQȩ2FCPCȩGQȩ PCEGML?Jȧ?LBȧQS@PCEGML?JȧJCTCJQȧRFCȧNMRCLRG?JȧGK-
?JQMȩ ?ȩ QRPMLEȩ A?QCȩ DMPȩ NMRCLRG?Jȩ GLCȑAGCLAGCQȩ GLȩ N?ARQ of AI deployment and to quickly put in place
the implied labour reallocation process, which are appropriate intervention. We return to this import-
JGICJWȩ RMȩ FSPRȩ QNCAGȏAȩ EPMSNQȩ BGQNPMNMPRGML?RC- ant issue in Chapter 13 on resilience.
ly. It is commonly agreed that policy measures
?PCȩLCACQQ?PWȩRMȩAMSLRCP?ARȩRFCȩLCE?RGTCȩCȎCARQȩ
of technological progress on equality. Strong
within-country variation in technological progress
 10 Cybersecurity Perspective 86

$ZCFSTFDVSJUZ
PQFSBUJPOT
XJMM
CFOFҨU
GSPN
"*
CVU
BMTP
GBDF
OFX
DIBMMFOHFT
GSPN
JU

SUMMARY

'LȩRFGQȩAF?NRCP
ȩUCȩPCȐCARȩMLȩRFCȩRP?LQDMPK?RGMLQȩ?LBȩGKN?ARQȩ@PMSEFRȩ@WȩRFCȩ?BTCLRȩ
MDȩ 'ȩ RMȩ AW@CPQCASPGRWȩ ?LB
ȩ AMLTCPQCJW
ȩ FMUȩ 'ȩ ?JEMPGRFKQȩ RFCKQCJTCQȩ ?PCȩ ?ȎCARCBȩ
by security concerns. When it comes to cybersecurity, it is crucial to be aware of the
fundamental dual nature of AI. It will empower human operators, professionals and
GLTCQRGE?RMPQȩ?LBȩGRȩUGJJȩ@PGLEȩ@CLCȏRQȩRMȩAW@CPQCASPGRWȩ&MUCTCP
ȩRFMQCȩUFMȩQCCIȩRMȩ
?RR?AIȩQWQRCKQȩ?LBȩLCRUMPIQȩ?PCȩCOS?JJWȩCVNCARCBȩRMȩR?ICȩ?BT?LR?ECQȩMDȩ'ȩRMȩA?PPWȩ
out faster and smarter cyber-attacks.

Moreover, AI systems based on ML techniques are often not robust enough against
malicious attacks, introducing a range of new and poorly understood vulnerabilities.
This could pose a danger for the security of the systems using them. To review all
RFCQCȩBCTCJMNKCLRQȩGLȩAMLRCVR
ȩUCȩBCQAPG@Cȩ?ȩP?LECȩMDȩPCACLRȩ'ȩ?NNJGA?RGMLQȩGLȩAW@CP-
QCASPGRW
ȩ MSRJGLCȩ RFCȩ QNCAGȏAȩ TSJLCP?@GJGRGCQȩ MDȩ 'ȩ QWQRCKQȩ ?LBȩ PCJ?RCBȩ ?RR?AIQ
ȩ ?LBȩ
suggest ways of increasing AI robustness and introduce security and safety by design
principles.
 87 10 Cybersecurity Perspective

CYBERSECURITY
PERSPECTIVE
 ?AIEPMSLB'?LBAW@CPQCASPGRW 'TFII?OFKD?BKBȭQP
!W@CPQCASPGRWȩGQȩRFCȩȏCJBȩ@PM?BJWȩAMLACPLCBȩUGRFȩRFCȩ to cybersecurity and
QEBȭDEQ>D>FKPQ
security of digital information, systems and services,
including the security of cyber-physical48 systems. It
encompasses notions such as SFTJMJFODF, EFUFSSFODF
against malicious actors or cybercrime and EFGFODF crime, but will also
(see Chapter 13 for a societal perspective on resili-
CLACȩ 'Lȩ ?ȩ PCACLRȩ HMGLRȩ !MKKSLGA?RGMLȩ RMȩ BCQAPG@Cȩ introduce new
@E>IIBKDBPFKQEBȭBIA
the technological and policy dimensions of cyber-
security, the European Union recognised its emerg-
ing importance for European societies (EC, 2017b).

The application of partly autonomous algorithms amplify the importance of cybersecurity even fur-
in cybersecurity is not entirely new, although trad- ther in the coming years.
itionally those systems were usually not referred
RMȩ?Qȩ~PRGȏAG?Jȩ'LRCJJGECLACȩ!W@CPQCASPGRWȩAMLRPMJQȩ 2FCȩK?GLȩJGLCQȩMDȩBCTCJMNKCLRȩUCȩA?LȩCVNCARȩGLȩRFCȩ
capable of functioning autonomously and taking coming years include:
intelligent decisions to protect information systems
?LBȩ QCPTGACQȩ F?TCȩ CVGQRCBȩ DMPȩ OSGRCȩ QMKCȩ RGKCȩ
They started to play a role many years ago with
the development of smart security applications,
for instance for deciding whether or not to allow
a certain network communication, to autonomous-
JWȩ ȏJRCPȩ QN?Kȩ KCQQ?ECQ
ȩ MPȩ ?B?NRȩ RMȩ LCUȩ AGPASK as well as investigative tools and capabilities
QR?LACQȩQSAFȩ?QȩRFCȩGBCLRGȏA?RGMLȩMDȩNPCTGMSQJWȩSL-
seen forms of cyber-attacks.
1GLACȩRFCL
ȩRFCȩȏCJBȩMDȩAW@CPQCASPGRWȩF?QȩSLBCPEMLCȩ
rapid transformation due to the developments in
ML, deep, and reinforcement learning, which have
resulted in notable successes in addressing com-
puter vision, NLP and autonomous decision-making
• /FX
DBQBCJMJUJFT enabling cybersecurity to do
things that were not possible until recently,
including those increasing the resilience of
systems and services (see Section 10.2 for
more details on some recent developments)
to assist in the deterrence of crime, e.g.
aiding law enforcement in the prosecution
of criminals.
• The integration of AI into widespread digital
products and applications, and into the con-
trols of cybersecurity itself, may introduce
new, and especially DPODFQUVBMMZ
EJҧFSFOU


QCCȩ CEȩ %MMBDCJJMUȩ CRȩ ?J
ȩ ȩ PSLB?ECȩ CRȩ ?J
ȩ
BOE
OPU
ZFU
GVMMZ
VOEFSTUPPE, vulnerabilities
ȩ-QM@?ȩCRȩ?J
ȩȩ+GQQGMLȩ4GJJ?LG
ȩȩ'RȩGQȩ ȩ RF?Rȩ A?Lȩ @Cȩ CVNJMGRCBȩ @W malicious actors
likely that these transformations will continue and (see Section 10.4 for an overview). These
 10 Cybersecurity Perspective 88

vulnerabilities, given the crucial importance of
data for AI systems, also raise the importance
of data protection.
• Challenges to private and public security
from the possible abuse of AI systems to
enhance cyber-attacks and malicious use.
ȩ 'KNMPR?LRȩ CV?KNJCQȩ GLAJSBCȩ GLAPC?QCBȩ ?LBȩ
ȩKCQQ?ECȩȏJRCPGLEȩMPȩGLȩ
ȩ KMPCȩ CVRCLBCBȩ AW@CP?RR?AIQ
ȩ KMPCȩ BGȑ assisting human security operators (e.g. Crawford
ȩ ASJRGCQȩRMȩ?RRPG@SRCȩ?RR?AIQȩRMȩ?ȩQNCAGȏAȩ?ARMP
et al., 2015; Radford 2018; Meidan 2018). In all
the targeting of human vulnerabilities
through autonomous social engineering, so-
cial media and propaganda manipulation,
the production of credible fake content in
With recent successes in DL, we are starting to
see a number of AI methods based on neural net-
works and DL gradually moving from research into
workable cybersecurity applications. They employ
deep neural network architectures and big data to
devise increasingly accurate cybersecurity learning
systems, e.g. for anomaly detection in malware and
AMKNSRCPȩLCRUMPIȩ?L?JWQGQ
likelihood, these trends will continue, but will also
come hand in hand with a proliferation of AI-backed
cyber-attacks (Brundage, 2018).
news or reporting, attacks on cyber-physical
systems such as autonomous vehicles, or the
development of autonomous weapon sys-
tems (see, e.g. Brundage 2018).
1 "CRCPPCLAC?LBȍEFR?E?GLQRAPGKC
Recent developments strengthen the resilience of
cybersecurity systems, reduce risks and increase
their capability to withstand threads and malicious
1 NNJGA?RGMLQMD'GLAW@CPQCASPGRW intent reliably. Another element is that research
into AI in cybersecurity systems plays a crucial part
#?PJWȩ CV?KNJCQȩ MDȩ 'ȩ QWQRCKQȩ ?NNJGCBȩ GLȩ AW@CP in developing cyber deterrence capabilities. Firstly,
security include systems for network intrusion AI systems are being and will increasingly be used
detection to identify previously unknown cyber- by cyber criminals (e.g. Brundage et al., 2018), and
?RR?AIQȩ .?VML
ȩ ȩ ?LBȩ ?LRGTGPSQȩ ?NNPM?AFCQȩ thus research into their capabilities and weakness-
RMȩBCRCARȩLCUȩK?JU?PCȩ1?LMI
ȩȩ2FCȩCVNJM- es (see Section 10.4) will play a crucial part in de-
sion in the practical use of ML in the early 2000s fending against such malicious usage of AI. Sec-
@PMSEFRȩDSPRFCPȩR?LEG@JCȩ@CLCȏRQȩRMȩRFCȩAW@CPQCAS- ondly, of course, law enforcement will increasingly
PGRWȩBMK?GLȩ2FCȩKMQRȩQRPGIGLEȩCV?KNJCQȩUCPCȩNPM@- engage in active usage of AI systems to reinforce
?@JWȩRFCȩȏPQRȩPM@SQRȩ?B?NRGTCȩQWQRCKQȩRMȩBCRCARȩ?LBȩ investigative capabilities and to strengthen digital
ȏJRCPȩQN?Kȩ?LBȩK?JGAGMSQȩKCQQ?ECQȩ?SRMLMKMSQJWȩ evidence-making in court. Most developments aim
QCCȩ?JQMȩ MVȩȩGLȩ!F?NRCPȩ at leveraging new tools for three main operative
EM?JQ
ȩRFCȩJMA?JGQ?RGMLȩMDȩ?ȩAPGKC
ȩRFCȩGBCLRGȏA?RGMLȩ
Most AI-based methods currently in use in opera- of victims and perpetrators, and the determination
tional controls are based on ML approaches, using a of the content of the crime.
FMQRȩMDȩUCJJRCQRCBȩ?JEMPGRFKQȩ?LBȩCVNCPRȩQAFCKCQȩ
and are usually embedded within a traditional cy- 2FCȩKMQRȩNPCT?JCLRȩCV?KNJCQȩA?Lȩ@CȩDMSLBȩGLȩBGEG
@CPQCASPGRWȩ?PAFGRCARSPCȩDMPȩ?ȩPCACLRȩCV?KNJCȩQCCȩ tal forensics and in the employment of biometric
IEEE 2017). They are largely TVQFSWJTFE
MFBSOJOH
QWQRCKQȩDMPȩGLTCQRGE?RGTCȩNSPNMQCQȩ2FCȩUGBCȩȏCJBȩ
methods, i.e. AI systems that evolve their models MDȩBGEGR?JȩDMPCLQGAQȩCLAMKN?QQCQȩTCPWȩBGȎCPCLRȩ?N-
SQGLEȩ?T?GJ?@JCȩJ?@CJJCBȩB?R?ȩ?QȩCV?KNJCQȩ2WNGA?J- plications, ranging from computer and hard-drive
JW
ȩRFCWȩ?PCȩNCPDMPKGLEȩAJ?QQGȏA?RGMLȩMDȩKCQQ?ECQ
ȩ forensics to fraud detection in large databases. The
digital events or computer code into acceptable or spread of online services and digital devices has
non-acceptable classes. These AI-based controls led to a proliferation of digital content relevant to
are capable of limited forms of adaptation, mostly the prosecution of criminal activities. In this con-
through retraining on new samples. RCVR
ȩ'ȩQWQRCKQ
ȩDMPȩGLQR?LACȩDMPȩGK?EC
ȩ?SBGMȩ?LBȩ
 89 10 Cybersecurity Perspective

RCVRȩ ?L?JWQGQ
ȩ MPȩ DMPȩ SLQSNCPTGQCBȩ ?L?JWQGQȩ MDȩ SL- use secure design principles, have a clear view of
ordered data volumes, have the potential to assist where vulnerabilities are likely to occur, and put in
law enforcement, not only in the investigation of
ȩ'ȩQWQ-
cyber-dependent and cyber-enabled crime, but also
ȩ CQNCAG?J-
in traditional forms of crime where digital content
can play a key role in the investigation.
NJ?ACȩCȎCARGTCȩQCASPGRWȩAMLRPMJQȩ'LȩAMLRP?QR
RCKQȩ ?PCȩ DSLB?KCLR?JJWȩ KMPCȩ AMKNJCV
ly if they are based on ML architecture. They are
composed of a possibly non-linear feedback loop
system between a human-designed algorithm and
1 0M@SQRLCQQMD'?JEMPGRFKQ?E?GLQR
 K?JGAGMSQ?ARGML
With an increasing number of AI systems employed
in cybersecurity, not only are we making progress
but we are also introducing new vulnerabilities that
then open the window for new types of attacks. For
this reason, these AI systems are now coming un-
der increased scrutiny.
In classical software development, the developer
uses a computer language to create the program
that implements a certain algorithm. This consti-
tutes a straight development process where we can
Figure 12. Paradigm shift in the cybersecurity of systems following the introduction of AI components
a human (or even AI) assembled dataset. Together,
these constitute a learning model that acts as the
actual ‘AI’ and is basically a new algorithm. These
AI models are then embedded into a program, of-
ten combined with other pieces of code or software
?LBȩNMQQG@JWȩGKNJCKCLRCBȩUGRFȩ?ȩFMQRȩMDȩBGȎCPCLRȩ
programming tools. The situation is depicted in
Figure 12.
LWȩTSJLCP?@GJGRWȩMPȩBCQGELȩȐ?UȩGLȩRFGQȩAMKNJCVȩQWQ-
RCKȩUGJJȩJGICJWȩ@CȩMDȩ?ȩTCPWȩBGȎCPCLRȩL?RSPCȩRF?LȩGLȩ
classical systems. Tracing an event and its cause,
such as an attack on the system and the vulner-
?@GJGRWȩ CVNJMGRCBȩ @CAMKCQȩ ?ȩ KSAFȩ KMPCȩ AMKNJCVȩ
 10 Cybersecurity Perspective 90

and demanding task (see Fig. 12). Moreover, once ȩȩ RMȩ@CȩKGQAJ?QQGȏCBȩMPȩRMȩDMMJȩRFCȩ?JEMPGRFKGAȩ

the AI system is embedded into parts of a classi- procedure in which the model is being used.
A?Jȩ QWQRCKȩ ?LBȩ BGȎCPCLRȩ AMK@GL?RGMLQȩ MDȩ MJBȩ ?LBȩ In Figure 12, this type of vulnerability would
new types of vulnerability come together we may conceptually be introduced in the right part of
CLBȩ SNȩ UGRFȩ LCUȩ M@QR?AJCQȩ ?JRMECRFCPȩ ȩ BGȎCPCLRȩ
?LEJCȩDPMKȩUFGAFȩRMȩQCCȩRFCȩAMKNJCVGRWȩNPM@JCKQȩ
BGQNJ?WCBȩ GLȩ $GESPCȩ ȩ GQȩ RF?Rȩ MDȩ ?JEMPGRFKGAȩ CV-
NJ?GL?@GJGRW
ȩ MPȩ GLRCPNPCR?@GJGRWȩ RMȩ UF?Rȩ CVRCLRȩ ?PCȩ
the results of a given algorithm still understand-
?@JCȩRMȩ?ȩFSK?LȩMPȩ?PCȩCTCLȩSLGOSCJWȩCVNJ?GL?@JCȩ
(e.g. because of non-linear functions employed in
K?LWȩ+*ȩKMBCJQ likelihood model that is known to be suscep-
the AI system, after the ‘data’ and the ‘model’
have interacted.
• %FTJHO
ҩBXT in the underlying mathematical
procedures that optimise the learning pro-
cess of the model, leading to model attacks.
For instance, the usage of a loss function or
tible to certain types of input could create an
#VNJ?GL?@GJGRWȩGQȩ?ȩICWȩ?PC?ȩMDȩDSLB?KCLR?JȩPCQC?PAFȩ
@CA?SQCȩ UCȩ LCCBȩ CVNJ?GL?@JCȩ ?LBȩ GLRCPNPCR?@JCȩ
models to increase the capabilities of AI systems
to abstract and transfer knowledge, understand
their robustness in terms of cybersecurity, assess
possible bias, and crucially, build and retain trust
in society (see Chapters 2 and 6 for a discussion
ȩ C?QWȩN?RFȩDMPȩ?BTCPQ?PG?JȩCV?KNJCQȩ
Moreover, insecure systems might allow
a malicious user to steal the mathematical
structure of a model. In Figure 12, this type
of vulnerability would conceptually be intro-
duced in the ‘model’ part of the AI system.
MLȩD?GPLCQQȩ?LBȩ?AAMSLR?@GJGRWȩGLȩRFCȩAMLRCVRȩMDȩ' There is growing evidence of how easily these vul-
LCP?@GJGRGCQȩ A?Lȩ @Cȩ CVNJMGRCBȩ GLȩ PC?Jȩ ?NNJGA?RGMLQ
ȩ
2FCȩCKCPEGLEȩȏCJBQȩMDȩ'ȩQ?DCRWȩ?LBȩ?BTCPQ?PG?Jȩ+*ȩ especially in ML-based models. Thus, designing
are actively concerned with these new challenges. 'QNCAGȏAȩ QCASPGRWȩ AMLRPMJQȩ GQȩ ?Lȩ ?ARGTC
ȩ ?J@CGRȩ
The former refers to the research branch concerned WMSLE
ȩ ȏCJBȩ MDȩ PCQC?PAFȩ 0CQC?PAFȩ MLȩ ?BTCPQ?PG?Jȩ
with designing safe AI systems from the start (see, ML currently focusses on two main measures de-
e.g. Amodei, 2016). The latter is concerned with the pending on circumstances, such as the intention of
LCUȩ RWNCQȩ MDȩ TSJLCP?@GJGRGCQȩ CVFG@GRCBȩ @Wȩ RFCȩ KMQRȩ the attack and the type of vulnerability (e.g. Paper-
prevalent AI systems employing ML (see, e.g. Huang not et al., 2015):
et al., 2011; Szegedy at al., 2013). Typical vulnera-
@GJGRGCQ
ȩCVAJSQGTCȩRMȩ'ȩQWQRCKQ
ȩGLAJSBCȩRFCȩDMJJMUGLEȩ • %BUB
TBOJUJTBUJPO: cleaning the training data
(see, e.g. Biggio et al., 2012; Szegedy et al., 2013; of all potentially malicious content before
Huang et al., 2017): training the model. In most realistic cases,
another AI system is employed to act as a
• %BUB
QPJTPOJOH, the deliberate introduction ȩ ȏJRCP
ȩ @SRȩ GLȩ TCPWȩ APGRGA?Jȩ AGPASKQR?LACQȩ
of false data in the training of the model, human intervention might be inevitable.
especially dangerous for autonomous AI
systems that employ online re-learning pro- • 3PCVTU
MFBSOJOH: redesigning the learning
ȩ ACBSPCQȩ2FGQȩCȎCARGTCJWȩKC?LQȩRF?RȩRFCȩB?R? procedure to be robust against malicious
ȩ GRQCJDȩ@CAMKCQȩ?ȩK?HMPȩTSJLCP?@GJGRWȩMDȩQSAF ȩ ?ARGML
ȩCQNCAG?JJWȩ?BTCPQ?PG?JȩCV?KNJCQȩ2FGQ
systems. In Figure 12, this type of vulnera-
bility would conceptually be introduced in the
ȩ MPȩ PCBCQGELGLEȩ RFC
‘data’ part of the AI system.
• "EWFSTBSJBM
FYBNQMFT, input in the trained
ML system, which is deliberately designed
ȩ CGRFCPȩCLR?GJQȩCVNJGAGRȩRP?GLGLEȩ?E?GLQRȩILMUL
ȩ ?BTCPQ?PG?Jȩ CV?KNJCQ
mathematical foundation of the algorithms
and employing techniques from statistics,
known as regularisation and robust infe-
rence. This approach actually follows the
 91 10 Cybersecurity Perspective

well-known security-by-design principle, i.e.

taking the security of a software or appli-

AI will not cation into account from the beginning of the

design process. It should be noted that for

replace humans ML based AI systems this approach usually

implies a degradation of performance.

in cybersecurity 1SKK?PW?LBAMLAJSQGMLQ
anytime soon 2FCȩ?BTCLRȩMDȩRFCȩ'ȩPCTMJSRGMLȩGQȩCVNCARCBȩRMȩF?TCȩ
?LȩGLAPC?QGLEȩGKN?ARȩMLȩAW@CPQCASPGRWȩ2FCȩCVGQRGLEȩ
but rather J?LBQA?NCȩUGJJȩ@CȩQS@HCARȩRMȩK?LWȩRP?LQDMPK?RGMLQȩ
GLȩRCPKQȩMDȩKMPCȩCȎCARGTCȩ?LBȩCȑAGCLRȩAW@CPȩBC-
empower them. fences, but at the same time more sophisticated
cyber-attacks. Key conclusions from these devel-

It is important opments are:

'ȧUGJJȧLMRȧPCNJ?ACȧFSK?LQȧGLȧAW@CPQCASPGRWȧ?LW
to understand RGKCȧ QMMLȧ @SRȧ P?RFCPȧ CKNMUCPȧ RFCKȩ #VNCAR
ations about AI are high, and for good reasons,
the limits of the new EGTCLȩ RFCȩ J?RCQRȩ ?BT?LACQȩ GLȩ RFCȩ ȏCJBȩ ?LBȩ RFCȩ
growing number of successful cases of its appli-
generation off AI cation in several domains, including cybersecu-
rity. However, it is important to understand the
algorithms. limits of the new generation of AI algorithms.
 10 Cybersecurity Perspective 92

We have summarised the key issues above, and
we clearly still need human operators to com-
plement the capabilities of AI for cybersecurity.
ȩQNCAGȏAȩRMȩ+*ȩ?JEM-
'ȧ UGJJȧ @PGLEȧ AJC?Pȧ @CLCȏRQȧ RMȧ AW@CPQCASPGRWȧ ?LBȧ
RFCȧ ȏEFRȧ ?E?GLQRȧ APGKC
ȧ @SRȧ UGJJȧ ?JQMȧ GLRPMBSACȧ UFGAFȩQNCAGȏAȩMPE?LGQ?RGML?Jȩ?LBȩRCAFLGA?JȩAMLRPMJQȩ
LCUȧAW@CPQCASPGRWȧAF?JJCLECQ. ML algorithms have
altered the classical paradigm of software devel-
opment. When using them, human developers no
longer program an algorithm to tell the comput-
ȩ QCASPGRWȩ ?SBGRQȩ ?LBȩ AW@CPQCASPGRWȩ
er IPX to solve a given problem but instead they
program it to MFBSO
IPX to solve the problem. As a
result, the AI system behaves as a new algorithm,
not directly programmed by the human developer:
this is the result of a learning process from training
data carried out by the ML algorithm originally de-
veloped by the human. Recently, it has been found
RF?R
ȩBCQNGRCȩRFCGPȩCȎCARGTCLCQQ
ȩ'ȧQWQRCKQȧ@?QCBȧ ties. This is of particular relevance when AI is used
MLȧ +*ȧ RCAFLGOSCQȧ ?PCȧ MDRCLȧ LMRȧ PM@SQRȧ ?E?GLQRȧ
K?JGAGMSQȧ ?RR?AIQȩ RF?Rȩ CVNJMGRȩ RFCȩ QNCAGȏAȩ U?Wȩ GLȩ
ȩJMA?JGQ?RGMLȩ?LBȩNPMQCASRGMLȩMDȩ
which the AI system behaves by presenting it with
specially crafted input, manipulating the data that
is used to train it, or abusing biases in the statistical
algorithms used to perform the training. These are
LCUȩAJ?QQCQȩMDȩTSJLCP?@GJGRGCQ
rithms, which need to be further understood and for
need to designed and implemented.
Secure software development practices, security
ACPRGȏA?RGMLQ
controls need to consider these new kinds of vul-
LCP?@GJGRGCQȩ QNCAGȏAȩ RMȩ RFCȩ 'ȩ AMKNMLCLRQȩ MDȩ RFCȩ
digital systems. Further research is needed in the
CKCPEGLEȩȏCJBȩMDȩ?BTCPQ?PG?Jȩ+*ȩGLȩMPBCPȩRMȩ@CRRCPȩ
understand the limitations in the robustness of ML
algorithms against malicious action and design
CȎCARGTCȩ QRP?RCEGCQ to mitigate these vulnerabili-
in cybersecurity controls or to assist human actors
GLȩRFCȩGBCLRGȏA?RGML
cybercrime.
 93 10 Cybersecurity Perspective

2FCPCȧGQȧ?ȧLCCBȧRMȧQCASPCȧRFCȧ?T?GJ?@GJGRWȧMDȧSN NS@JGAȩPCQC?PAFȩGQȩAMLBSARCBȩGLȩRFCȩȏCJBȩMDȩ'ȩ?LBȩ
RMB?RCȧ FGEFOS?JGRWȧ B?R?QCRQȧ GLȧ RFCȧ BMK?GLȧ MDȧ AW@CPQCASPGRW
ȩ?JQMȩAMLQGBCPGLEȩRFCȩMȎCLQGTCȩQGBC
ȩRMȩ
AW@CPQCASPGRWȧ The latest techniques in ML such study emerging and potentially AI-powered cyber-
as DL or reinforced learning can produce impres- attacks and devise corresponding countermeasures.
sive results but require huge quantities of train-
GLEȩB?R?ȩRMȩUMPIȩ?LBȩ@CȩCȎCARGTCȩQȩGQȩRFCȩA?QCȩGLȩ Furthermore, the research landscape on AI is al-
other domains, there is a scarcity of large enough ready highly dominated by the USA and China (see
FGEFOS?JGRWȩB?R?QCRQȩGLȩRFCȩȏCJBȩMDȩAW@CPQCASPGRWȩ Chapter 3) and the ongoing general drain of re-
2FCȩ ?T?GJ?@GJGRWȩ MDȩ QSAFȩ B?R?QCRQȩ RMȩ RFCȩ QAGCLRGȏAȩ search and researchers from Europe needs to be
community is crucial to ensure future advances in PCTCPQCB
ȩCQNCAG?JJWȩGLȩȏCJBQȩ?JPC?BWȩJ?AIGLEȩILMU
RFCȩȏCJB JCBEC?@JCȩ CVNCPRQȩ GLȩ RFCȩ K?PICR
ȩ QSAFȩ ?Qȩ AW@CP-
security and AI. Focusing research on the ethical
Moreover, the creation of up-to-date reference use of AI and the role of regulation can become an
datasets in the several areas of the cybersecur- asset to strengthen European leadership in these
ity domain can also boost research and innovation ?PC?Q
ȩ?LBȩPCR?GLȩCVNCPRGQCȩGLȩ'ȩ?LBȩAW@CPQCASPGRW
@WȩCL?@JGLEȩ?EGJCȩ?LBȩCȎCARGTCȩ@CLAFK?PIGLEȩ?LBȩ
evaluation of state of the art AI techniques. 'ȧ GLȧ AW@CPQCASPGRWȧ PCOSGPCQȧ ?ȧ KSJRGBGQAGNJGL?PWȧ
?NNPM?AFȩ-LȩRFCȩFGEFJWȩRCAFLGA?JȩK?RRCPQȩRF?RȩQSP-
The dual nature of AI.ȧ 'Rȧ GQȧ APSAG?Jȧ RMȧ @Cȧ ?U?PCȧ round the application of AI in cybersecurity, strong
MDȧ RFCȧ DSLB?KCLR?Jȧ BS?Jȧ L?RSPCȧ MDȧ RCAFLGOSCQ
ȧ interactions of policymakers with researchers and
RMMJQȧ ?LBȧ QWQRCKQȧ CKNJMWCBȧ GLȧ AW@CPQCASPGRWȧ CLEGLCCPQȩQCCKȩGKKCLQCJWȩGKNMPR?LRȩ2FCȩȏCJBQȩMDȩ
?LBȧ', especially when both have to be considered AI and cybersecurity are rapidly changing, and re-
together. AI can improve security, but cybercrim- ports will quickly become outdated, making a con-
GL?JQȩ?PCȩCOS?JJWȩCVNCARCBȩRMȩR?ICȩ?BT?LR?ECȩMDȩRFCȩ QR?LRȩ GLDMPK?RGMLȩ ȐMUȩ DPMKȩ RFCȩ RCAFLGA?Jȩ JCTCJȩ RMȩ
latest advances in AI techniques to carry out faster BCAGQGMLȩ@MBGCQȩ?JKMQRȩK?LB?RMPWȩQȩ?LȩCVRCLQGMLȩ
and smarter cyber-attacks. of this, we also need to foster a culture of cross-
collaboration between AI and cybersecurity devel-
This situation does not necessarily require restrict- opers and, on a broader level, researchers from
GLEȩ PCQC?PAFȩ Lȩ CȎCARGTCȩ BCDCLACȩ A?Lȩ MLJWȩ @Cȩ MRFCPȩPCJ?RCBȩȏCJBQȩQSAFȩ?QȩNFGJMQMNFW
ȩCAMLMKGAQȩ
designed knowing how attacks might take place. or social science.
Therefore, it is important to ensure that further
 11Computer Processing and Energy Perspective 94

*OUFOTJWF
DPNQVUJOH
OFFET
UP
CF
FOFSHZ
FҪDJFOU

SUMMARY

In this chapter, we analyse one of the often-ignored aspects of AI, i.e. that the opportunities
created by great data volumes and advanced analytics come at an increasing environmen-
R?J
ȩCAMLMKGA
ȩ?LBȩECMNMJGRGA?JȩAMQRȩCLCPEWȩ2FCȩGLRPMBSARGMLȩMDȩRFCȩ'M2ȩ?LBȩ%ȩLCRUMPIQȩUGJJȩ
GLAPC?QCȩRFCȩNPM@JCKȩQGELGȏA?LRJWȩ#LCPEWȩGQȩPCAMELGQCBȩ?Qȩ?ȩK?GLȩBCRCPKGL?LRȩMDȩRFCȩJMLE
term sustainability of AI. Therefore, we need to take early action to address this aspect of the
development of AI and its related technologies. While some actions can be done using today’s
RCAFLMJMEW
ȩKMPCȩNMUCPPCBSARGMLȩGLLMT?RGMLȩGQȩLCCBCBȩRMȩDSJȏJȩDSRSPCȩBCK?LBȩ'RȩGQȩLCACQQ?PWȩ
to introduce advanced technologies in the hardware, software, and services areas, to develop
intensive-computing ecosystems that are ‘green’.

These ecosystem solutions must face four important energy-consumption areas: reduce
RFCȩ CLCPEWȩ AMLQSKCBȩ @Wȩ RFCȩ CVGQRGLEȩ @GEȩ B?R?ȩ ACLRPCQȩ JGKGRȩ B?R?ȩ KMTCKCLRȩ PCBSAGLEȩ RFCȩ
energy consumed from machine-to-machine communications; reduce the embodied energy;
PCBSACȩRFCȩM@QMJCQACLACȩMDȩBGEGR?JȩRCAFLMJMEGCQȩQCC
ȩDMPȩCV?KNJC
ȩRFCȩQFMPRȩJGDCQN?LȩMDȩKM@GJCȩ
phones, replaced every one to three years), which is perhaps the main energy consumption
voice for digital technology.
 95 11Computer Processing and Energy Perspective

COMPUTER
PROCESSING
AND ENERGY
PERSPECTIVE
 'LRPMBSARGML There is a paradigm
There are several studies (SIA, 2015; Mayer 2009;
IEE, 2012) showing that the digital economy is
PEFȱFK@LJMRQFKD
facing an increasing problem: a severe mismatch
between the processing and storage needs of the
to bring processing
escalating volumes of data available, and a sus-
tainable energy footprint. A report prepared for
at the level of smart
%PCCLNC?ACȩȩAJ?GKCBȩRF?RȩGDȩRFCȩAJMSBȩUCPCȩ
?ȩ AMSLRPW
ȩ GRȩ UMSJBȩ F?TCȩ RFCȩ ȏDRFȩ J?PECQRȩ CLCPEWȩ
sensors, opening
demand in the world, while Vidal (2017) suggested
RF?RȩRFCȩB?R?ȩRQSL?KGȩAMSJBȩAMLQSKCȩMLCȩȏDRFȩMDȩ
an opportunity for
global electricity by 2025.
European investment,
2FCQCȩCQRGK?RCQȩK?W@CȩQMKCUF?RȩMȎȩRFCȩK?PI
ȩ@SRȩ
there seems to be a consensus that B?R?ȧEPMURFȧ
regulatory
UGJJȧ J?PECJWȧ MSRN?ACȧ DMPCQCC?@JCȧ GKNPMTCKCLRQȧ
GLȧAMKNSR?RGML?JȧNMUCPȧ?LBȧRFCȧ?QQMAG?RCBȧCL-
frameworks, and
CPEWȧ DMMRNPGLRȩ $MPȩ CV?KNJC
ȩ @CRUCCLȩ ȩ ?LBȩ

ȩB?R?ȩRP?ȑAȩEPCUȩDMJB
ȩAMKN?PCBȩRMȩ?ȩAMP-
standards to shape
responding 16-fold increase in computing power
(largely tracking Moore’s law) as shown in Figure
these developments.
13 (Short et al., 2011; and Kambatla et al., 2014).
2FCȩ BCTCJMNKCLRȩ MDȩ RFCȩ 'M2ȩ GQȩ CVNCARCBȩ RMȩ UGBCLȩ
ȩUCȩ
the fork between data and processing power in the
coming years with an increasing energy bill. There-
fore, the computing and energy consumption gen-
erated by AI and IoT (including devices, networks,
and data centres to which they are connected)
KSQRȩ@CȩMNRGKGQCBȩQGELGȏA?LRJWȩ'LȩRFGQȩAF?NRCP
CVNJMPCȩRFPCCȩK?GLȩ?NNPM?AFCQȩRMȩPCBSACȩRFCȩCLCP-
gy footprint after reporting in more detail the en-
ergy challenges of one important part of the data
chain: data centres.
 11Computer Processing and Energy Perspective 96

Figure 13. Projection of data and computing growths (logarithmic scale)

Source: JRC based on Kambatla et al., 2014

 QQCQQKCLRMDB?R?ACLRPC"!
 CLCPEWAMLQSKNRGML
AI requires a considerable amount of data and
processing power, which often implies the need
for dedicated facilities such as data centres (DC).
DCs host IT equipment for data storage and pro-
ACQQGLEȩ CK?GJ
ȩ ȏL?LAG?Jȩ RP?LQ?ARGMLQ
ȩ QMAG?Jȩ KC- consumption.51 In Europe, the number of DC has
dia, computing, etc.) including computers (servers)
and other components, such as communication and
storage equipment. Normally the DC is hosted in
a controlled environment requiring air cooling and
AMLBGRGMLGLE
ȩGLȩMPBCPȩRMȩCVRP?ARȩRFCȩFC?RȩECLCP?RCBȩ
in processors. In addition, power has to be distrib-
uted to the IT equipment, including conversion from
AC to DC, battery storage and uninterruptable pow-
er supplies (UPS) to cope with very short power dis-
ruptions. There is a high level of redundancy in DCs
for both power distribution and cooling to increase
power availability and reliability, including provision
of annual global electricity consumption, i.e. more
than the electricity consumption of several EU
Member States.
DC energy consumption in the USA (largest mar-
ket in the world) has been estimated at 70 TWh,
representing, about 1.8 % of total US electricity
dramatically increased in the last 10 years, particu-
larly with the development of cloud computing. DC
energy consumption in the EU (second or third mar-
ket in the world) was estimated in 2014-2015 to be
at between 43 and 49 TWh i.e. between 1.5 % and
1.8 % of EU total consumption. This is a calculation
@WȩRFCȩ(0!ȩ@?QCBȩMLȩ?Lȩ?QQCQQKCLRȩMDȩL?RGML?JȩCV-
perts and published reports.
These estimates do not include the energy consump-
tion to transfer the data from one DC to another or
to the end-users. This is part of the broadband op-
for standby generation (usually diesel-powered). CP?RMPQȩAMLQSKNRGMLȩCEȩRCJCAMKQ
ȩMTCPȩȏ@PC
ȩ"1*
ȩ
These redundancies and the need for a controlled A?@JCȩ MPȩ KM@GJC%ȩ ?LBȩ CLBSQCPQȩ CEȩ KMBCKQ
ȩ
environment are typically doubling the power and PMSRCPQ
ȩ QUGRAFCQȩ ?Rȩ FMKCȩ ?LBȩ MȑACQȩ 2FCȩ B?R?ȩ
energy consumption of the IT part (i.e. twice the transmission energy consumption is comparable to
IT power is required by a typical data centre).49 In the total DC energy consumption. JRMECRFCP
ȧRFCPC-
CVRPCKCJWȩCȑAGCLRȩ"!QȩRFGQȩ?BBGRGML?JȩAMLQSKNRGMLȩ DMPC
ȧ"!Qȧ?LBȧB?R?ȧRP?LQKGQQGMLȧAMSJBȧ?AAMSLRȧDMPȧ
is only 10 %. ȧϡȧMDȧ?JJȧ#3ȧNMUCPȧAMLQSKNRGMLȧ

The global energy consumption of DC was esti- DC consolidation, outsourcing and cloud computing
mated at 194 TWh in 201450, which is around 1 % ?PCȩFCJNGLEȩRMȩICCNȩCLCPEWȩAMLQSKNRGMLȩGLȩ"!QȩȐ?R
ȩ
 97 11Computer Processing and Energy Perspective

notwithstanding the increase of data and process- which is used for ML applications such as
GLE
ȩ ?Qȩ J?PECPȩ "!Qȩ RCLBȩ RMȩ @Cȩ KMPCȩ CȑAGCLRJWȩ BC- ȩ LCSP?Jȩ LCRUMPIQȩ -RFCPȩ 'ȩ ?AACJCP?RMPȩ BC
signed and managed. signs are starting to appear from other
ȩ TCLBMPQȩ?GKCB
ȩDMPȩCV?KNJC
ȩ?RȩRFCȩPM@MR
 -NRGMLQRMGKNPMTCRFCCLCPEW ics markets.
 CȏAGCLAWMDRFCGLAPC?QGLE
 BCK?LBDMP&.! Ȍ
"EWBODFE
DMPVE
NVMUJDPSF

DIJQT that com-
ȩ @GLCȩRFCȩ?BT?LR?ECQȩMDȩ!.3QȩUGRFȩ%.%.3Q

2FCPCȩ ?PCȩ BGȎCPCLRȩ ?NNPM?AFCQȩ RMȩ ?BBPCQQȩ RFGQȩ and specialised AI chip. This new breed of
problem: chips promises to reduce ten-fold the
times the processing power per watt (e.g.
(a) advance CPU capabilities to process big ȩ RFPMSEFȩ ?ȩ ȏLCEP?GLCBȩ NMUCPȩ K?L?EC
ȩ B?R?
ȩ PSLȩ '
+*ȩ HM@Q
ȩ ?LBȩ AMLQSKCȩ JCQQȩ ment) to be capable of running the most
energy; ȩ AMKNJCVȩAMKNSRGLEȩR?QIQ
(b) improve the energy consumption of DCs
necessary to cool the systems; Ȍ
&NCFEEFE
QSPDFTTPST for smart sensors
(c) move the intelligence from the centre to by using new microcontrollers designed for
the edge by optimising (cloud-based) com- AI, which are the brains of billions of ‘smart
puting infrastructures and bring AI closer to things’. Trained DL models are com-
the data. piled to generate optimised neural net-
work code that is embedded in advanced
!.3?BT?LACKCLRQCLCPEWQ?TGLE
 ?LBN?P?JJCJGQ?RGMLAMKNSRGLE
With respect to CPU advancement, there are sev-
eral lines of action:
• (SBQIJDT
1SPDFTTJOH
6OJU
(16
: a specialised
electronic circuit designed to accelerate
the creation of images intended for output
microcontrollers. Embedded processors are
usually simpler in design and have minimal
power requirements. In addition, moving the
computation from the centre to the edge of
the network contributes to reducing the
energy footprint of the IoT.
Ȍ
/FVSPNPSQIJD
DPNQVUJOH: a technology
based on the use of very-large-scale inte-
ȩ RMȩ?ȩBGQNJ?WȩBCTGACȩ%.3Qȩ?PCȩSQCBȩGLȩKM@GJCȩ gration systems that mimic neuro-biologic-
phones, personal computers, workstations, al structures present in the nervous system.
ȩ ?LBȩE?KCȩAMLQMJCQȩ+MBCPLȩ%.3Qȩ?PCȩTCPWȩ
ȩ CȑAGCLRȩ?RȩK?LGNSJ?RGLEȩAMKNSRCPȩEP?NFGAQ
and image processing, and their highly
ȩ N?P?JJCJȩQRPSARSPCȩK?ICQȩRFCKȩKMPCȩCȑAGCLRȩ
than general-purpose Central Processing
Units (CPUs) for algorithms where the pro-
cessing of large blocks of data is done in
parallel, as for ML and DL.
Ȍ
"*
BDDFMFSBUPST

UIF
5FOTPS
1SPDFTTJOH
6OJU
conserve energy.

516
ȩ?Lȩ'ȩ?AACJCP?RMPȩBCTCJMNCBȩ@Wȩ%MM
Recent solutions focus on models of neural
systems for perception, motor control, or
multisensory integration.
Ȍ
3FWFSTJCMF
DPNQVUJOH: a model of computing
that tries to minimise the amount of energy
dissipated as heat during processing. They
are based on low power circuits, known
as ‘adiabatic’ circuits, that are designed to
ȩ EJCȩ QNCAGȏA?JJWȩ DMPȩ LCSP?Jȩ LCRUMPIȩ +*ȩ Ȍ
2VBOUVN
DPNQVUJOH: a form of computing
ȩ 2FCȩAFGNȩF?Qȩ@CCLȩBCQGELCBȩRMȩSQCȩ%MMEJCQ ȩ CVNJMGRGLEȩQMKCȩMDȩRFCȩNFWQGA?JȩNPMNCPRGCQȩMD
TensorFlow framework, an open math library quantum mechanics. Classic computers
 11Computer Processing and Energy Perspective 98

store information in bits that can either be 0

or 1. Qubits, which are the quantum equi-
valent of transistors, are capable of ‘super
position’, i.e. they can be both 0 and 1 at the
same time, and are therefore able to store
The expected
a lot more information. ‘Entanglement’ is
the other key property, which means that
growth in data,
ȩ
qubits are connected to each other, so that
MNCP?RGLEȩ MLȩ MLCȩ ?ȎCARQȩ RFMQCȩ RF?Rȩ ?PCȩ
computing
connected. This property allows for very fast
parallel processing. Quantum computing is
processing,
already used for cryptography, but other
applications are starting to emerge.
and energy
'LLMT?RGTC?LBKMPCCȏAGCLR consumption
 AMMJGLEQWQRCKQ
CLEGLCCPGLE
 QMJSRGMLQ is changing the
&ҪDJFOU
DPPMJOHȩRFCȩȏPQRȩQMJSRGMLȩGQȩRMȩP?GQCȩRFCȩ"!ȩ
temperature so that there is no need to over-cool
present paradigm.
servers and other equipment. Another is to use
fresh outdoor air (free cooling, direct or indirect) or
cold water provided by lakes, rivers, the sea, or un-
BCPEPMSLBȩ ?OSGDCPQȩ RFCȩ ȏL?Jȩ AMMJGLEȩ RMȩ RFCȩ P?AIQȩ
GQȩ QRGJJȩ NPMTGBCBȩ @Wȩ ?GPȩ RFPMSEFȩ ?ȩ FC?Rȩ CVAF?LECPȩ
Finally, ‘adiabatic’ or evaporative cooling can also 'LLMT?RGTCGLDP?QRPSARSPCQ

be a good alternative to compressor-based cooling.  ?PAFGRCARSPCQ?LBAMLȍESP?RGMLQ
There are an increasing number of DCs with no-
compressor-based cooling system, running all the AI and IoT are going to change the nature of the
RGKCȩMLȩDPCCȩAMMJGLEȩKMBC
ȩUFGAFȩGQȩFGEFJWȩCȑAGCLRȩ internet as we know it. 2FCȧ CVNCARCBȧ EPMURFȧ GLȧ
For high-performance computing, the air cooling is
ȧ AMKNSRGLEȧ NPMACQQGLE
ȧ ?LBȧ CLCPEWȧ AML-
limited because of the poor conductivity of air and
the very large power densities and heat generated
by the processors. The use of water cooling can dra-
matically reduce the cooling energy consumption.
0UIFS
TPMVUJPOT: removing old servers and install-
B?R?
QSKNRGMLȧ?PCȧAF?LEGLEȧRFCȧNPCQCLRȧN?P?BGEK ‘do
everything in the cloud’ to a new one that can be
called ‘bring computing to where the data are’.
Edge and fog computing: many use the terms fog
computing and edge computing interchangeably,
GLEȩ LCUCPȩ ?LBȩ KMPCȩ CȑAGCLRȩ MLCQȩ CEȩ UGRFȩ #L- as both involve bringing intelligence and processing
CPEWȩ 1R?P
ȩ GLQR?JJGLEȩ KMPCȩ CȑAGCLRȩ QMDRU?PC
ȩ ?LBȩ AJMQCPȩRMȩUFCPCȩRFCȩB?R?ȩ?PCȩAPC?RCBȩ2FCȩICWȩBGȎCP
KMPCȩCȑAGCLRȩNMUCPȩBGQRPG@SRGML
ȩCEȩ@?RRCPGCQȩMLȩ ence between the two is where the intelligence and
QCPTCPQȩ#ȑAGCLAWȩF?Qȩ?JQMȩGKNPMTCBȩRF?LIQȩRMȩRFCȩ
#3ȩ !MBCȩ MDȩ !MLBSARȩ MLȩ #LCPEWȩ #ȑAGCLAWȩ GLȩ "?R?ȩ
Centres52, adopted voluntarily by more than 350
DCs in Europe.
ȧ LC?Pȧ RFCȧ QMSPACȧ MDȧ RFCȧ B?R?. For
computing power is placed.
•

&EHF
DPNQVUJOH: a methodology for optimis-
ing cloud computing systems by NCPDMPKGLEȧ
ȧ B?R?ȧ NPMACQQGLEȧ ?Rȧ RFCȧ CBECȧ MDȧ RFCȧ LCR
ȧ UMPI
 99 11Computer Processing and Energy Perspective

and acted upon. In other words, fog comput-

ing acts as midway in the processing chain
between the edge (e.g. sensors, mobile phones)
and the cloud.

Fog and edge computing architectures have been

introduced recently recognising the potentially
K?QQGTCȩGLAPC?QCȩGLȩB?R?ȩRP?ȑAȩRMȩ@CȩCVNCARCBȩUGRFȩ
the adoption of the IoT, where billions of devices
used on a daily basis will be connected to each-
other using new-generation wireless technology
?LBȩCK@CBBCBȩ'ȩ2FCȩ@CLCȏRQȩMDȩBCACLRP?JGQ?RGMLȩ
include: scalability (less data moved through the
internet), latency (AI closer to data sources), data
security (local scope), and cost. Fog and edge com-
NSRGLEȩ?PCȩBGȎCPCLRȩDPMKȩRP?BGRGML?JȩAJMSBȩAMKNSR-
GLEȩ?LB
ȩFCLAC
ȩRFCWȩUGJJȩPCOSGPCȩBGȎCPCLRȩQMJSRGMLQȩ
ȩ CV?KNJC
ȩNCPDMPKGLEȩKMPCȩAMKNSR?RGMLȩ?RȩRFC
level of the sensors capturing the data, or
mobile devices like mobile phones. In this
way, there is less need to transfer data to
centralised servers or clouds, and only the
result of the processing is then transferred,
2FCWȩ GLRPMBSACȩ ?ȩ BGȎCPCLRȩ AJMSBGȏA?RGMLȩ @SQGLCQQȩ
level dealing with more local and less powerful
clouds, which are closer to the smart edge devic-
CQȩ ?LBȩ RMȩ RFCȩ SQCPQȩ zȩ RFCWȩ ?PCȩ A?JJCBȩ CBECȩ AJMSBQȩ
or cloudlets. This type of infrastructure, which will
be crucial for the future development of IoT and AI
in Europe, is only emerging now. The approach and
frameworks commonly used for data processing and
AI in the cloud are not good for the cloudlets. We
have therefore a (narrow) window of opportunity to
make sure European industry takes a lead in devel-
oping this new infrastructure and collectively (Eu-
ropean Commission, Member States, and industry)
ȩ PCBSAGLEȩ RFCȩ B?R?ȩ RP?ȑAȩ AMLQGBCP?@JWȩ 2FGQ
approach can also leverage resources that
are not continuously connected to a network,
such as laptops, smart phones, tablets and
sensors.
• 'PH
DPNQVUJOH implements a decentralised
computing infrastructure in which data, com-
puting, storage and applications are distri-
ȩ @SRCBȩ GLȩ RFCȩ KMQRȩ JMEGA?J
ȩ CȑAGCLRȩ NJ?ACȩ • Respecting European values (transparency,
between the data source and the cloud. Fog
sets the rules of the game, including:
• A European open reference architecture for
fog and edge computing, complementing
the traditional cloud (DCs) approach.
• Reinforcing research on transportable DCs,
to support cloudlets and edge architectures.
privacy, and ethics) at the fog computing
ȩ AMKNSRGLEȩ CQQCLRG?JJWȩ CVRCLBQȩ AJMSBȩ AMK ȩ JCTCJȩzȩGCȩDMEȩLMBCQȩ?LBȩAJMSBJCRQ
puting and services to the edge of the
network, bringing the advantages and power
of the cloud closer to where data is created
 11Computer Processing and Energy Perspective 100

• As to cloudlets’ energy consumption, small

amounts of energy are consumed in a (b) Limit data movement reducing the energy
ȩ NJSP?JGRWȩMDȩJMA?RGMLQȩ-DRCLȩRFCȩMPE?LGQ?RGMLȩ consumed from machine-to-machine com-
paying for the energy may not be the same munications; this can also be achieved by
one running the infrastructure or the service. pushing data and computing towards the
ȩ 2FCPCDMPC
ȩ CLCPEWȩ CȑAGCLAWȩ NMJGAGCQȩ LCCB edge of the network. Important improve-
to consider new business models and pos- ments in this direction come from smart
sibly new regulations. sensors (e.g. embedded-AI systems) and
fog and edge clouds (cloudlets).
 1SKK?PW?LBAMLAJSQGML
(c) Reduce the embodied energy (i.e. energy
In this chapter, we have analysed one of the often- consumed in the production of goods) of
ignored aspects of AI, i.e. that RFCȧ MNNMPRSLGRGCQȧ digital technology by implementing the
APC?RCBȧ @Wȧ EPC?Rȧ B?R?ȧ TMJSKCQȧ ?LBȧ ?BT?LACBȧ Industry 4.0 Smart Factory principles and
?L?JWRGAQȧAMKCȧ?Rȧ?LȧGLAPC?QGLEȧCLTGPMLKCLR?J
ȧ in particular the smart energy pillar.
CAMLMKGA
ȧ ?LBȧ ECMNMJGRGA?Jȧ AMQRȧ CLCPEW. The in-
RPMBSARGMLȩMDȩRFCȩ'M2
ȩ?LBȩ%ȩUGJJȩGLAPC?QCȩRFCȩNPM@- (d) Reduce the obsolescence of digital tech-
JCKȩ QGELGȏA?LRJWȩ #LCPEWȩ GQȩ PCAMELGQCBȩ ?Qȩ ?ȩ K?GLȩ
ȩ DMPȩ CV?KNJC
ȩ RFCȩ QFMPRȩ JGDC
determinant of the long-term sustainability of AI.
We need therefore to take early action to address
this aspect of the development of AI and its relat-
ed technologies. While some actions can be taken
using today’s technology, more power-reduction
ȩ LMJMEGCQȩ QCC
span of mobile phones, replaced every one
to three years), which is perhaps the main
energy consumption voice for digital tech-
nology.
GLLMT?RGMLȩGQȩLCCBCBȩRMȩDSJȏJȩDSRSPCȩBCK?LBȩ'RȩGQȩ
necessary to introduce advanced technologies in
ȩQSN-
the hardware, software, and services areas, for
realising intensive-computing ecosystems that
are ‘green’.
ȩ QK?PRȩ QCLQMPQ
ȩ CȑAGCLRȩ
These ecosystem solutions must face four import-
We need to accelerate the speed of energy con-
QSKNRGMLȩPCBSARGMLȩRFPMSEFȩCLCPEWȩCȑAGCLAW
porting a holistic and multi-disciplinary approach
e.g. by introducing and using innovative semi-
AMLBSARMPȩ RCAFLMJMEGCQ
data transmission and DCs, and advanced comput-
ing architectures, to compensate for the speed of
ant energy consumption areas: B?R?ȩRP?ȑAȩGLAPC?QCȩ$MPȩRF?RȩUCȩLCCBȩRMȩA?PCDSJJWȩ
estimate the amount of energy required by AI (and
(a) 0CBSACȩRFCȩCLCPEWȩAMLQSKCBȩ@WȩRFCȩCVGQRGLE the IoT), taking into consideration all the required
ȩ @GEȩB?R?ȩACLRPCQ
ȩDMPȩCV?KNJC
ȩPCBSAGLEȩB?R?ȩ NF?QCQȩzȩCEȩB?R?ȩECLCP?RGML
ȩKMTGLE
ȩQRMPGLE
ȩNPM-
centralisation and pushing data and com- cessing, and insight generation and provision.
puting distribution towards the edge of the
cloud. Another promising area is the im-
ȩ NPMTCKCLRȩ MDȩ RFCȩ CVGQRGLEȩ AMMJGLEȩ RCAF
nologies.
101 11Computer Processing and Energy Perspective
 12 Data perspective 102

(PPE
RVBMJUZ
TIBSFE
EBUB
JT
FTTFOUJBM
UP
EFWFMPQ
TPDJBMMZ
SFTQPOTJWF
"*

SUMMARY

In this chapter, we review the key economic characteristics of digital data, including
CAMLMKGCQ
ȩ MDȩ QA?JC
ȩ MDȩ QAMNC
ȩ ?LBȩ LMLPGT?JPWȩ 2FCȩ ȏPQRȩ RUMȩ NMGLRȩ RMȩ ?Lȩ GLAPC?QCȩ
concentration of data, and hence information and power, in the hands of a few actors
in the internet economy, while non-rivalry creates potential tensions between opening
?AACQQȩRMȩRFCȩB?R?ȩQMȩRF?RȩQMAGCRWȩ@CLCȏRQȩRFCȩKMQR
ȩ?LBȩPCQRPGARGLEȩ?AACQQȩQMȩRF?Rȩ
RFCȩB?R?ȩFMJBCPȩ@CLCȏRQȩKMQRȩ2FCQCȩRCLQGMLQȩ?PCȩPCȐCARCBȩGLȩRFCȩJCE?JȩDP?KCUMPI
ȩ
as discussed earlier, and make the development of data strategies in an open and
globalised environment a particular challenge. We argue, however, that if we apply
the lessons of successful internet companies to the European public sector and de-
velop ecosystems based around public platforms, it is possible to create large pools
of shared data that are semantically well-structured and labelled, and can fuel new AI
?NNJGA?RGMLQȩGLȩBGȎCPCLRȩBMK?GLQȩ'LȩRFGQȩU?W
ȩUCȩA?LȩMNCLȩ?AACQQȩRMȩRFCȩB?R?
ȩBCTCJMNȩ
the market, serve the public, and enrich the data commons at the same time.
 103 12 Data perspective

DATA PERSPECTIVE
 2FCJ?U?LBCAMLMKGAQMD?AACQQ Data is the lifeline of
 RMB?R?

ML is a data-driven technology, as discussed in

AI. Europe is data rich
Chapter 2. It is necessary for us to make sense
of the enormous volumes and range of data be-
but the economics
ing produced daily, but at the same time ML needs
large volumes of data to learn. For this reason, as
and legal framework
we have seen in Chapter 4, all the countries devel-
oping AI strategies put access, and adding value to
create complex
data as key concerns.
@E>IIBKDBP-MBKFKD
Developing such strategies is not easy because
they need to balance the often-competing inter-
access to data and
ests between opening access to data to promote
innovation and transparency, and restricting it to
building interactions
NPMRCARȩ NPGT?AWȩ ?LBȩ AMKKCPAG?Jȩ AMLȏBCLRG?JGRWȩ
#SPMNCȩ GQȩ ?JQMȩ GLȩ ?ȩ N?PRGASJ?PJWȩ BGȑASJRȩ NMQGRGMLȩ
because the global data market is characterised
by a few dominant non-European players and plat-
forms that strengthen their power as they gather
more and more data from European customers and
governments.
There are many facets to this problem. Therefore,
in this chapter, we outline the key economic char-
acteristics of digital data, review the EU regulatory
QCRRGLEQȩ ?LBȩ RFCGPȩ GLȐSCLACȩ MLȩ RFCȩ RP?BCMȎQȩ @C-
ȩ?Lȩ?SRMK?RCBȩBPGTGLEȩ?JEMPGRFKȩK?WȩPC-
tween openness and restriction, and then put for-
ward the elements for a data-sharing strategy for
among participants
is key to succeeding.
taset increases and variation in the data decreases.
5GRFȩ?ȩFGEFȩT?PGCRWȩMDȩM@QCPT?RGMLQȩ?LBȩK?LWȩCV-
planatory variables in the dataset, achieving robust
predictions may require very large datasets. Typ-
ically, ML algorithms need much more data than
humans in order to train for comparable skills. For
CV?KNJC
quire millions of kilometres of driving data while a
human driver needs only a few thousand kilometres
the public sector. MDȩCVNCPGCLACȩRMȩ@CAMKCȩ?ȩNPMȏAGCLRȩBPGTCPȩ!MJJCAR-
GLEȩ ?LBȩ ?LLMR?RGLEȩ J?PECȩ B?R?QCRQȩ GQȩ BGȑASJR
ȩ RFCȩ
quality of raw data is often poor and needs clean-
2FCCAMLMKGAAF?P?ARCPGQRGAQ ing and structuring, and a lot of the data are often
 MDB?R? useless. Investing in high-quality datasets for train-
GLEȩ+*ȩ?JEMPGRFKQȩGQȩAMQRJWȩ?LBȩGKNJGCQȩFGEFȩȏVCBȩ
Digital data has three main economic characteristics:
ȩFMUCTCP
ȩRFCȩK?PEGL-
&DPOPNJFT
PG
TDBMF: ML shares many similarities
with statistical modelling. Statistical estimates be-
come more reliable as the size of the underlying da-
AMQRQȩ-LACȩRFCWȩ?PCȩRP?GLCB
al cost of additional use of the algorithm can be
very low. Common ML methods do not perform well
with small data sample sizes, especially when data
are noisy, i.e. contain many irrelevant variables.
 12 Data perspective 104

+*ȩKMBCJQȩRCLBȩRMȩ~MTCPȏRȩGLȩRFCQCȩA?QCQ
ȩPCBSAGLEȩ data by installing sensors and setting up servers to
RFCȩCȑAGCLAWȩMDȩ?JEMPGRFKQȩMLACȩRFCWȩ?PCȩ?NNJGCBȩ collect, store, transmit and process the data. With-
GLȩPC?JJGDCȩQCRRGLEQȩ-LCȩU?WȩRMȩ?BBPCQQȩRFGQȩGQȩRMȩ
try to understand the underlying mechanisms in
the data through causal inference. This potentially
increases the interpretability and fairness of algo-
rithms (Kilbertus et al., 2017; Athey, 2017).
&DPOPNJFT
PG
TDPQFȩRFCQCȩMAASPȩUFCLȩRFCȩ@CLCȏRQȩ
out the ability to monetise revenue from the data
there would be no incentive to collect them. This
NPM@JCKȩU?QȩȏPQRȩBGQAMTCPCBȩ@Wȩ?SRFMPQȩ?LBȩNS@-
lishers of books and by inventors of new technolo-
gies. All these innovative ideas and artworks could
easily be copied by others, leaving the original cre-
?RMPȩ UGRFMSRȩ ?LWȩ PCKSLCP?RGMLȩ DMPȩ FGQ
FCPȩ CȎMPRQȩ
of analysing a merged dataset are higher than the Intellectual property rights (IPR) were introduced to
?L?JWQGQȩMDȩC?AFȩB?R?QCRȩQCN?P?RCJWȩ2FCQCȩ@CLCȏRQȩ avoid these situations.
only occur when there is a relationship between the
two sets, i.e. when they are not completely separa- Should we apply the same economic principles
ble and data that pertain to one situation may also MDȩ'.0ȩRMȩB?R?ȩ?LBȩ?QQGELȩCVAJSQGTCȩPGEFRQȩRMȩB?R?ȩ
@CȩPCJCT?LRȩDMPȩ?LMRFCPȩQGRS?RGMLȩ$MPȩCV?KNJC
ȩUC@ȩ FMJBCPQ
ȩNMQQG@JWȩCTCLȩDSJJȩMULCPQFGNȩPGEFRQȩQȩGLȩ
QSPȏLEȩB?R?ȩK?WȩNPMBSACȩGLQGEFRQȩMLȩAMLQSKCPȩ@C-
haviour. Merging these with mobile phone data may
produce more insights when compared to studying
both datasets separately. Merging with pay data in
shops adds further insights, etc. Economies of scope
CVNJ?GLȩ UFWȩ B?R?BPGTCLȩ ȏPKQȩ ?PCȩ QMȩ B?R?FSLEPWȩ
and collect all the data they can get: the wider the
the case of IPR, these rights should keep a @?J?LACȧ
@CRUCCLȧEP?LRGLEȧCVAJSQGTCȧPGEFRQȧ?Qȧ?LȧGLACL-
RGTCȧDMPȧRFCȧNPMBSARGMLȧMDȧB?R?ȧ?LBȧCLQSPGLEȧ?A-
ACQQȧRMȧB?R?ȧ?LBȧ?JEMPGRFKQȧGLȧMPBCPȧRMȧNPMBSACȧ
UGBCPȧ@CLCȏRQȧDMPȧQMAGCRW. The applicability of IPR
to ML algorithms was discussed in Section 7.3. We
come back to legal protection of data rights be-
coverage of the available datasets in terms of varie- JMUȩ GLȩ ȩ $GPQR
ȩ UCȩ CVNJMPCȩ T?PGMSQȩ KC?LGLEQȩ
ty of situations and observations, the more accurate of ‘data access’.
and insightful the predictions that can be made with
these datasets.
NNJWGLECAMLMKGAPC?QMLGLE
/POSJWBMSZ: whilst goods and services are rival  RMB?R??AACQQ
products, i.e. they can only be used by one person
at a time, ?LȧGBC?
ȧ?ȧNGCACȧMDȧGLDMPK?RGMLȧMPȧ?Lȧ?J- The modalities and conditions for access to data
EMPGRFKȧ?PCȧLMLPGT?JȧRFCWȧA?Lȧ@CȧSQCBȧ@WȧK?LWȧ T?PWȩ AMLQGBCP?@JW
ȩ UGRFȩ NMRCLRG?JJWȩ TCPWȩ BGȎCPCLRȩ
NCMNJCȧ?RȧRFCȧQ?KCȧRGKCȧQȩ?ȩPCQSJR
ȩRFCȩ@CLCȏRQȩ economic outcomes for individuals and for society
derived from data are higher when used by a group as a whole (Bergmann and Morris, 2017). We should
of people compared to used by an individual. From therefore be careful and more precise in what we
a societal perspective, it may therefore be better mean when we use the word ‘access’.
to share data and algorithms as widely as possible
P?RFCPȩRF?LȩICCNGLEȩRFCKȩNPGT?RCȩ&MUCTCP
ȩRFCȩȐGNȩ ȩB?R?ȩFMJBCPȩK?WȩLMRȩNCPACGTCȩ?LWȩ@CLCȏRQȩGLȩBGP
side of wide sharing and easy access is the lack ectly sharing or selling data to others and decide to
of incentives to invest in production of data and keep them private, protected by legal and/or tech-
algorithms. Data may be a by-product of ongoing nical protection measures. Rather than selling data
activities and require no additional incentives. For directly, they can be used indirectly as input in the
GLQR?LAC
ȩCAMKKCPACȩȏPKQȩQRMPCȩAMLQSKCPȩ@CF?T- NPMBSARGMLȩ MDȩ ?ȩ AMKKCPAG?Jȩ QCPTGACȩ $MPȩ CV?KNJC
ȩ
iour data on their websites as part of their online %MMEJCȩ BMCQȩ LMRȩ BGPCARJWȩ QCJJȩ MPȩ QF?PCȩ AMLQSKCPȩ
transactions. Public administrations collect loads of data but uses them for selling advertising slots.
data while carrying out their duties. In other cases !MLQSKCPQȩ @CLCȏRȩ DPMKȩ RFCȩ GLDMPK?RGMLȩ QCPTGACQȩ
FMUCTCP
ȩȏPKQȩK?WȩBCJG@CP?RCJWȩGLTCQRȩGLȩAMJJCARGLEȩ while advertisers pay for the advertising service.
 105 12 Data perspective

In this way, even non-accessible data can generate LMRFCPȩ QACL?PGMȩ MAASPQȩ UFCLȩ B?R?ȩ ?PCȩ CVRP?ARCBȩ
@CLCȏRQȩDMPȩGLBGTGBS?JQȩ?LBȩDMPȩQMAGCRWȩ from data holders against their will, possibly with-
out their knowledge and at a cost to them, but may
Data holders can also make the data directly avail- @CȩMDȩ@CLCȏRȩRMȩQMAGCRWȩ#V?KNJCQȩGLAJSBCȩK?LB?-
?@JCȩ'DȩRFCȩB?R?ȩ?PCȩP?RFCPȩCVAJSQGTCȩ?LBȩT?JS?@JC
ȩ tory open data access for environmental reporting
they may fetch a price in the data market. Since and access to private data to detect or prevent
data are non-rival they can be sold many times. crime. Individual data holders may incur costs but
+MLMNMJGQRGAȩQCJJCPQȩMDȩCVAJSQGTCȩB?R?ȩA?LȩCGRFCPȩQCJJȩ
?Rȩ?ȩȏVCBȩNPGACȩMPȩ?BHSQRȩNPGACQȩ?AAMPBGLEȩRMȩC?AFȩ
buyer’s willingness to pay. In that case, everyone
has access and there is no welfare loss to society.
2FCȩB?R?ȩKMLMNMJGQRȩK?VGKGQCQȩKMLCR?PWȩ@CLCȏRQȩ
from the data. A welfare distribution issue may oc-
ȩUGRFMSRȩILMUGLEȩGR
ȩ@SRȩRFCȩȏPKȩK?Wȩ@CLCȏRȩ
RFCȩ LCRȩ @CLCȏRȩ RMȩ QMAGCRWȩ GQȩ NMQGRGTCȩ LMRFCPȩ CV-
?KNJCȩGQȩMLJGLCȩȏPKQȩRF?RȩRCQRȩT?PGMSQȩAMKKCPAG?Jȩ
strategies on a small sample of consumers before
deploying them on a larger scale. The consumers in
RFCȩRCQRȩQ?KNJCȩK?WȩE?GLȩMPȩJMQCȩDPMKȩRFCȩCVNCPG-
KCLR
ASPȩUFCLȩKMQRȩMDȩRFCȩ@CLCȏRQȩ?AAPSCȩGLȩKMLCRGQCBȩ DPMKȩCVRP?ARGLEȩ@CF?TGMSP?JȩGLDMPK?RGMLȩDPMKȩSQCPQȩ
form to the data holder and less so to data users or
the person whose behaviour generated the data. If 2PSQRȧGQȧ?LȧGKNMPR?LRȧD?ARMPȧGLȧB?R?ȧQF?PGLEȧ?LBȧ
RFCȩB?R?ȩ?PCȩJCQQȩCVAJSQGTCȩ?LBȩAMKNCRGLEȩQMSPACQȩ ?AACQQȧA variety of initiatives have been deployed
for (partial) substitute data are available, the mo- RMȩ QRPCLERFCLȩ RPSQRȩ UFGJCȩ QF?PGLEȩ B?R?ȩ #VNCPG-
LMNMJWȩA?Lȩ@CȩUC?ICLCBȩ?LBȩKMPCȩ@CLCȏRQȩUGJJȩ?A- mental personal information management spaces
crue to data users. The choice between direct and (PIMS) give individuals more control over the use
indirect data sales is often not straightforward for of their personal data. Similarly, ‘data trusts’ seek
B?R?ȩFMJBCPQȩQMKCȩK?WȩBMȩ@MRFȩGLȩN?P?JJCJȩ$MPȩCV- RMȩNPMTGBCȩRPSQRCBȩCVAF?LECȩKCAF?LGQKQȩTG?ȩJCE?Jȩ
?KNJC
ȩK?LWȩB?R?ȩȏPKQȩQCJJȩB?R?ȩBGPCARJWȩ?LBȩNPM- conditions, storage formats and mutual oversight
vide data analytics services. organisations.

ȩBGȎCPCLRȩQGRS?RGMLȩMAASPQȩUGRFȩMNCLȩMPȩDPCCȩ?AACQQȩ
that makes data available to all potential users at .MJGAWGLRCPTCLRGMLGLRFCB?R?
XCPMȩAMQRȩ CLCȏRQȩ?PCȩBGQRPG@SRCBȩ?APMQQȩRFCȩUGBCQRȩ  K?PICR
possible group of users, similar to the perfect price
BGQAPGKGL?RGMLȩA?QCȩ?@MTC
ȩCVACNRȩRF?Rȩ?JJȩ@CLCȏRQȩ #VRP?ARGLEȩGLQGEFRQȩDPMKȩB?R?QCRQȩ@WȩKC?LQȩMDȩ+*ȩ
accrue to data users, not the data holder. If data MPȩMRFCPȩRCAFLGOSCQȩA?Lȩ@CȩTCPWȩ@CLCȏAG?JȩDMPȩQMKCȩ
are a by-product of the data holder’s ordinary ac- @SRȩRFCȩGLQGEFRQȩA?Lȩ?JQMȩ@CȩCVNJMGRCBȩRMȩPCBGQRPG@SRC
tivities and can be collected at low or zero cost,
DPCCȩ?AACQQȩUGJJȩLMRȩ?ȎCARȩRFCȩUCJD?PCȩMDȩRFCȩB?R?ȩ
holder. Even when publishing the data carries a cost
to the data holder, open access may prove bene-
ȏAG?JȩRMȩRFCȩFMJBCPȩ@CA?SQCȩEPC?RCPȩSQCȩMDȩRFCȩB?R?ȩ
can act as an advertisement, and improve data
ȩ RFGQȩ P?GQCQȩ RFCȩ OSCQRGMLȩ
quality through validation and error detection at
limited cost for the data holder. However, if data
collection is costly and their use is very valuable,
the holder may not have an incentive to grant free
and open access. Losing revenue may also reduce
incentives for data collectors to invest in innovative
services that could be produced with the data.
@CLCȏRQȩ @CRUCCLȩ N?PRGCQȩ 1F?PGLEȩ ?AACQQȩ RMȩ B?R?ȩ
A?LȩNMRCLRG?JJWȩGLAPC?QCȩRFCȩ@CLCȏRQȩ@SRȩK?Wȩ?JQMȩ
result in more redistribution. Should policymakers
intervene in data ownership and access markets,
MPȩ QFMSJBȩ RFCWȩ @Cȩ JCDRȩ RMȩ MNCP?RCȩ DPCCJWȩ 2P?LQJ?R-
CBȩ GLRMȩ CAMLMKGAȩ H?PEML
of potential data market failures that reduce social
welfare and require policy intervention. Such fail-
ures may not only occur in primary data markets
but also in downstream markets for goods and ser-
vices where transactions depend on access to data.
2FCȩCVRCLRȩMDȩ?AACQQȩ?LBȩAMKNCRGRGMLȩGLȩSNQRPC?Kȩ
data markets may have implications for competition
in downstream goods and services markets.
 12 Data perspective 106

"?R?ȩ K?Wȩ F?TCȩ TCPWȩ @CLCȏAG?Jȩ SQCQȩ DMPȩ QMAGCRWȩ GLȩ 1MKCȩNPMTGQGMLQȩGLȩRFCȩ%".0ȩPCBSACȩ?AACQQȩRMȩNCP-
terms of innovation and welfare. However, data can QML?Jȩ B?R?ȩ $MPȩ CV?KNJC
ȩ RFCWȩ NPMFG@GRȩ RFCȩ SQCȩ MDȩ
also be used to reduce competition in markets and personal data for purposes other than the original-
nudge consumers into decisions that are not neces- ly intended purpose, unless they are anonymised.
Q?PGJWȩRMȩRFCGPȩ?BT?LR?ECȩ2FCPCȩGQȩMDRCLȩ?ȩRP?BCMȎȩ 2FGQȩK?ICQȩGRȩBGȑASJRȩDMPȩFMJBCPQȩMDȩNCPQML?JȩB?R?ȩ
@CRUCCLȩQFMPRRCPKȩNPGACȩCȑAGCLAWȩ?LBȩJMLECPRCPKȩ to aggregate them into larger databases or share
innovation incentives. them with other types of data services for the pur-
pose of aggregation and analysis by means of ML
0CESJ?RMPQȧA?LȧAPC?RCȧCVAJSQGTCȧAMLRPMJȧPGEFRQȧDMPȧ ?JEMPGRFKQȩ-RFCPȩNPMTGQGMLQȩGLȩRFCȩ%".0ȩD?AGJGR?RCȩ
B?R?ȧFMJBCPQȧMPȧRFCWȧA?LȧQMDRCLȧRFCQCȧCVAJSQGTCȧ ?AACQQȩRMȩB?R?ȩ$MPȩCV?KNJC
ȩPRGAJCȩȩK?LB?RCQȩ
PGEFRQȧ?LBȧNSLAFȧQMKCȧFMJCQȧGLȧRFCKȧRMȧD?AGJGR?RCȧ portability of personal data. At the request of the
?AACQQȧRMȧRFCȧB?R? In reality, regulators often do B?R?ȩ QS@HCAR
ȩ RFCȩ B?R?ȩ FMJBCPȩ F?Qȩ RMȩ RP?LQDCPȩ RFCȩ
@MRFȩ?RȩRFCȩQ?KCȩRGKCȩzȩ?QȩUCȩUGJJȩLMUȩGJJSQRP?RCȩ personal data to a destination of choice of the data
for the case of the EU. QS@HCARȩ2FCȩB?R?ȩQS@HCARȩA?LȩRFSQȩAFMMQCȩRMȩEGTCȩ
wider access to his/her data. Anonymous data can
In the EU, there are two general horizontal legal be transferred or used for other purposes. A good
instruments that have something to say on data CV?KNJCȩGQȩL?TGE?RGMLȩ?NNQȩLȩGKNMPR?LRȩQMSPACȩMDȩ
PGEFRQȩ RFCȩ %CLCP?Jȩ "?R?ȩ .PMRCARGMLȩ 0CESJ?RGMLȩ PM?BȩRP?ȑAȩB?R?ȩGQȩRFCȩKMTCKCLRȩMDȩKM@GJCȩNFMLCQȩ
%".0ȩ#!ȩ?ȩDMPȩNCPQML?JȩB?R?ȩMDȩL?RSP?JȩNCP- in cars. Telecom service operators track the pos-
sons and the Data Base Directive (DBD) (EC, 1996) itioning of mobile phones as a technical measure
DMPȩ?LWȩB?R?ȩAMJJCARCBȩ@WȩȏPKQȩQCCȩ1CARGMLȩ to keep the phones connected to the antennas of
RFCȩ KM@GJCȩ NFMLCȩ LCRUMPIȩ 2FCȩ %".0ȩ ?JJMUQȩ RFCȩ
2FCȩ #3ȩ %".0ȩ QCRQȩ QMKCȩ PSJCQȩ PCE?PBGLEȩ RFCȩ SQCȩ transmission of anonymous data about movement
of personal data. Any data that can be linked to a MDȩKM@GJCȩNFMLCQȩ2FCȩ%".0ȩNSRQȩLMȩPCQRPGARGMLQȩMLȩ
natural person is considered to be personal data, cross-border transfers of data within the EU. How-
irrespective of the source, storage or transmission ever, RP?LQDCPQȧMSRQGBCȧRFCȧ#3ȧA?LȧMLJWȧ@CȧK?BCȧ
KCAF?LGQKȩ 2FCȩ %".0ȩ BMCQȩ LMRȩ EP?LRȩ MULCPQFGNȩ
PGEFRQȩ RMȩ NCPQML?Jȩ B?R?ȩ @SRȩ EGTCQȩ B?R?ȩ QS@HCARQȩ ?ȩ
number of control rights over their personal data,
including the right to obtain consent from the data
QS@HCARȩ@CDMPCȩ?AACQQGLEȩRFCȩB?R?
ȩRMȩRFCȩPGEFRȩDMPȩ right. Copyright protects the structure of data-
B?R?ȩ QS@HCARQȩ RMȩ ?AACQQȩ RFCGPȩ NCPQML?Jȩ B?R?ȩ ?LBȩ
delete them, to port their data to other uses, etc.
Privacy Personal data protection is considered to be
a basic and inalienable human right that cannot be
traded away. For this reason, personal data are as-
QSKCBȩLMRȩRMȩ@CȩRFCȩQS@HCARȩMDȩ?LȩCAMLMKGAȩRP?LQ-
action. However, a frequently used online business
KMBCJȩ GLTMJTCQȩ RFCȩ CVAF?LECȩ MDȩ NCPQML?Jȩ B?R?ȩ GLȩ
RMȧMNCP?RMPQȧRF?RȧPCQNCARȧRFCȧ%".0
The EU Database Directive (DBD) (EC, 1996) con-
tains two provisions: copyright and a TVJ
HFOFSJT
bases which, if original, constitutes the author’s
own intellectual creation. By contrast, the more
controversial TVJ
HFOFSJT right
protects databases
regardless of their originality, as long as there has
been ‘substantial investment in obtaining, verifying
or presenting the contents’. Both the copyright and
the TVJ
HFOFSJT
right
in the DBD put the database
MULCPȩGLȩ?ȩNMQGRGMLȩMDȩCVAJSQGTCȩ?AACQQȩ?LBȩRFCPC-
PCRSPLȩ DMPȩ ?ȩ B?R?@?QCBȩ QCPTGACȩ 2FCȩ B?R?ȩ QS@HCARȩ fore reduce access to the data. A recent evaluation
A?Lȩ ?AACNRȩ MPȩ PCDSQCȩ RFCȩ CVAF?LEC
ȩ BCNCLBGLEȩ MLȩ of the DBD (EC, 2018g) concluded that it has been
whether s/he perceives it as a fair and credible deal CȎCARGTCȩGLȩF?PKMLGQGLEȩRFCȩ#3ȩJCEGQJ?RGMLȩMLȩB?R?-
RF?RȩUGJJȩ@PGLEȩKMPCȩ@CLCȏRQȩRF?LȩNMRCLRG?JȩF?PKQȩ bases and avoiding fragmentation, and that it still
from reduced privacy. provides an appropriate balance between protect-
GMLȩ MDȩ GLTCQRKCLRȩ ?LBȩ GLRCPCQRQȩ MDȩ SQCPQȩ -Lȩ RFCȩ
other hand, it also notes that as a result of a 2004
 107 12 Data perspective

decision of the Court of Justice53 clarifying that the RMȩCVRP?ARȩ?BBGRGML?JȩGLDMPK?RGMLȩ?LB MPȩQCPTCȩRFCȩ

TVJ
HFOFSJT right does not apply to databases that community better. A 2015 study funded by the EC54
?PCȩRFCȩ@WNPMBSARQȩMDȩRFCȩK?GLȩ?ARGTGRWȩMDȩ?ȩȏPK
ȩ supports the view that public administrations will
it must be assumed that these rights do not apply @Cȩ@WȩD?PȩRFCȩJ?PECQRȩ@CLCȏAG?PGCQȩMDȩMNCLGLEȩSNȩRFCȩ
broadly to the data economy (machine-generated data they hold.
data, IoT devices, big data, AI, etc.). The evaluation
concluded that at this stage reforming the TVJ
HFO 1GLACȧB?R?ȧ?PCȧLMLPGT?Jȧ?LBȧA?Lȧ@CȧRP?LQDCPPCBȧ
eris rights would be largely disproportionate. The ?RȧTGPRS?JJWȧXCPMȧAMQRȧ?JJȧMTCPȧRFCȧUMPJB
ȧRFCȧOSCQ-
BC@?RCȩ?KMLEȩQAFMJ?PQȩMLȩRFCȩ@CLCȏRQȩMDȩPCAMELGQ- RGMLȧMDȧ~B?R?ȧQMTCPCGELRWȧF?Qȧ@CAMKCȧ?LȧGQQSCȧ
ing such rights particularly for machine-generated 2FGQȩPCTMJTCQȩ?PMSLBȩRFCȩGBC?ȩRF?RȩB?R?ȩ?PCȩQS@HCARȩ
data is still very open (see Section 7.2.). to the laws and governance structures in the na-
tion where it is collected55ȩ5GRFGLȩRFCȩ#3
ȩRFCȩ%".0ȩ
2FCȩRCVRȩ?LBȩB?R?ȩKGLGLEȩ2"+ȩCVACNRGMLȩRMȩAMNW allows transfer of personal data between Member
right protection could also be considered as a States. The recent EU Free Flow of Data policy in-
measure that opens access to data. In the EU, the GRG?RGTCȩ?GKQȩRMȩCVRCLBȩRFGQȩRMȩ?JJȩRWNCQȩMDȩB?R?ȩGLȩ
2"+ȩCVACNRGMLȩGQȩJGKGRCBȩRMȩLMLAMKKCPAG?JȩSQCȩMDȩ order to avoid national fragmentation of data mar-
the data by public sector research institutes and kets. Personal data transfers outside the EU re-
organisations. That is an opening combined with a OSGPCȩRF?RȩRFCȩPCAGNGCLRȩAMKNJGCQȩUGRFȩ%".0ȩPSJCQ
restriction. In the USA, the ‘fair use’ provisions al-
low for wider use of data obtained through TDM,
provided they do not compete with the services of-
fered by the data holder (Section 7.2).

2FCȧ#3ȧ.S@JGAȧ1CARMPȧ'LDMPK?RGMLȧ.1'ȧ"GPCARGTCȧ
#!
ȧ ȧ QCCIQȧ RMȧ K?ICȧ B?R?ȧ FCJBȧ @Wȧ NS@JGAȧ Since data are
QCARMPȧ MPE?LGQ?RGMLQȧ KMPCȧ ?AACQQG@JCȧ RMȧ AGRG-
XCLQȧ?LBȧȏPKQȩ2MȩRFCȩCVRCLRȩRF?RȩQMKCȩMDȩRFCQCȩ
organisations may be involved in the production of
non-rival and
AMKKCPAG?Jȩ QCPTGACQȩ zȩ DMPȩ GLQR?LACȩ NS@JGAȩ SRGJGRGCQȩ
GLȩRP?LQNMPRȩ?LBȩCLCPEWȩzȩGRȩK?Wȩ?JQMȩAMTCPȩQMKCȩ
can be transferred
commercial operations. Providing public access to
these commercially sensitive data may entail op-
at virtually zero
portunity costs for the data holders. The Directive
suggests that pricing of data access should incur
cost all over
marginal costs. Digital datasets are often charac-
RCPGQCBȩ@WȩRFCȩFGEFȩȏVCBȩAMQRQȩMDȩAPC?RGLEȩRFCȩB?R?-
the world
set and virtually zero marginal costs of replicating
and transmitting it. As a result, the marginal cost
the question of
rule may not cover the actual cost of producing
public-sector data but, as argued earlier, much of
‘Data Sovereignty’
RFGQȩB?R?ȩGQȩECLCP?RCBȩRMȩDSJȏJȩ?ȩJCE?JȩK?LB?RCȩGLȩ
RFCȩȏPQRȩNJ?AC
ȩQMȩGRQȩPCSQCȩ@CAMKCQȩ?ȩ@WNPMBSAR
ȩ
has become
?LBȩ A?Lȩ @CLCȏRȩ RFCȩ NS@JGAȩ ?BKGLGQRP?RGMLȩ KSAFȩ
more than keeping the data closed. This is the case, an issue.
DMPȩ CV?KNJC
ȩ GDȩ RFCȩ B?R?ȩ ECLCP?RCQȩ LCUȩ NPMBSARQȩ
and services that the public administration can use
 12 Data perspective 108

 1SKK?PW 2MU?PBQ?B?R?QRP?RCEWDMPNS@JGA

 ?BKGLGQRP?RGMLQ
(Big) datasets are a necessary input into ML algo-
rithms. Access to data is important for the develop-
ment of AI. Access and wide data sharing can occur
in markets for direct and indirect data sales, across
a wide spectrum of modalities and economic con-
BGRGMLQ
ȩCTCLȩUFCLȩB?R?ȩFMJBCPQȩF?TCȩCVAJSQGTCȩAML- eral and ML in particular. Robust algorithms need
trol over their data.
#3ȧPCESJ?RMPWȧGLRCPTCLRGMLQȧGLȧB?R?ȧK?PICRQȧ?PCȧ
A?SEFRȧ@CRUCCLȧRUMȧNMJCQȧ?ȧMȎCPGLEȧKMPCȧCV-
AJSQGTCȧ PGEFRQ as an incentive for data producers
and holders to invest more in data collection &
analysis, and @ȧK?IGLEȧB?R?ȧKMPCȧUGBCJWȧ?T?GJ-
ȩDMPȩCV?KNJCȩ
?@JCȩ ?LBȩ ?AACQQG@JCȩ RMȩ D?AGJGR?RCȩ RFCȩ CVRP?ARGMLȩ
of new insights from data, including with AI/ML.
There is a wide-open unregulated space between
NCPQML?Jȩ B?R?ȩ PGEFRQȩ SLBCPȩ RFCȩ %".0ȩ ?LBȩ ECLCP?Jȩ
ȩ?LBȩBCTCJMNGLEȩNMJGAGCQȩ
database ownership rights under the DBD where
K?PICR@?QCBȩ B?R?ȩ CVAF?LECQ
ȩ @GJ?RCP?Jȩ AMLRP?ARQȩ mentation in analogy to the way ML adapts based
and technical protection measures rule. There is an
ongoing debate in the EU as to whether this open
space contains some market failures that need to
1CRRGLERFCQACLC
At several points in this report we have highlighted
that data is crucial to the development of AI in gen-
large volumes of good-quality data for training,
where ‘good’ means data that is well organised, se-
mantically labelled, free of bias and open, as dis-
cussed in Chapters 2 and 6.
.S@JGAȩ ?BKGLGQRP?RGMLQȩ AMSJBȩ @CLCȏRȩ QGELGȏA?LRJWȩ
DPMKȩRFCȩSQCȩMDȩB?R?ȩQAGCLACȩ?LBȩ'
to design policy that is targeted to groups and geo-
graphical areas that are in greatest need of inter-
vention, modelling the potential cumulative impact
MDȩNMJGAWȩGLȩBGȎCPCLRȩ?PC?Q
that ‘learn’ from feedback loops from policy imple-
on data feeds (Craglia, Hradec and Troussard,
2019). For this reason, most national AI strategies
summarised in Chapter 4 consider AI as a great op-
@Cȩ ȏJJCBȩ SNȩ UGRFȩ DSPRFCPȩ PCESJ?RGMLȩ 2F?Rȩ BC@?RCȩ
P?LECQȩ DPMKȩ CVN?LBGLEȩ CVAJSQGTCȩ B?R?ȩ MULCPQFGNȩ
rights (European Commission, January 2017) to
facilitating access to data (European Commission,
April 2018). The rapid rise of AI/ML as a very prom-
GQGLEȩ B?R?NPMACQQGLEȩ RCAFLMJMEWȩ RMȩ CVRP?ARȩ KMPCȩ
insights and value from data is fuelling this debate.
When data and data analytics technologies be-
portunity to modernise public administration, public
services and research centres like the Alan Turing
Institute in the UK, which has developed an active
public policy programme.
In the previous section, we have highlighted the
tensions between opening up data and restricting
access and use, from both an economic and legal
come cheaper and more widely available, the need
for protection as an incentive to investors may di-
minish in favour of more access to stimulate innov-
ation. Finding a new balance between these two op-
ȩ K?XML
ȩ $?AC@MMI
ȩ NNJCȩ
posite poles in the debate requires more societal
and policy debate.56
'LȩRFCȩLCVRȩQCARGMLȩUCȩAMLQGBCPȩQMKCȩNMQQG@JCȩQRP?R-
CEGCQȩRMȩCVRP?ARȩT?JSCȩDPMKȩRFCȩB?R?ȩUFGJCȩQF?PGLEȩ
ȩQR?WQȩGLȩRFCGPȩPCQNCARGTCȩAMSLRPGCQȩ
them.
NCPQNCARGTCȩ2FCQCȩRCLQGMLQȩ?PCȩCV?ACP@?RCBȩ@WȩRFCȩ
international dimension: in the global competition
between the USA and China, and among the key in-
BSQRPG?Jȩ NJ?WCPQȩ %MMEJC
and Microsoft on the one hand, and Baidu, Alibaba,
Tencent and Xiaomi on the other), Europe is notably
absent. As noted by the Mission Villani (2018), whilst
most data captured in the USA and China, and to
?LȩCVRCLRȩ0SQQG?
because of market dominance, policy, or both, in Eur-
MNCȩRFCȩT?QRȩK?HMPGRWȩMDȩB?R?ȩNPMTGBCBȩ@Wȩ#SPMNC?Lȩ
citizens, companies and governments ends up out-
side Europe. Potentially, the more the public and re-
search sectors pursue an open data policy, the more
 109 12 Data perspective

RFCWȩ DCCBȩ RFGQȩ ȐMUȩ MDȩ B?R?ȩ MSRQGBCȩ #SPMNCȩ 'Lȩ RFGQȩ 3)ȩ #VNJMPGLEȩLCUȩDMPKQȩMDȩN?PRLCPQFGNȩUGRFȩRFC
AMLRCVR
ȩ UCȩ D?ACȩ ?ȩ PC?Jȩ AF?JJCLEC
ȩ GDȩ U?LRȩ RMȩ QR?Wȩ public to share personal data, leveraging
open, in developing a data strategy for Europe that their increased control on access to such
increases the pool of European data, adds value to ȩ B?R?ȩNPMTGBCBȩ@WȩRFCȩ%".0
it by structuring and semantically annotating it, and
RP?GLQȩ'ȩ?NNJGA?RGMLQȩRF?Rȩ@CLCȏRȩ#SPMNC?LȩQMAGCRWȩ -NCLGLEȩ?LBȩQF?PGLEȩB?R?ȩ?APMQQȩ?JJȩICWȩQR?ICFMJBCPQȩ
Many countries are grappling with this challenge, as zȩNS@JGAȩQCARMP
ȩAMKKCPAG?JȩQCARMPȩ?LBȩRFCȩNS@JGAȩzȩGQȩ
GLBGA?RCBȩGLȩ!F?NRCPȩȩ-LCȩAMSLRPWȩRF?RȩF?Qȩ@CRRCPȩ very important and already a challenge. Adding value
articulated a possible response is France where the to the data while sharing it is an additional crucial
report by the Mission Villani (2018) put forward a dimension to support the development of AI bene-
RFPCCNPMLECBȩB?R?ȩQRP?RCEWȩQCCȩ MVȩ ȏAG?JȩRMȩ#SPMNCȩ'LȩRFGQȩPCQNCAR
ȩUCȩA?LȩJC?PLȩDPMKȩRFCȩ
NPGT?RCȩQCARMPȩ?QȩMSRJGLCBȩGLȩRFCȩLCVRȩQCARGML

ɎɛɤDZ Summary data strategy from the Mission Villani Report (2018)

#LAMSP?ECAMKN?LGCQRMNMMJ?LBQF?PCRFCGPB?R?
The government must encourage the creation of data commons
and support an alternative data production and governance
model based on reciprocity, cooperation and sharing. The goal
is to boost data sharing between actors in the same sector.
?V@>PB>KAPB@QLOPMB@Fȭ@?>PFPCLOMR?IF@FKQBOBPQOB>PLKP1EFP
P
could be in one of two ways: by making the data accessible onlyy
to the government, or by making the data more widely available, e,
for example, to other economic actors.
The government must also encourage data sharing between 1SNNMPRRFCPGEFRRMB?R?NMPR?@GJGRW
private actors, and assist businesses in this respect. It must The right to data portability is one of the most importantt
organise for certain data held by private entities to be released FKKLS>QFLKPFKOB@BKQ$OBK@E>KA#ROLMB>KQBUQP'QTFIIDFSB>KVV
LK>@>PB?V@>PB?>PFP
>KAPRMMLOQA>Q>>KAQBUQJFKFKD individual the ability to migrate from one service ecosystem m
practices without delay. to another without losing their data history. This right could bee
BUQBKABAQL>II@FQFWBK@BKQOBA>OQFȭ@F>IFKQBIIFDBK@B>MMIF@>QFLKP
P
!PC?RCB?R?RF?RGQGLRFCNS@JGAGLRCPCQR In this case, it would involve making personal data available too
Most of the actors heard by the mission were in favour of government authorities or researchers.
MOLDOBPPFSBIVLMBKFKDRM>@@BPPQLPLJBA>Q>PBQPLK>@>PB

This proposed strategy is interesting because it
builds on three key dimensions:
1) Using the EU legal framework, including the
ȩ %".0
ȩ ?Qȩ ?Lȩ MNNMPRSLGRWȩ LMRȩ HSQRȩ RMȩ Q?DC suppliers, advertisers, software developers, etc.)
guard European interests but also to nurture
ȩQCPTGACQȩ?LBȩGL-
European companies.
2) Framing the opening of public-sector data in
a two-way relationship of reciprocity with
the commercial sector to gain increased
access to some commercial data of public
interest.
@MSRNJ?RDMPKQ?LB.'Q
From an economic perspective, a (digital) platform
is where two or more types of users (consumers,
AMKCȩRMECRFCPȩRMȩCVAF?LECȩEMMBQ
formation. Successful platforms make use of net-
UMPIȩCȎCARQȩRMȩ?RRP?ARȩKMPCȩSQCPQȩMLȩQCTCP?JȩQGBCQȩ
of the market: more consumers attract more sup-
pliers and vice versa. They also leverage the data
that they collect on user behaviour on the plat-
form to reinforce their own position. Economies of
scope in data collection and analysis give them an
 12 Data perspective 110

informational advantage over individual suppliers

and consumers (see Martens, 2016).
ɎɛɤǩǨ#V?KNJCQMD.'QGLJMA?J
government
From a computation perspective, a (web) platform
is an evolution of the web from a medium to pub-
"CLK?PIQ BBPCQQCQ 5C@ .' "5 DAWA is a service ce
lish information to an infrastructure to build ap-
allowing public authorities, businesses and citizens to use usse
NJGA?RGMLQȩ -LCȩ MDȩ RFCȩ K?HMPȩ AMKNMLCLRQȩ MDȩ UC@ȩ
>RQELOFQ>QFSB >AAOBPPBP FK QEBFO '1 PVPQBJP 'K 
 FQ OB@BFSBA
B@BFSBBA
digital platforms is the use of web application pro-
?FIIFLKOBNRBPQP>KA>MMOLUFJ>QBIVRKFNRBRPBOPMBO PBOP MBO
gramming interfaces (APIs). APIs allow third-party
TBBH1EBKRJ?BOLCÄ'OBNRBPQPFPIFJFQBAQLOBNRBPQPMBO BPQPMBO
developers to access selected data and functionali-
PB@LKA1EBOB>OB>MMOLUFJ>QBIV'1PVPQBJPAO>TFKDA>Q> KDA>Q>
ties available on the platform, and build new appli-
concerning Danish addresses through DAWA.
cations. APIs let the data provider keep better con-
trol of how the data is used, and let the third-party
+?BPGB+M@GJGRW*?@Qthis is an open and interoperable platform
latform
developer build applications without the need to
from the city council allowing third parties to develop applications
cations
download the entire set of data.
FKQEBȭBIALCRO?>KJL?FIFQV'QOB@BFSBPJFIIFLKOBNRBPQPMBO
BPQPMBO
VB>OABSBILMBOP>OBOBDFPQBOBAFKQEBPVPQBJ
>KA>OLRKA
>OLRKA
ȩEMMBȩCV?KNJCȩGQȩ%MMEJC+?NQ
ȩUFGAFȩGQȩ?AACQQG@JCȩ
>MMPE>SB>IOB>AV?BBKABSBILMBA
to developers via an API. In this way, the developer,
Source: ELISE project, Joint Research Centre.
DMPȩCV?KNJC
ȩMDȩ?ȩFMRCJQȩUC@QGRCȩBMCQȩLMRȩLCCBȩRMȩ
download the map data showing where the hotel is
JMA?RCBȩMPȩFMUȩRMȩECRȩRFCPCȩ'RȩCK@CBQȩRFCȩ%MME-
leMaps API in the application to call its functional- $PMKMNCLGLEB?R?
GRGCQ
ȩ?LBȩQFMUQȩRFCȩJMA?RGMLQȩ?RȩBGȎCPCLRȩQA?JCQȩMPȩ  RMQK?PRQF?PGLE
directions to get there from anywhere in the world.
This provides a clear advantage to the developer, Using APIs to make the data available to third par-
@SRȩ?JQMȩRMȩRFCȩ%MMEJCȩNJ?RDMPKȩ@CA?SQCȩGRȩA?LȩPC?Bȩ ties is an essential step towards developing the
every call on the application and see how popular data economy, but is not enough to underpin an AI
the hotel is, from where it gets its customers, how it strategy. 5F?RȧGQȧLCACQQ?PWȧGQȧDMPȧRFCȧ#SPMNC?Lȧ
is associated with other facilities, and so on. In oth- NS@JGAȧQCARMP
ȧGLȧN?PRGASJ?P
ȧRMȧKMTCȧDPMKȧNSQFGLEȧ
er words, the platform gains intelligence by moni- RFCȧB?R?ȧMSRȧzȧ?ȧAQN@CB@RSHMFȧKMBCJȧzȧRMȧBP?UGLEȧ
toring the use of the API. This links to the economic RFCȧSQCPQȧGL
ȧ@MRFȧRFGPBN?PRWȧBCTCJMNCPQȧ?LBȧRFCȧ
perspective above: the more users of the APIs, the NS@JGA
ȧ?LBȧRMȧBCTCJMNȧGLRCP?ARGML
ȧRFCȧHMSDQ@BSHUD«
more intelligence the platform gathers, the more it
reinforces its position, and the more it attracts new
developers and customers.
Public administrations across Europe are starting
to use APIs to provide access to their data in ma-
chine-readable form. When they do, they often see
?ȩQGELGȏA?LRȩGLAPC?QCȩGLȩRFCȩSQCȩMDȩRFCGPȩB?R?ȩAMK-
pared to publishing it as downloadable datasets
KMBCJ
5CȩQCCȩDPMKȩRFCȩCV?KNJCȩMDȩAMKKCPAG?JȩNJ?RDMPKQȩ
RF?RȩSQCPQȩGLRCP?ARȩGLȩBGȎCPCLRȩU?WQ
• Providing new content.
• Annotating/tagging/classifying the content
already available on the platform.
QCCȩ MVȩ
• 3QGLEȩRFCȩQCPTGACQȩMȎCPCB

• Building/using new apps and functionalities

based on the platform APIs.
 111 12 Data perspective

Users’ behaviour (interaction) and content generat- contract among all the key stakeholders so that
CBȩMPȩCVAF?LECBȩAMLRPG@SRCȩRMȩDMPKȩRFCȩCLF?LACBȩ each agree on the sharing and use of their data for
data that is so important for AI. Raw data as such is AJC?PȩNSPNMQCQȩ?LBȩAJC?Pȩ@CLCȏRQȩ
frequently of little value, but this increases when it
GQȩAMLRCVRS?JGQCBȩ?LBȩAPMQQPCDCPCLACBȩ$?AC@MMIȩGQȩ
the biggest content manager in the world, but does
not produce any content. Its users do it.
ȩ NPGT?RCȩ QCARMPȩ ?LBȩ RFCȩ NS@JGA
ȩ DMPȩ CV?KNJCȩ
If we want to take advantage of the enormous di-
versity and richness of European data, we should
JC?PLȩ DPMKȩ RFCȩ CV?KNJCQȩ MDȩ QSAACQQDSJȩ GLRCPLCRȩ
companies and develop European data ecosystems
bringing together the public sector, the commercial
sector, academia and the third sector and the gen-
eral public. Developing such ecosystems requires a
LCUȩN?PRLCPQFGN
ȩDMPȩCV?KNJC
ȩGLȩRFCȩDMPKȩMDȩ?ȩQMAG?Jȩ cations must be based on, and reinforce the Euro-
ȩ
We should learn
We have seen in the review of national AI strate-
gies in Chapter 4 that several countries are already
PCȐCARGLEȩMLȩFMUȩRMȩQF?PCȩB?R?ȩ@CRUCCLȩRFCȩNS@JGAȩ
QCARMP
through data trusts (UK), social contracts (France)
or MyData (Finland). Each country may have its
own model but we need to share the best practices
and synergies and agree a framework to then en-
rich this data for the public good.
There is general agreement that European AI appli-
pean values enshrined in the Treaties. These values
and the European legal framework must be seen as
a source of strength and global competitive advan-
tage, not a limitation for AI. They must permeate
?JJȩRFCȩD?ACRQȩMDȩ?ȩ#SPMNC?Lȩ~B?R?ȏA?RGMLȩNPMACQQ
i.e. the process of transforming the information we
generate into new forms of value or insights. From
from the examples RFCȩCVGQRGLEȩQAGCLRGȏAȩJGRCP?RSPCȩCEȩ2?LȩCRȩ?J
ȩȩ
0MWȩCRȩ?J
ȩȩ-NFCPȩCRȩ?J
ȩȩUCȩQCCȩRF?RȩRFGQȩ
of successful process can be organised into three main phases,
(see Fig. 14): data collection and aggregation, data
internet companies transformation/enhancement, and insight genera-
tion and provision.

and develop European #@S@«BNKKDBSHNM«@MC«@FFQDF@SHNM«OG@RD

data ecosystems 2FCȩȏPQRȩB?R?ȏA?RGMLȩNF?QCȩGLTMJTCQȩQCRRGLEȩSNȩCD-

bringing together ȏAGCLRȩ ?LBȩ JMLERCPKȩ B?R?ȩ AMJJCARGMLȩ ?LBȩ QRMP?ECȩ
solutions. This is essential to provide a stable foun-

the public sector, the dation for any analytical process. Infrastructure is
necessary to facilitate and support the formation
MDȩ AMJJ?@MP?RGTCȩ CAMQWQRCKQ
ȩ ?Rȩ BGȎCPCLRȩ JCTCJQȩ
commercial sector, (national, European, and community or sectoral
MLCQ
ȩRMȩAMLQRGRSRCȩ?ȩPCJG?@JC
ȩȐCVG@JC
ȩ?LBȩQA?J?@JCȩ
academia and the third supply-chain system for data, and processing. Valu-
?@JCȩCV?KNJCQȩAMSJBȩGLAJSBCȩNS@JGAȩ?BKGLGQRP?RGMLȩ
sector, and the general data ecosystems, industrial data ecosystems, for
CV?KNJC
ȩGLȩRFCȩQN?ACȩQCARMP
ȩ'LBSQRPWȩ
ȩFC?JRF
ȩ
public. etc. In the latter, data collected and provided by the
public would also be crucial.
 12 Data perspective 112

Figure 14.'DWDÀFDWLRQSURFHVVHQDEOLQJWKH$,UHYROXWLRQ
 113 12 Data perspective

As discussed in Chapter 13, if we want to use the  1SKK?PW?LBAMLAJSQGMLQ

MNNMPRSLGRGCQȩ?ȎMPBCBȩ@Wȩ'ȩRMȩ@SGJBȩAMKNCRCLACQ
ȩ
skills, and resilience at the local level, particularly in In this chapter, we review the key economic char-
peripheral regions, we also need to consider local- acteristics of digital data, including economies of
level ecosystems bringing together public adminis- QA?JC
ȩMDȩQAMNC
ȩ?LBȩLMLPGT?JPWȩ2FCȩȏPQRȩRUMȩNMGLRȩ
RP?RGMLQ
ȩ JMA?Jȩ ȏPKQ
ȩ ?A?BCKG?
ȩ ?LBȩ JMA?Jȩ AMKKS to an increase in the concentration of data, and
nities to share and enrich local data and build appli- hence information and power, in the hands of a few
cations addressing local needs. actors in the internet economy, while non-rivalry
creates potential tensions between opening access
#@S@«SQ@MRENQL@SHNMDMG@MBDLDMS«OG@RD RMȩRFCȩB?R?ȩQMȩRF?RȩQMAGCRWȩ@CLCȏRQȩRFCȩKMQR
ȩ?LBȩ
restricting access so that the data holder bene-
'LȩRFGQȩNF?QC
ȩBGȎCPCLRȩB?R?ȩQRPC?KQȩ?LBȩRWNCQȩLCCBȩ ȏRQȩKMQRȩ2FCQCȩRCLQGMLQȩ?PCȩPCȐCARCBȩGLȩRFCȩJCE?Jȩ
to be transformed, harmonised and aggregated to framework, as discussed in Chapter 7, and make
increase the value of the data collected and make it the development of data strategies in an open and
useful for AI application. This transformation phase globalised environment a particular challenge.
PCOSGPCQȩBMK?GLȩCVNCPRGQC
ȩUFGAFȩK?WȩLMRȩ@Cȩ?T?GJ-
able everywhere in the same way. Therefore, it is To address this challenge, we have put forward
important to develop the transformation algorithms some key elements of a possible QRP?RCEWȧRMȧK?ICȧ
from the outset so that they are transparent and RFCȧ @CQRȧ SQCȧ MDȧ RFCȧ T?QRȧ PGAFLCQQȧ MDȧ #SPMNC?Lȧ
reusable. This is the key level where we need to ad- QMAGCRWȧ ?LBȧ B?R? We suggest that the tensions
BPCQQȩ RFCȩ AMLACPLQȩ MDȩ RP?LQN?PCLAW
ȩ CVNJ?GL?@GJGRW
ȩ and dilemmas between openness and restrictions
?AAMSLR?@GJGRWȩ?LBȩD?GPLCQQȩCVNPCQQCBȩGLȩ!F?NRCPȩȩ
and also also discussed in Chapters 6. To do so, re-
quires the adoption at the European level of proto-
cols to document, share and reuse algorithms, and
agreed evaluation frameworks for algorithms.
(MRHFGS«FDMDQ@SHNM«@MC«OQNUHRHNM«SGQNTFG«
SGD«OK@SENQL
This last development phase is essential to match
demand and supply and create real economic val-
ȧ RP?LQDMPK?RGMLȧ ?LBȧ CLF?LACKCLR
ȧ GLQGEFRȧ
outlined in Section 12.1 can be navigated by learn-
ing from the most successful companies on the in-
ternet. We need to add value to the data we have,
particularly in the public sector, by moving from the
broadcasting model to the interactive one in which
users contribute to enrich the data whilst using it,
and providers gain intelligence from the feedback
they receive. This means BCTCJMNGLEȧ#SPMNC?LȧB?R?
NJ?RDMPKQȧUGRFȧ?ȧAJC?PȧB?R?ȏA?RGMLȧQRP?RCEWȧ?P-
RGASJ?RCBȧ ?PMSLBȧ B?R?ȧ AMJJCARGMLȧ ?LBȧ ?EEPCE?-
RGML
SCȩ'LȩRFGQȩAMLRCVR
ȩ?ȩNJ?RDMPKȩGQȩRFCȩNJ?ACȩRF?RȩCL- ECLCP?RGMLȧ ?LBȧ NPMTGQGML Through these phases
ables interactions between customers and service we can also address the concerns over transpar-
providers, where the latter can be the public admin- CLAW
ȩ ?AAMSLR?@GJGRW
ȩ ?LBȩ D?GPLCQQȩ CVNPCQQCBȩ GLȩ
istration or third-party developers. In this phase, !F?NRCPQȩ ȩ ?LBȩ ȩ ?LBȩ ȏLBȩ ?ȩ #SPMNC?Lȩ U?Wȩ RMȩ 'ȩ
ML algorithms are developed and trained with development that is consistent with our values.
the pool of data available to respond to user de-
mand or local needs, tested, and deployed to pro-
vide a service. Important aspects to achieve a Euro
pean-level strategy for data services and boost the
European AI economy would include developing a
framework for sharing and reusing AI/ML training
B?R?QCRQȩ?LBȩ?JEMPGRFKQ
ȩBCȏLGLEȩKGLGKSKȩOS?JGRWȩ
of service, and ensuring platform interoperability
and portability of data.
 13 Societal resilience perspective 114

Ȅ*O
XJOUFS
TOPX
GBMMT
PO
UIF
CBNCPP
USFFT
5IFZ
CFOE
CVU
OFWFS
CSFBL
BOE
DPNF
TQSJOH
UIF
USFFT

TUBSU
UP
HSPX
TUSBJHIU
BHBJO

"EESFTT
CZ
/PSJP
.JUTVZB
1BSMJBNFOUBSZ
4FOJPS
7JDF.JOJTUFS
.JOJTUSZ
PG
'PSFJHO
"ҧBJST
+BQBO
.$.
$IBJS

0&$%
4PVUIFBTU
"TJB
3FHJPOBM


'PSVN

.BSDI
#BMJ
*OEPOFTJB

IUUQTXXXNPGBHPKQNPGBKҨMFTQEG

SUMMARY

In this chapter, we introduce the concept of resilience which is useful to frame a

#SPMNC?Lȩ?NNPM?AFȩRMȩ'ȩ?BBPCQQGLEȩBGȎCPCLRȩNF?QCQȩNPCTCLRGML
ȩ?LRGAGN?RGML
ȩNPCN?-
P?RGML
ȩ@SRȩ?JQMȩ?B?NR?RGMLȩ?LBȩRP?LQDMPK?RGMLȩRMȩ@MSLACȩ@?AIȩDPMKȩRFCȩCȎCARQȩMDȩ?ȩ
shock. It is important to both measure and monitor resilience, and we introduce some
CV?KNJCQȩMDȩFMUȩRFGQȩAMSJBȩ@CȩBMLCȩCTCLȩRFMSEFȩRFCPCȩ?PCȩ?ȩLSK@CPȩMDȩKCRFMBMJMEG-
cal and data-related limitations that need to be addressed.

As we show in Chapter 9, we do not know what impacts AI will have on work and
society, but we do know we need to increase our resilience to AI, particularly in regions
of Europe already more vulnerable. We put forward some elements of a possible
approach centred on the network of Digital Innovation Hubs. There are already several
FSLBPCBȩFS@Qȩ?RȩRFCȩNPCQCLRȩRGKCȩUFGAFȩ?PCȩCVN?LBGLEȩP?NGBJWȩUGRFȩ?ȩR?PECRȩMDȩMLCȩ
for each region in Europe, many of which are focusing on AI. Their purpose is to faci-
JGR?RCȩ?AACQQȩRMȩRCAFLMJMEWȩ?LBȩILMUFMUȩRMȩNS@JGAȩ?BKGLGQRP?RGMLQȩ?LBȩȏPKQ
ȩ1+#Qȩ
in particular, in combination with the planned AI-on-demand-platform. We suggest
putting these hubs at the centre of local ecosystems which could create local pools of
shared data so that the AI skills developed/upgraded locally could be put to good use
in developing AI algorithms and solutions trained on local data to address local needs.
 115 13 Societal resilience perspective

SOCIETAL RESILIENCE
PERSPECTIVE
2FCLCCBDMPPCQGJGCLAC Harnessing the
As we discussed in the previous chapters of this re- richness and diversity
NMPR
ȩ 'ȩ MȎCPQȩ K?LWȩ NMRCLRG?Jȩ MNNMPRSLGRGCQȩ RMȩ GK-
prove our lives but also raises a number of challeng- of European regions
es to society. We are preparing ethical guidelines for
the development of AI systems that are account- and data is crucial
able, responsible and transparent but we must also
recognise that future change may be unpredictable to build resilience
and sudden. Therefore, we should not only try and
shape the development of technology so that it bene- and develop AI
ȏRQȩ MSPȩ QMAGCRW
ȩ @SRȩ ?JQMȩ NPCN?PCȧ MSPȧ GLQRGRSRGMLQ
ȧ
NMJGAGCQ
ȧ NCMNJCȧ ?LBȧ QMAGCRWȧ ?Rȧ J?PECȧ RMȧ @CAMKCȧ applications
KMPCȧȐCVG@JC
ȧ?B?NR?@JCȧ?LBȧPC?BWȧRMȧRP?LQDMPK.
Basically, we need to become more resilient. responsive to
Interest in resilience has been rising rapidly dur- European
ing the last 20 years as a response to increasing
challenges like climate change, globalisation, im- and local needs.
migration, and digital transformation that test our
coping capacities as individuals, regions, countries This framework may prove very useful in addressing
and institutions. the multiple challenges and opportunities of AI. For
CV?KNJC
ȩ K?LWȩ UMPICPQȩ UGJJȩ LCCBȩ @MRFȩ NPMRCARGMLȩ
Resilience is a key concept in the policy discourse of (absorptive capacity), and support for their potential
the EU57, and the JRC has developed a framework retraining (adaptive capacity). This also highlights
(Manca et al., 2017) for societal resilience that in- the centrality of individual resilience capacities.
cludes not only prevention, anticipation and prepa- .PCN?P?RGMLȩ GQȩ ?JQMȩ ?ȩ K?HMPȩ CJCKCLR
ȩ GLAJSBGLEȩ RFCȩ
ration but also adaptation and transformation, i.e. right policy frameworks (as advocated in the con-
the need to reform and potentially bounce back af- RCVRȩMDȩEJM@?JGQ?RGMLȩ@Wȩ ?JBUGLȩ?LBȩ4GFPG?J?
ȩ
ȩ
ter a shock. According to this framework, a system MPȩ%GJCQ
ȩȩDMPȩ?LRGRPSQRȩNMJGAGCQȩ?LBȩRFCȩLCACQ-
(or society) is resilient if it can face shocks and per- sary public debates about the ways forward as a so-
sistent structural changes without losing its ability AGCRWȩ?QȩP?GQCBȩ@W
ȩ?KMLEȩMRFCPQ
ȩ%SGJJĸLȩ?LBȩ0CBBW
ȩ
to deliver societal well-being in a sustainable way 2018; University of Melbourne, 2018).
(i.e. deliver societal well-being without compromis-
ing that of future generations). 5Cȩ?JQMȩLCCBȩRMȩAMLQGBCPȩRF?RȩRFCȩCȎCARQȩMDȩ'ȩK?Wȩ
RP?LQACLBȩECMEP?NFGCQȩ?LBȩQCARMPQȩ'RȩGQȩCVNCARCBȩ
to give a new boost to globalisation (Baldwin and
 13 Societal resilience perspective 116

Vihriala, 2017), but could also lead to a repatri- BGPCARJWȩ-LCȩQSAFȩCV?KNJCȩGQȩ(0!Qȩ',$-0+ȩ'LBCVȩ

?RGMLȩ MDȩ HM@Qȩ NPCTGMSQJWȩ BCJMA?JGQCBȩ GLȩ AFC?NCPȩ DMPȩ0GQIȩ+?L?ECKCLRȩ"Cȩ%PMCTCȩCRȩ?J
ȩ
AMSLRPGCQȩ'Lȩ?LWȩA?QC
ȩHM@ȩJMQQCQȩ?LBȩPC?JJMA?RGMLQȩ
UGJJȩPCQF?NCȩAMKKSLGRGCQȩJCV?LBCP
ȩȩQȩ?Lȩ 'Lȩ RFCȩ A?QCȩ MDȩ '
ȩ CVNCPR@?QCBȩ GLBGA?RMPQȩ ?PCȩ ?J-
CV?KNJC
ȩQCJDBPGTGLEȩA?PQȩUGJJȩLMRȩHSQRȩCL?@JCȩBPGT- PC?BWȩ @CGLEȩ BCTCJMNCBȩ -LCȩ CV?KNJCȩ GQȩ .?Sȩ CRȩ ?Jȩ
ing without a steering wheel, but will change a ma- (2017) who looked at a set of Asian countries, and
HMPȩCAMQWQRCKȩ@SGJRȩMLȩA?PȩMULCPQFGNȩ?LBȩSQCȩ$GLL
ȩ prepared two composite indicators, for AI prepared-
2018). Inequality may have an even stronger global ness and resilience. Preparedness is ‘the ability of
component (Lee, 2017). This calls for a IPMJTUJD
BOE
companies and talent to capitalise on opportunities
DPNQMFY
BQQSPBDI
UIF
TZTUFN
WJFX
. @PMSEFRȩ?@MSRȩ@Wȩ'
ȩUFGJCȩPCQGJGCLACȩzȩGLȩRFCGPȩ?N-
NPM?AFȩ zȩ GQȩ RFCȩ ~?@GJGRWȩ RMȩ ?B?NRȩ RMȩ ?LBȩ UGRFQR?LBȩ
Moreover, the development and use of AI is set to structural changes brought about by AI’.
rise and continue for a long time. We need to look
into the future, ?QQCQQȧGRQȧCȎCARQȧMLȧNCMNJC
ȧAMK- The main ingredients for preparedness are start-
KSLGRGCQ
ȧAMSLRPGCQȧ?LBȧFSK?LGRWȧQMAGCR?JȧUCJJ up activity, AI funding (venture capital), students
@CGLE
ȧ?LBȧUCȧLCCBȧRMȧGLȐSCLACȧGRQȧDSRSPCȧBCTCJ- enrolled in science, technology, engineering and
MNKCLRȧRFCȧBWL?KGAȧPC?ARGML
wisely. As argued K?RFCK?RGAQȩ 12#+ȩ QS@HCARQȩ ?Rȩ RMNP?LICBȩ SLG-
by Melvin Kranzberg, ‘technology is neither good nor versities, and AI publication volume. Resilience is
bad; nor is it neutral’, so it is crucial to actively steer assessed using relevant government policies and
the development of AI to ‘improve the quality of our RFCȩHM@ȩQRPSARSPCȩ'LȩMSPȩRCPKGLMJMEW
ȩNPCN?PCBLCQQȩ
NPMBSARQȩ?LBȩMSPȩQR?LB?PBȩMDȩJGTGLEȩ%MMJQ@CC
ȩ can be viewed as a part of resilience, as it signals
the ability to respond well to AI penetration. Below
Finally, we need not only to become more resili- UCȩNPCQCLRȩRFPCCȩGJJSQRP?RGTCȩCV?KNJCQȩMDȩQGKGJ?PȩGL-
ent in the face of AI but also to use AI to become dicators for the EU.
KMPCȩ PCQGJGCLR
ȩ DMPȩ CV?KNJCȩ @Wȩ CVRCLBGLEȩ RFCȩ SQCȩ
of AI-enabled predictive analytics in public policy, As an attempt to assess the penetration of, and the
infrastructure maintenance, disaster risk manage- dynamic response to AI in Europe, Figure 15 shows
KCLR
ȩ?EPGASJRSPC
ȩAPGKCȩNPCTCLRGMLȩMPȩRP?ȑAȩAMLRPMJȩ the shares of manufacturing versus healthcare in
(Prakash, 2017). AI can also provide improved tech- employment, comparing the USA and the EU (un-
nologies to foster citizen engagement in building up JGICȩRFCȩQG?LȩCV?KNJC
ȩUCȩJMMIȩ?RȩRFCȩQRPSARSPCȩ@Wȩ
a resilient EU society. QCARMPQȩ ?LBȩ LMRȩ @Wȩ HM@ȩ RWNCQȩ ȩ K?HMPȩ BCAJGLCȩ GLȩ
manufacturing and a rise in healthcare can be seen
+C?QSPGLEPCQGJGCLACRM' in both entities. The EU still has a larger share of
manufacturing and a smaller share of healthcare.
It is important to both measure and monitor resili- This indicates the possibility of further declines in
CLACȩ JCQQGȩ CRȩ ?Jȩ ȩ NPMTGBCȩ MLCȩ MDȩ RFCȩ ȏPQRȩ manufacturing, but also in the absorption capacity
CV?KNJCQȩMDȩ?ȩDSJJȩKC?QSPCKCLRȩQRP?RCEWȩ?JGELCBȩ of healthcare. The degree of relative shift was larger
with the JRC conceptual framework focusing on the in the USA (6.8 percentage points, as opposed to
ȩ ȏL?LAG?Jȩ ?LBȩ CAMLMKGAȩ APGQCQȩ 2FCȩ RUM 
ȩUFGAFȩK?WȩCVNJ?GLȩRFCȩFGEFCPȩBCEPCCȩMDȩNCQ-
step approach adopted by Alessi and colleagues
creates resilience indicators based on the proper-
ties of the system’s components, and then looks
DMPȩ QR?RGQRGA?JJWȩ QGELGȏA?LRȩ ?LBȩ PM@SQRȩ NPCBGARMPQȩ
(resilience characteristics)58 of resilient behaviour.
The more widespread approach is to identify, typ-
GA?JJWȩ@WȩCVNCPRȩ?QQCQQKCLRQ
ȩQSAFȩAF?P?ARCPGQRGAQȩ
simism (and also the change in inequality) in the
USA. Although the underlying mechanisms require
further analysis, the decline in manufacturing is
informative about the penetration of automation,
while the increase in healthcare and its further ab-
sorption capacity can be linked to resilience.
 117 13 Societal resilience perspective

Figure 15. Employment in manufacturing and healthcare: EU and USA

Source: Eurostat for the EU, OECD for the USA

Figure 16 shows a second measurement attempt, for possible future action. In Figure 18, we bring
this time looking mostly at preparedness at the together both Digital Innovation Hubs and current
Member State level. It combines various aspects of ?A?BCKGAȩQSNNJWȩMDȩQRSBWȩRMNGAQȩMLȩ'
ȩ?QȩOS?JGȏCBȩ
'MPGCLRCBȩEMTCPLKCLRȩCȎMPRQȩ@?QCBȩMLȩRFCȩȩ GLȩ!F?NRCPȩȩ'LȩRFGQȩU?W
ȩUCȩA?LȩQR?PRȩPCȐCARGLEȩMLȩ
CBGRGMLȩMDȩRFCȩ-NCLȩ"?R?ȩ ?PMKCRCPȩ?LBȩB?R?ȩMLȩ potential synergies, gaps and priorities to develop
RFCȩ QF?PCȩ MDȩ PCJCT?LRȩ ȏCJBQȩ GLȩ FGEFCPȩ CBSA?RGML59 such local ecosystems.
2FCȩȏESPCȩQFMUQȩRF?RȩDMPȩ?@MSRȩF?JDȩMDȩRFCȩAMSL-
tries all variables point to the same direction60, so
EMTCPLKCLRȩCȎMPRQȩ?LBȩCBSA?RGMLȩN?RRCPLQȩNMPRP?Wȩ
a coherent picture of preparedness.

Finally, Figure 17 looks at preparedness at the level of

EU regions, showing the share of people employed
in science and technology61, as a percent of the eco-
nomically active population. Some regions in the
Southern and Eastern part of the EU look particu-
larly unprepared. 2FGQȧA?JJQȧDMPȧSNQIGJJGLE
ȧ@SRȧ?JQMȧ
?ARGMLȧRMȧPCBSACȧRFCȧ@P?GLȧBP?GLȧDPMKȧRFCȧ1MSRFȧ
?LBȧ #?QRȧ RMȧ RFCȧ ,MPRF. The network of Digital In-
LMT?RGMLȩ &S@Q
ȩ UFGAFȩ GQȩ ASPPCLRJWȩ CVN?LBGLEȩ DPMKȩ
DCUȩFSLBPCBQȩRMȩNMRCLRG?JJWȩMTCPȩȩ
ȩAMSJBȩMȎCPȩ
an opportunity to bring together the training supply
from academia and technical centres, with poten-
tial employers in the private and public sector, and
local banks of shared data. In this way, JMA?Jȧ'ȧQIGJJQȧ
?LBȧ?NNJGA?RGMLQȧAMSJBȧ@CȧBCTCJMNCBȧRMECRFCPȧRMȧ
?BBPCQQȧJMA?JȧLCCBQ
ȧ?LBȧQRPCLERFCLȧPCQGJGCLACȧ?Rȧ
RFCȧQ?KCȧRGKC. It is worth noting that at the present
time, some of the countries with the low science and
technology employment regions, like Bulgaria and Figure 16. AI preparedness in EU Member States
Romania, still lack such hubs, suggesting directions Source: Source Open Data Barometer and Eurostat.
 13 Societal resilience perspective 118

Figure 17. Human resources in science and technology and Digital Innovation Hubs
Source: Eurostat regional yearbook 2018 and JRC.

While we are making some progress in understand- operation by a robot, while a somewhat bigger share
ing how to measure resilience to AI, there are some (47 % in 2012 and 2014, but declining to 35 % in
BGȑASJRGCQȩGLȩ?NNJWGLEȩRMȩ'ȩRFCȩA?SQ?J
ȩRUMQRCNȩ?N- 2017) would be comfortable with robots at their
proach that Alessi et al. (2018) used to measure the UMPINJ?ACȩ -Lȩ ?ȩ KMPCȩ NMQGRGTCȩ LMRCȩ ?LBȩ JMMIGLEȩ ?Rȩ
PCQNMLQCȩRMȩRFCȩȏL?LAG?JȩAPGQGQȩ$MPȩRFCȩȏPQRȩQRCNȩMDȩ the broader public perception of digital transform-
the approach, which involves assessing the strength ations, the 2017 Eurobarometer survey (SP460), in-
of the phenomenon, we lack a long-term time series dicates that three quarters of the respondents think
MDȩ 'ȩ ?BMNRGMLȩ GLȩ BGȎCPCLRȩ AMSLRPGCQȩ 2FCȧ 'ȧ 'LBCVȧ the digital transformation has a positive impact on
?LBȧGRQȧNJ?LLCBȧ#3ȧCVRCLQGMLȧ@WȧRFCȧ(0!ȧK?WȧMȎCPȧ the economy. Two thirds (67 %) think that the im-
?ȧQR?PRGLEȧNMGLRȧRMȧ?BBPCQQȧRFGQȧJGKGR?RGML. Anoth- pact is also positive on their quality of life, while
er possibility is to use corporate- and citizen-level
QSPTCWQȩ +MPGI?U?ȩ ȩ NPCQCLRQȩ QSAFȩ CV?KNJCQȩ
for Japan, while various special issues of the Euro-
barometer survey (SP382 in 2012, SP427 in 2014
?LBȩ 1.ȩ GLȩ 
ȩ DMPȩ CV?KNJCȩ ?PCȩ GLBGA?RGTCȩ MDȩ
ȩϤȩ?JQMȩȏLBȩ?ȩNMQGRGTCȩGKN?ARȩMLȩQMAGCRWȩ2?ICLȩ
together, these results make it possible to start per-
DMPKGLEȩQMKCȩȏPQRȩPCQGJGCLACȩ?L?JWQGQȩUFGJCȩ@SGJBGLEȩ
up the time series on AI adoption.
NS@JGAȩ MNGLGMLȩ GLȩ BGȎCPCLRȩ #SPMNC?Lȩ AMSLRPGCQȩ $MPȩ Another set of challenges relates to the second step
CV?KNJC
ȩRFCȩRFPCCȩQSPTCWQȩ?@MTCȩQFMUȩRF?RȩMLJWȩ?ȩ of the Alessi et al. model, which involved robust stat-
small share of EU citizens (around 22 %, quite sta- istical predictors of resilient behaviour. Here the
ble in time) would be comfortable having a medical main problem is that AI is already pervasive across
 119 13 Societal resilience perspective

Figure 18. Human resources in science and technology, Digital Innovation Hubs, and academic supply in study topics related to AI
Source: Eurostat regional yearbook 2018 and JRC.

many sectors of the economy and society and there- *CQQMLQDPMKRFC?L?JWQGQ
DMPCȩ GRȩ GQȩ TCPWȩ BGȑASJRȩ RMȩ BGQCLR?LEJCȩ NPMACQQȩ DPMKȩ  MD'DMPSLBCPQR?LBGLEPCQGJGCLAC
response, and attribute causality. The way we re-
QNMLBȩRMȩRFCȩCȎCARQȩMDȩ'ȩK?WȩGLȐSCLACȩ?LBȩQRCCPȩ Looking at AI from a resilience perspective does not
RFCȩ CVRCLRȩ ?LBȩ QCTCPGRWȩ MDȩ RFCȩ CȎCARQȩ RFCKQCJTCQȩ
This steering process is a key aspect of resilience.
-LCȩ NMQQG@GJGRWȩ DMPȩ KC?QSPGLEȩ RFCȩ CȎCARGTCLCQQȩ MDȩ
steering is to keep track of predictions and then
compare them to realised outcomes some time af-
only enrich our understanding of, and preparedness
for AI, but also advances the analysis of resilience.
2FCȩA?QCȩMDȩ'ȩGQȩMLCȩMDȩRFCȩȏPQRȩCV?KNJCQȩ?RȩRFCȩ(0!ȩ
of considering a persistent structural change, which
A?JJQȩDMPȩ?Lȩ?BHSQRCBȩAMLACNRS?Jȩ?LBȩKC?QSPCKCLRȩ
RCPU?PBQȩ $MPȩ CV?KNJC
ȩ RFCȩ NPCBGARGMLQȩ MDȩ HM@ȩ BGQ- ?NNPM?AFȩ'RȩNPMTGBCQȩ?LȩGLRCPCQRGLEȩCV?KNJCȩMDȩRFCȩ
placements and replacements, which are based on LCCBȩ RMȩ RFGLIȩ ?@MSRȩ BGȎCPCLRȩ A?N?AGRGCQȩ ?LBȩ GLRCP-
current forecasts of technology developments, could TCLRGMLQȩ ?Rȩ BGȎCPCLRȩ RGKCȩ FMPGXMLQȩ 'Rȩ CKNF?QGQCQȩ
@CȩPCAMPBCBȩ1WQRCK?RGAȩBGȎCPCLACQȩ@CRUCCLȩRFCȩPC- the role of power relationships, and the issue of po-
alised and forecast numbers may be indicative of tential updates of the ‘social contract’. Finally, it calls
AF?LECQȩGLȩRFCȩ?SRMK?RGMLȩNPMACQQȩ-RFCPȩNMQQG@GJG- DMPȩQRCCPGLEȩRFCȩDSRSPCȩN?RFȩMDȩ'ȩRMȩRFCȩ@CLCȏRȩMDȩ
ties may emerge as we apply the resilience frame- humankind, and thereby coming out stronger from
work to AI and progress in practice. the adoption of AI.
 14 Summary and conclusions 120

SUMMARY
AND CONCLUSIONS
5Cȧ?PCȧMLJWȧ?RȧRFCȧ@CEGLLGLEȧMDȧ?ȧP?NGBȧNCPGMBȧMDȧ
RP?LQDMPK?RGMLȧMDȧMSPȧCAMLMKWȧ?LBȧQMAGCRW due
to the convergence of many digital technologies.
PRGȏAG?Jȩ 'LRCJJGECLACȩ 'ȩ GQȩ ACLRP?Jȩ RMȩ RFGQȩ AF?LECȩ
The EU Member States and the European Com-
mission are developing coordinated national and
European strategies, recognising that we can only
succeed together. We can build on our areas of
?LBȩMȎCPQȩK?HMPȩMNNMPRSLGRGCQȩRMȩGKNPMTCȩMSPȩJGTCQȩ
ȩ GLAJSBGLEȩ CVACJJCLRȩ PCQC?PAF
ȩ JC?BCPQFGNȩ
The recent developments in AI are the result of in-
creased processing power, improvements in algo-
PGRFKQȩ ?LBȩ RFCȩ CVNMLCLRG?Jȩ EPMURFȩ GLȩ RFCȩ TMJSKCȩ
and variety of digital data. Many applications of AI
have started to appear in our everyday lives, from
machine translations, to image recognition, and
music generation, and are increasingly deployed in
industry, government and commerce. Connected
and autonomous vehicles, and AI-supported med-
ical diagnostics are areas of application that will
soon be commonplace.
There is strong global competition on AI among the
ȧin addition to centralised facilities. This
USA, China and Europe. The USA leads for now but
QRPCLERF
in some industrial sectors like automotive and
robotics, a solid legal and regulatory framework,
and very rich cultural diversity also at the regional
and sub-regional levels.
'RȩGQȩECLCP?JJWȩPCAMELGQCBȩRF?Rȩ'ȩA?LȩȐMSPGQFȩMLJWȩGDȩ
supported by a robust computing infrastructure and
good quality data:
• With respect to computing, UCȧF?TCȧGBCLRGȏCBȧ
ȧ ?ȧUGLBMUȧMDȧMNNMPRSLGRWȧDMPȧ#SPMNCȧRMȧGLTCQR
ȧ GLȧRFCȧCKCPEGLEȧLCUȧN?P?BGEKȧMDȧAMKNSRGLEȧ
ȧ BGQRPG@SRCBȧRMU?PBQȧRFCȧCBECQȧMDȧRFCȧLCR
ȧ UMPI
will also support the future deployment of
China is catching up fast and aims to lead by 2030.
For the EU, it is not so much a question of winning or
JMQGLEȩRFCȩP?ACȩ@SRȩMDȩȏLBGLEȩ?ȩU?WȩRMȩCK@P?ACȩRFCȩMN-
NMPRSLGRGCQȩMȎCPCBȩ@Wȩ'ȩGLȩ?ȩU?WȩRF?RȩGQȩFSK?LACL-
tred, ethical, secure, and true to our core values.
ȩ %ȩ?LBȩRFCȩ'M2
• With respect to data, we argue in favour of
learning from successful internet companies,
opening access to data and developing
 121 14 Summary and conclusions

ȩ GLRCP?ARGTGRWȩ UGRFȩ RFCȩ SQCPQȩ P?RFCPȩ RF?Lȩ HSQR

broadcasting data. In this way, we can BCTCJ
ȧ MNȧ CAMQWQRCKQȧ MDȧ NS@JGAȧ ?BKGLGQRP?RGMLQ
ȧ
ȧ ȏPKQ
ȧ?LBȧAGTGJȧQMAGCRW
ȧCLPGAFGLEȧRFCȧB?R?ȧ
to make iRȧȏRȧDMPȧ'ȧ?NNJGA?RGMLQȧPCQNMLBGLEȧ
We need
to #SPMNC?LȧLCCBQȧ to act together
5CȩQFMSJBȩCK@P?ACȩRFCȩMNNMPRSLGRGCQȩ?ȎMPBCBȩ@Wȩ
'ȩ@SRȩLMRȩSLAPGRGA?JJWȩ2FCȩ@J?AIȩ@MVȩAF?P?ARCPGQRGAQȩ
and act fast.
of most leading AI techniques make them opaque
even to specialists. AI systems are currently limited
RMȩ L?PPMUȩ ?LBȩ UCJJBCȏLCBȩ R?QIQ
ȩ ?LBȩ RFCGPȩ RCAF-
nologies inherit imperfections from their human We still know very little about how AI will impact
APC?RMPQ
ȩ QSAFȩ ?Qȩ RFCȩ UCJJPCAMELGQCBȩ @G?Qȩ CȎCARȩ the way we think, make decisions, relate to each
present in data. We should challenge the short- MRFCP
ȩ?LBȩFMUȩGRȩUGJJȩ?ȎCARȩMSPȩHM@Qȩ2FGQȩSLACPR?GL-
comings of AI and work towards strong evaluation ty can be a source of concern but is also a sign of
strategies, transparent and reliable systems, and
good human-AI interactions.
#RFGA?Jȧ ?LBȧ QCASPC@WBCQGELȧ ?JEMPGRFKQȧ ?PCȧ
APSAG?JȧRMȧ@SGJBȧRPSQRȧGLȧRFGQȧBGQPSNRGTCȧRCAFLMJMEW
opportunity. 2FCȧDSRSPCȧGQȧLMRȧWCRȧUPGRRCLȧ5CȧA?Lȧ
QF?NCȧ GRȧ @?QCBȧ MLȧ MSPȧ AMJJCARGTCȧ TGQGMLȧ MDȧ UF?Rȧ
DSRSPCȧUCȧUMSJBȧJGICȧRMȧF?TC. But we need to act
together and act fast.

but we also need the broader engagement of civil

society in the values to be embedded in AI and the
directions for future development.

This social engagement should be part of the ef-

fort to strengthen our resilience at all levels from
local, to national and European, across institutions,
industry and civil society. Developing local ecosys-
tems of skills, computing, data and applications
can foster the engagement of local communities,
respond to their needs, harness local creativity and
knowledge, and build a human-centred, diverse,
and socially driven AI.
 Footnotes 122

FOOTNOTES
1 4B—ABȵKB—MI>VBOP—>P—OBPB>O@E—@BKQOBP
—>@>ABJF@—FKPQFQRQFLKP
—>KA—@LJM>KFBP—TEF@E—E>SB—M>OQF@FM>QBA—FK—LKB—LO—JLOB—
— LC—QEB—CLIILTFKD—B@LKLJF@—>@QFSFQFBP—/"—MOL@BPPBP
—FKARPQOF>I—MOLAR@QFLK—>KA—J>OHBQFKD
—PMB@Fȵ@—'OBI>QBA—PBOSF@BP
— PBB— LU—
 1B@EKF@>IIV
—QEFP—AF>DO>J—J>FKIV—OBCBOP—QL—PRMBOSFPBA—IB>OKFKD
 ,BRO>I—KBQTLOH—JLABIP—CLO—@LJMRQFKD—TBOB—FKQOLAR@BA—FK—QEB—P—>KA—P—FKPMFOBA—?V—KBROL?FLILDV—*B>OKFKD—
— FP—JLABIIBA—>P—>—DO>AR>I—@E>KDB—FK—QEB—@LJMRQFKD—KBQTLOH—@LKKB@QFLKP—FK—>K>ILDV—QL—QEB—PVK>MQF@—JLAFȵ@>QFLKP
— Q>HFKD—MI>@B—FK—QEB—?FLILDF@>I—KBRO>I—KBQTLOH
 https://medium.com/politics-ai/an-overview-of-national-ai-strategies-2a70ec6edfd
 FRRNQ UUUPCSRCPQAMK ?PRGAJC SQSQ??PRGĻAG?JGLRCJJGECLAC RPSKN?BKGLGQRP?RGMLUGJJ?JJMU?GRMDPCCJWBCTCJMN
– GLSQMĸAG?JGB31) ,' $–
6 FRRNQ UUUB?PN?KGJ LCUQCTCLRQ 
7 FRRNQ UUUL?QB?OAMK ?PRGAJC SQQRMAIK?PICRGQ@GEECQRKMQRCVNCLQGTCGLUMPJB@SRSQCAMLMKWGQLMR
– RFCKMQRNPMBSARGTCAK–
8 FRRNQ UUUAL@AAMK    RCAFLMJMEWLMUK?ICQSN?OS?PRCPMDRFCQRMAIK?PICRFRKJ–
9 FRRNQ UUUQR?RGQR?AMK QR?RGQRGAQ  K?PICRA?NGR?JGX?RGMLMDSQRCAF?LBGLRCPLCRAMKN?LGCQ –
10 FRRNQ UUUUFGRCFMSQCEMT UNAMLRCLR SNJM?BQ   ?N==PCQC?PAFDWNBD–
11 FRRNQ UUUQR?RGQR?AMK QR?RGQRGAQ  UMPJBUGBCK?PICRQF?PCMDQC?PAFCLEGLCQ –
 http://www.internetlivestats.com/google-search-statistics/
 FRRNQ UUUQR?RGQR?AMK QR?RGQRGAQ  LSK@CPMDKMLRFJW?ARGTCD?AC@MMISQCPQUMPJBUGBC –
 ––FRRNQ UUUQJGBCQF?PCLCR '$PMLRGCPQ HCķBC?LRPCLBQ?LBBCTCJMNKCLRQGLBCCNJC?PLGLEPCQC?PAF–
 https://www.datacenterknowledge.com/google-alphabet/google-switching-self-driving-data-center-
– K?L?ECKCLRQWQRCK–
16 https://www.contino.io/insights/whos-using-aws
17 FRRNQ UUUAGD?PA? ?G N?LA?L?BG?L?PRGĻAG?JGLRCJJGECLACQRP?RCEW
18 'K—QEFP—SBOPFLK
—QEB—1#0—>K>IVPFP—LC—QEB—'—B@LPVPQBJ—FP—?>PBA—LK—A>Q>—@LJFKD—COLJ—JLOB—QE>K——A>Q>—PLRO@BP
—
— TEF@E—FK@IRAB—OUFS
—PD>OA
—! 'KPFDEQP
—!ORK@E?>PB
—"LT(LKBP
—#2—!-/"'0
—#Ç-—Ç>QPQ>Q
— ROB>R—3>K—"FGH—-O?FP
—
— 0Q>QFPQ>
—>KA—LQEBO—JFKLO—LKBP
19 FRRN CACSPMN?CS BGEGR?JQGLEJCK?PICR CL LCUQ CSKCK@CPQR?RCQQGELAMMNCP?RC?PRGĻAG?JGLRCJJGECLAC
FRRN CSPMN?CS P?NGB NPCQQPCJC?QC='.=CLFRK
——-K—QEFP—PR?GB@Q
—PBB—>IPL—!E>MQBO——FK—QEFP—OBMLOQ
———RKF@LOK—FP—>—MOFS>QBIVEBIA—PQ>OQRM—@LJM>KV—S>IRBA—>Q—LSBO—20——?FIIFLK
FRRNQ UUUQR?RGQR?AMK QR?RGQRGAQ  LSK@CPMD?ARGTCUCAF?RKCQQCLECP?AAMSLRQ –
FRRNQ KNUCGVGLOOAMK Q 6-'0O4-1)(1* *UGK*–
— $ILOFAF——J>HBP—>—AFPQFK@QFLK—?BQTBBK—E>OA—>KA—PLCQ—AFDFQ>I—BQEF@P
—TEBOB—E>OA—MOB@BABP—QEB—CLOJRI>QFLK—
— LC—KBT—OBDRI>QFLKP—LO—@E>IIBKDBP—QEB—BUFPQFKD—LKBP
—>KA—PLCQ—@LJBP—>CQBO—QEB—@LJMIF>K@B—?V—@LKPFABOFKD—TE>Q—
— @LRIA—?B—ALKB—|LSBO—>KA—>?LSB—QEB—BUFPQFKD—OBDRI>QFLK}—1ERP
—PLCQ—BQEF@P—FP—>—DLLA—>MMOL>@E—TEBOB—>—KLOJ>QFSB—
— CO>JBTLOH
—PR@E—>P—OBPMB@Q—LC—ERJ>K—OFDEQP
—FP—>IOB>AV—FK—MI>@B.
 — 1EFP—PB@QFLK—E>P—@LKPFABOBA—PQO>QBDFBP—>KA—MR?IF@—MLIF@V—AL@RJBKQP—OBIB>PBA—?V—#2—FKPQFQRQFLKP—>KA—MR?IF@—BKQFQFBP
— FK—$O>K@B—>KA—QEB—2)—
 P—@LAFȵBA—FK—OQ——1#2—|1EB—2KFLK—FP—CLRKABA—LK—QEB—S>IRBP—LC—OBPMB@Q—CLO—ERJ>K—AFDKFQV
—COBBALJ
—ABJL@O>@V
—
— BNR>IFQV
—QEB—ORIB—LC—I>T—>KA—OBPMB@Q—CLO—ERJ>K—OFDEQP
—FK@IRAFKD—QEB—OFDEQP—LC—MBOPLKP—?BILKDFKD—QL—JFKLOFQFBP—
 123 Footnotes

— 1EBPB—S>IRBP—>OB—@LJJLK—QL—QEB—+BJ?BO—0Q>QBP—FK—>—PL@FBQV—FK—TEF@E—MIRO>IFPJ
—KLKAFP@OFJFK>QFLK
—QLIBO>K@B
—
— GRPQF@B
—PLIFA>OFQV—>KA—BNR>IFQV—?BQTBBK—TLJBK—>KA—JBK—MOBS>FI}—1EB—#ROLMB>K—!E>OQBO—LC—$RKA>JBKQ>I—
— /FDEQP
—OB@LDKFPBP
—>JLKD—LQEBOP
—QEB—OFDEQ—QL—ERJ>K—AFDKFQV
—MOFS>@V
—MOLQB@QFLK—LC—MBOPLK>I—A>Q>
—COBBALJ
— LC—BUMOBPPFLK—>KA—FKCLOJ>QFLK
—BNR>IFQV
—KL—AFP@OFJFK>QFLK
—QEB—OFDEQP—LC—QEB—@EFIA—>KA—QEB—BIABOIV—>KA—MBOPLKP—
— TFQE—AFP>?FIFQFBP—>P—TBII—>P—QEB—OFDEQP—QL—@LKPRJBO—MOLQB@QFLK
—DLLA—>AJFKFPQO>QFLK—>KA—>—C>FO—QOF>I
 — ,LQB—QE>Q—BSBK—FC—OQ——%"Ç/—OB@LDKFPBP—BU@BMQFLKP—QL—QEFP—OFDEQ
—QEB—A>Q>—PR?GB@Q—TFII—>IT>VP—HBBM—QEB—OFDEQ—
— zQL—L?Q>FK—ERJ>K—FKQBOSBKQFLK—LK—QEB—M>OQ—LC—QEB—@LKQOLIIBO
—QL—BUMOBPP—EFP—LO—EBO—MLFKQ—LC—SFBT—>KA—QL—@LKQBPQ—
— QEB—AB@FPFLK{
 http://www.algoaware.eu/
 — !LRK@FI—"FOB@QFSB—  ##!—LC——(RIV——LK—QEB—>MMOLUFJ>QFLK—LC—QEB—I>TP
—OBDRI>QFLKP—>KA—>AJFKFPQO>QFSB—
— MOLSFPFLKP—LC—QEB—+BJ?BO—0Q>QBP—@LK@BOKFKD—IF>?FIFQV—CLO—ABCB@QFSB—MOLAR@QP
—-(—*——
 — JLKD—QEB—MOL?IBJ>QF@—@>QBDLOFBP—>OB—QEB—P@LMB—OBCBOBK@B—QL—MOLAR@QP—>KA— !—OBI>QFLKP—>P—TBII—>P—
— QEB—OBNRFOBJBKQP—RMLK—QEB—@LKPRJBO—QL—?BKBȵQ—COLJ—QEB—PQOF@Q—IF>?FIFQV—OBDFJB—KLQ>?IV—QEB—?ROABK—LC—QEB—MOLLC—
— @LK@BOKFKD—QEB—ABCB@Q—LC—QEB—MOLAR@Q
—QEB—>@QR>I—A>J>DB—>KA—QEB—@>RP>I—IFKH—?BQTBBK—QELPB—
 — !>II—CLO—>MMIF@>QFLKP—CLO—QEB—PBIB@QFLK—LC—JBJ?BOP—LC—QEB—#UMBOQ—%OLRM—LK—*F>?FIFQV—>KA—,BT—1B@EKLILDFBP
$LO—>—ABQ>FIBA—LSBOSFBT—LC—QEB—IBD>I—@E>IIBKDBP—@LK@BOKFKD—IF>?FIFQV
—PBB—0Q>ȴ—4LOHFKD—"L@RJBKQ—LK—IF>?FIFQV—
— CLO—BJBODFKD—AFDFQ>I—QB@EKLILDFBP—04"——>@@LJM>KVFKD—QEB—!LJJRKF@>QFLK—LK—OQFȵ@F>I—FKQBIIFDBK@B—
— CLO—#ROLMB—!-+——#!—?—"FOB@QIV—@LKKB@QBA—QL—QEB—AFP@RPPFLK—LK—IF>?FIFQV
—FQ—JRPQ—>IPL—?B—KLQBA—
— QE>Q
—FK—FQP—/BPLIRQFLK
—QEB—#ǗFKSFQBP—QEB—#!—QL—@LKPFABO—>—PMB@Fȵ@—IBD>I—PQ>QRP—CLO—OL?LQP
—TEF@E—@LRIA—IB>A—
— QL—OB@LDKFPFKD—IBD>I—MBOPLK>IFQV—CLO—>RQLKLJLRP—OL?LQP—#Ǘ
 Q—QEB—PB@QLO>I—IBSBI
—QEB—#ǗE>P—@>IIBA—LK—QEB—!LJJFPPFLK—QL—MR?IFPE—>—IBDFPI>QFSB—MOLMLP>I—LK—>@@BPP—QL—FKSBEF@IB
— A>Q>—#Ǘ>—4FQE—QEB—OBIB>PB—LC—QEB—1EFOA—+L?FIFQV—Ç>@H>DB
—QEB—!LJJFPPFLK—E>P—>KKLRK@BA—>—/B@LJJBKA>
— QFLK—QE>Q
—>JLKD—LQEBO—QEFKDP
—TFII—AB>I—TFQE—|>—A>Q>—DLSBOK>K@B—CO>JBTLOH—QE>Q—BK>?IBP—A>Q>—PE>OFKD
—FK—IFKB—TFQE—
— QEB—FKFQF>QFSBP—LC—QEB——">Q>—Ç>@H>DB
—>KA—TFQE—A>Q>—MOLQB@QFLK—>KA—MOFS>@V—IBDFPI>QFLK}
 1EB—MOLMLP>I—FP—KLT—?BFKD—AFP@RPPBA—?V—QEB—!LJJFPPFLK
—QEB—!LRK@FI—>KA—QEB—Ç>OIF>JBKQ—'Q—JRPQ—?B—KLQBA—QE>Q—
— ?LQE—QEB—#Ǘ>KA—QEB—!LRK@FI—E>SB—MOLMLPBA—QL—BKI>ODB—QEB—P@LMB—QL—@RIQRO>I—FKPQFQRQFLKP—>KA—QL—@OB>QB
— >—KBT—LMQFLK>I—BU@BMQFLK—FK—>AAFQFLK—QL—QE>Q—MOLMLPBA—?V—QEB—#!—
 0BB—!LRK@FI—LC—#ROLMB
—#UMI>K>QLOV—OBMLOQ
—ÇOLQL@LI—>JBKAFKD—QEB—!LKSBKQFLK—CLO—QEB—ÇOLQB@QFLK—LC—'KAFSFAR>IP—
— TFQE—OBD>OA—QL—RQLJ>QF@—ÇOL@BPPFKD—LC—ÇBOPLK>I—">Q>—#10—,L—
—A—EL@—!LJJFQQBB—LK—">Q>—ÇOLQB@QFLK—
— !&"1
—QE—0BPPFLK—LC—QEB—!LJJFQQBB—LC—+FKFPQBOP—#IPFKLOB
—"BKJ>OH
——+>V——1EB—#UMILO>QLOV
— /BMLOQ—OB>AP—z1EB—KLQFLK—LC—|FABKQFȵ>?IB}—OBCBOP—KLQ—LKIV—QL—QEB—FKAFSFAR>I}P—@FSFI—LO—IBD>I—FABKQFQV—>P—PR@E
—?RQ—>IPL—
— QL—TE>Q—J>V—>IILT—QL—|FKAFSFAR>IFPB}—LO—PFKDIB—LRQ—>KA—QERP—>IILT—QL—QOB>Q—AFȴBOBKQIV—LKB—MBOPLK—COLJ—LQEBOP—
— 1EFP—|FKAFSFAR>IFP>QFLK}—@LRIA—?B—ALKB
—CLO—FKPQ>K@B
—?V—OBCBOOFKD—QL—EFJ—LO—EBO—PMB@Fȵ@>IIV
—LO—QL—>—ABSF@B—LO—
— >—@LJ?FK>QFLK—LC—ABSF@BP—@LJMRQBO
—JL?FIB—MELKB
—@>JBO>
—D>JFKD—ABSF@BP
—BQ@—LK—QEB—?>PFP—LC—>K—FABKQFȵ@>QFLK—
— KRJ?BO
—>—MPBRALKVJ
—?FLJBQOF@—LO—DBKBQF@—A>Q>
—IL@>QFLK—A>Q>
—>K—'Ǘ>AAOBPP
—LO—LQEBO—FABKQFȵBO—1EB—RPB—
— LC—>—MPBRALKVJ—LO—LC—>KV—AFDFQ>I—FABKQFȵBO AFDFQ>I—FABKQFQV—ALBP—KLQ—IB>A—QL—>KLKVJFP>QFLK—LC—QEB—A>Q>—>P—QEB—A>Q>
— PR?GB@Q—@>K—PQFII—?B—FABKQFȵ>?IB—LO—FKAFSFAR>IFPBA—ÇPBRALKVJLRP—A>Q>—FP—QERP—QL—?B—@LKPFABOBA—>P—MBOPLK>I—A>Q>—
— >KA—FP—@LSBOBA—?V—QEB—MOLSFPFLKP—LC—QEB—!LKSBKQFLK—1EB—NR>IFQV—LC—QEB—MPBRALKVJFP>QFLK—QB@EKFNRBP—>MMIFBA—PELRIA
— ?B—ARIV—Q>HBK—FKQL—>@@LRKQ—TEBK—>PPBPPFKD—QEB—>MMOLMOF>QBKBPP—LC—P>CBDR>OAP—FJMIBJBKQBA—QL—JFQFD>QB—QEB—OFPHP—
— QL—A>Q>—PR?GB@QP{
 FRRNQ UUUCNMMPE J?UNP?ARGAC JCE?JRCVRQ FRKJ ESGBCJGLCQ C E=GG===FRK–
 —0B—2)—%LSBOKJBKQ}P—OBPMLKPB—QL—QEB—*LOAP—0BIB@Q—!LJJFQQBB—'—OBMLOQ
—(RKB—
 — MMOL>@EBP—ABPFDKBA—QL—QBPQ—QEB—?LRKA>OFBP—LC—@LJMRQBO—MOLDO>JJFKD—I>KDR>DB—ABPFDK
—>P—>—MOLLC—LC—@LK@BMQ
 FRRNQ UUU@JMMK@CPEAMK LCUQ ?PRGAJCQ  HSQRFMUQF?JJMUGQRFC?PRGĻAG?JGLRCJJGECLACR?JCLRNMMJ
 https://www.elementsofai.com
0BB
—CLO—BU>JMIB
—&LBHPQO>—>KA—S>K—0IRGFP——
 Footnotes 124

/B>ABOP—J>V—>IPL—?B—FKQBOBPQBA—FK—|&2+',1}
—>K—FKQBOAFP@FMIFK>OV—(/!—MOLGB@Q—>FJFKD—QL—RKABOPQ>KA—QEB—FJM>@Q—
— LC—J>@EFKB—FKQBIIFDBK@B—LK—ERJ>K—?BE>SFLRO
—TFQE—>—CL@RP—LK—@LDKFQFSB—>KA—PL@FLBJLQFLK>I—@>M>?FIFQFBP—
— >KA—AB@FPFLKJ>HFKD—PBB—FRRNQ CACSPMN?CS HPA AMKKSLGRGCQ AMKKSLGRW FSK?GLR
 $LO—>—JLOB—@LJMOBEBKPFSB—LSBOSFBT—>KA—>PPBPPJBKQ—LC—QEB—OBIBS>KQ—OBPB>O@E—IFQBO>QROB—>KA—>K—BJMFOF@>I—
— FKSBPQFD>QFLK—FKQL—@E>KDFKD—L@@RM>QFLK>I—M>QQBOKP—FK—#ROLMB
—PBB—QEB—(/!}P—CLOQE@LJFKD—OBMLOQ—LK—QEB—CRQROB—LC—TLOH
 IQELRDE—QEB—P@>IB—LC—OBAFPQOF?RQFLK—CLO—B@LKLJF@—DOLTQE—QL—@LKQFKRB—>Q—>—PQB>AV—M>@B—TLRIA—E>SB—QL—?B—QORIV—
— BKLOJLRP
—PL—BKLOJLRP—>P—QL—CRIIV—@LJMBKP>QB—QEB—>IJLPQ—@LJMIBQB—PR?PQFQRQFLK—LC—ERJ>K—I>?LRO—?V—J>@EFKBP
 0FKDRI>OFQV—OBCBOP—QL—QEB—P@BK>OFL—LC—PBICIB>OKFKD—>KA—PBICMOLDO>JJFKD—J>@EFKBP—PQ>OQFKD—>—ORK>T>V—PMFO>I—
— LC—FJMOLSBJBKQP—QE>Q—LSBOQ>HB—ERJ>K—FKQBIIFDBK@B—TFQE—@>Q>PQOLMEF@—@LKPBNRBK@BP—
 —DBKBO>IxMROMLPB—QB@EKLILDV—OBCBOP—QL—>—HBV—BK>?IFKD—QB@EKLILDV—?OFKDFKD—J>GLO—@E>KDBP—QL—PL@FBQV
—
— PR@E—>P—BIB@QOF@FQV—LO—'1
 1EB—QBOJ—@V?BOMEVPF@>I—PVPQBJP—OBCBOP—QL—MEVPF@>I—L?GB@QP
—LCQBK—FKQBO@LKKB@QBA—LSBO—AFDFQ>I—KBQTLOHP
—
— QE>Q—>OB—M>OQIV—@LKQOLIIBA—?V—PLCQT>OB—>IDLOFQEJP—>KA—>OB—@>M>?IB—LC—FKQBO>@QFKD—TFQE—QEB—MEVPF@>I—TLOIA—
— BD—>RQLKLJLRP—SBEF@IBP
 'K—'0-—PQ>KA>OAP
—QEB—O>QFL—?BQTBBK—QEB—QLQ>I—A>Q>—@BKQOB—FKMRQ—MLTBO—>KA—QEB—'1—MLTBO—FP—ABȵKBA—>P—MLTBO—
— RQFIFP>QFLK—BȴB@QFSBKBPP—Ç2#—1VMF@>I—A>Q>—@BKQOBP—E>SB—>—Ç2#—LC—
—TEFIB—QEB—JLPQ—Bȷ@FBKQ—LKBP—E>SB
— >—Ç2#—LC—
FRRN UUUGC?MPE NS@JGA?RGMLQ DPCCNS@JGA?RGMLQ NS@JGA?RGML "GEGR?JGX?RGML?LB#LCPEWNBD–
FRRNQ CR?J@JEMT NS@JGA?RGMLQ SLGRCBQR?RCQB?R?ACLRCPCLCPEW
1EFP—FP—>—#ROLMB>K—FKFQF>QFSB—J>K>DBA—?V—QEB—(/!
—FRRNQ CACSPMN?CS HPA CL CLCPEWCĸAGCLAW AMBCAMLBSAR
datacentres
 $FUQROBP—+>OHBQFKD—*QA—S—-V—3BFHH>RP—?—! 
—  
—$FUQROBP—+>OHBQFKD—*QA—S—0SBKPH>—0MBI—?—
— ! 
—  — OFQFPE—&LOPBO>@FKD— L>OA—*QA—S—4FIIF>J—&FII—! 
—  —$FUQROBP—+>OHBQFKD—*QA—S—
— -ÇǗ! 
—  —
 FRRNQ UUUCSPMNC?LB?R?NMPR?JCS QGRCQ BCD?SJR ĻJCQ CBN=APC?RGLE=T?JSC=RFPMSEF=MNCL=B?R?=NBD–
 0BB—FRRNQ CLUGIGNCBG?MPE UGIG "?R?=QMTCPCGELRW
 1EFP—@E>MQBO—FP—?>PBA—LK—"R@E OLTK
—+>OQBKP—>KA—+RBIIBO*>KDBO——$LO—JLOB—FKABMQE—AFP@RPPFLK—
— LK—QEB—FPPRBP—O>FPBA—MIB>PB—OBCBO—QL—QE>Q—MR?IF@>QFLK
 #U>JMIBP—FK@IRAB—QEB—/LJB—"B@I>O>QFLK—LC—
—QEB—/BȶB@QFLK—M>MBO—LK—&>OKBPPFKD—%IL?>IFP>QFLK
—QEB—(LFKQ—
— !LJJRKF@>QFLK—LK—|—0QO>QBDF@—MMOL>@E—QL—/BPFIFBK@B—FK—QEB—#2}P—#UQBOK>I—@QFLK}
 1EBV—>OB—RPR>IIV—FKAF@>QBA—>P—OBPFIFBK@B—FKAF@>QLOP—1EB—CLOQE@LJFKD—OBMLOQ—LK—/BPFIFBK@B—OBCBOP—QL—
— QEBJ—>P—|BUMBOQ—?>PBA—FKAF@>QLOP}
—>KA—LȴBOP—>—JLOB—FKABMQE—AFP@RPPFLK
%LSBOKJBKQ—MLIF@FBP—TBOB—BS>IR>QBA—?V—RPFKD—QEB—>SBO>DB—O>KHFKD—>@OLPP—QEB——PR?>PMB@QP—LC—QEB—QEOBB—J>FK—
— AFJBKPFLKP—OB>AFKBPP
—FJMIBJBKQ>QFLK—>KA—FJM>@Q
—PBB https://opendatabarometer.org/leadersedition/
– KCRFMBMJMEW –DMP–BCR?GJQ).—1EB—BAR@>QFLK—ȵBIAP—MOLUV—QEB—@LK@BMQ—LC—01#+—?V—@LKPFABOFKD—K>QRO>I—P@FBK@BP
—
— J>QEBJ>QF@P—>KA—PQ>QFPQF@P
—FKCLOJ>QFLK—>KA—@LJJRKF@>QFLK—QB@EKLILDFBP
—A>Q>?>PB—>KA—KBQTLOH—ABPFDK—
— >KA—>AJFKFPQO>QFLK
—>KA—PLCQT>OB—>KA—>MMIF@>QFLKP—ABSBILMJBKQ—>KA—>K>IVPFP—1EB—PE>OBP—TBOB—>SBO>DBA—LSBO
— QEB—VB>OP——
60 $LO—B>@E—S>OF>?IB
—>—S>IRB—LC—LKB—FKAF@>QBP—QE>Q—QEB—RKABOIVFKD—JB>PROB—FP—>?LSB—QEB—@OLPP@LRKQOV—>SBO>DB
—TEFIB
— >—JFKRP—LKB—JB>KP—QE>Q—FQ—FP—?BILT—1EB—PBOFBP—>OB—QEBK—PQ>@HBA—KBUQ—QL—B>@E—LQEBO—1EBOBCLOB
—>—S>IRB—LC——
— JB>KP—?BFKD—?BILT—>SBO>DB—LK—>II—QEOBB—S>OF>?IBP
61 "BȵKBA—>P—E>SFKD—BFQEBO—PR@@BPPCRIIV—@LJMIBQBA—>—QBOQF>OVIBSBI—BAR@>QFLK
—LO—KLQ—CLOJ>IIV—NR>IFȵBA—?RQ—BJMILVBA—
— FK—>K—01—L@@RM>QFLK—TEBOB—QBOQF>OV—IBSBI—NR>IFȵ@>QFLKP—>OB—KLOJ>IIV—OBNRFOBA
 – FRRN QNJ?RDMPKHPACACSPMN?CS BGEGR?JGLLMT?RGMLFS@Q
 125 References

REFERENCES
@BJLDIR
—"—>KA—RQLO
—"
——|0HFIIP
—1>PHP—>KA—1B@EKLILDFBP—'JMIF@>QFLKP—CLO—#JMILVJBKQ—>KA—#>OKFKDP}—'K Handbook
– MD–*?@MP–#AMLMKGAQ
——, #/—4LOHFKD—Ç>MBO—0BOFBP
—#IPBSFBO
——@BJLDIR
—"—>KA—/BPQOBML
—/
——2FC–0?AC– CRUCCL–+?AFGLC–?LB–+?L–'KNJGA?RGMLQ–MD–2CAFLMJMEW–DMP–%PMURF
–$?ARMP–1F?PCQ–
– ?LB–#KNJMWKCLR–4LOHFKD—Ç>MBO——,>QFLK>I— ROB>R—LC—#@LKLJF@—/BPB>O@E https://doi.org/10.3386/w22252
——@BJLDIR
—"—>KA—/BPQOBML
—/
—–SRMK?RGML–?LB–*MA?J–*?@MP–+?PICRQ https://doi.org/10.3386/w23285
@BJLDIR
—"—>KA—/BPQOBML
—/
—?—|1EB—/>@B—?BQTBBK—+>K—>KA—+>@EFKB—'JMIF@>QFLKP—LC—1B@EKLILDV—CLO—%OLTQE
—$>@QLO—
— 0E>OBP
—>KA—#JMILVJBKQ}–KCPGA?L–#AMLMKGA–0CTGCU—
—MM—. https://www.aeaweb.org/articles?id=10.1257/
aer.20160696
——@BJLDIR
—"—>KA—/BPQOBML
—/
—@ SRMK?RGML–?LB–LCU–R?QIQ–RFC–GKNJGA?RGMLQ–MD–RFC–R?QI–AMLRCLR–MD–RCAFLMJMEW–DMP–J?@MSP–
– BCK?LB
–JFJBL
—, #/
—RDRPQ—
——@BJLDIR
—"—>KA—/BPQOBML
—Ç
—>—|OQFȵ@F>I—'KQBIIFDBK@B
—RQLJ>QFLK—>KA—4LOH}. 110,–#JCARPMLGA–(MSPL?J
https://doi.org/10.2139/ssrn.3098384
——DEFLK
—Ç
—(LKBP
— —(—>KA—(LKBP
—!
——|OQFȵ@F>I—'KQBIIFDBK@B—>KA—#@LKLJF@—%OLTQE}—,?RGML?J– SPC?S–MD–#AMLMKGA–
– 0CQC?PAF
—KL—T
——'—,LT——/BMLOQ https://ainowinstitute.org/AI_Now_2017_Report.pdf
——'—,LT——IDLOFQEJF@—'JM>@Q—PPBPPJBKQP——ÇO>@QF@>I—$O>JBTLOH—CLO—ÇR?IF@—DBK@V—@@LRKQ>?FIFQV—
https://ainowinstitute.org/aiareport2018.pdf
——IABTBOBIA
—&—+
—"FDKRJ
—3—>KA—1>K
—6——|"BPFDK—CLO—S>IRBP—FK—PLCQT>OB—ABSBILMJBKQ}—'K—S>K—ABK—&LSBK—(
—3BOJ>>P—
— Ç—#—>KA—S>K—AB—ÇLBI
—'—#AP—&?LB@MMI–MD–#RFGAQ
–4?JSCQ
–?LB–2CAFLMJMEGA?J–"CQGEL
–QMSPACQ
–RFCMPW
–T?JSCQ–?LB–?NNJGA?RGML–
domains
MM—— BOIFK—0MOFKDBO
——IBPPF
—*
— BK@WRO
—Ç
—!>JMLILKDL
—$—BQ—>I— 2FC–PCQGJGCLAC–MD–#3–+CK@CP–1R?RCQ–RM–RFC–ĻL?LAG?J–?LB–CAMLMKGA–APGQGQ–
What are the characteristics of resilient behaviour?—(/!—0@FBK@B—CLO—ÇLIF@V—/BMLOQ—http://publications.jrc.ec.europa.eu/
– PCNMQGRMPW @GRQRPC?K (0! HPA=PCQGJGCLAC=APGQGQ=NGJMR=UGRFGBCLRGĺCPQNBD
IBU>KABO
— 
——|4EV—OL?LQP—EBIMBA—"LK>IA—1ORJM—TFK}—+'2–2CAFLMJMEW–0CTGCU–https://www.technologyreview.
com/s/611422/why-robots-helped-donald-trump-win/
JLABF
—"
—-I>E
—!
—0QBFKE>OAQ
—(—BQ—>I——!MLAPCRC–.PM@JCKQ–GL–'–1?DCRW–https://arxiv.org/abs/1606.06565
OKQW
—+
—%OBDLOV
—1
—>KA—7FBO>EK
—2——|1EB—/FPH—LC—RQLJ>QFLK—CLO—(L?P—FK—-#!"—!LRKQOFBP——!LJM>O>QFSB—K>IVPFP}
-#!"–1MAG?J
–#KNJMWKCLR–?LB–+GEP?RGML–5MPIGLE–.?NCPQ–——. https://doi.org/10.1787/5jlz9h56dvq7-en
QEBV
—0——|1EB—FJM>@Q—LC—J>@EFKB—IB>OKFKD—LK—B@LKLJF@P}
—#AMLMKGAQ–MD–PRGĻAG?J–'LRCJJGECLAC–3LGTCPQGRW–
– MD–!FGA?EM–.PCQQ
RQLO
—"—&
—*BSV
—$—>KA—+ROK>KB
—/—(——|1EB—0HFII—!LKQBKQ—LC—/B@BKQ—1B@EKLILDF@>I—!E>KDB—K—#JMFOF@>I—#UMILO>QFLK}
The /S?PRCPJW–(MSPL?J–MD–#AMLMKGAs. https://doi.org/10.1162/003355303322552801
——RQLO
—"
—>KA—0>ILJLKP
———|/L?L@>IVMPB—,LTx"LBP—ÇOLAR@QFSFQV—%OLTQE—1EOB>QBK—#JMILVJBKQ}—#! –$MPSK–
– ML–!CLRP?J– ?LIGLE–—(RKB—— https://conference.nber.org/confer/2017/AIf17/Autor.pdf
>@@>I>
—+
—!ROO>K
—!
—%>OBQQ
—"
—*FHBKP
—0
—/>L
—
—/RDDIBP
—
—>KA—0EBE>?
—+——–'–NPCBGARGMLQ––GLQGEFRQ
– RM–QF?NC–@SQGLCQQ–QRP?RCEW–ÇT!—https://www.pwc.com/us/en/advisory-services/assets/ai-predictions-2018-report.pdf
—— >H>OAGFBS>
—+—>KA—%BEI
—/——|!OFQF@>I—MMOL>@EBP—QL—!LJJRKF@>QFLK—1B@EKLILDV—1EB—Ç>PQ—$FSB—6B>OP}—Annals
– MD–RFC–'LRCPL?RGML?J–!MKKSLGA?RGML–QQMAG?RGML–—
—-KIFKB—ȵOPQ—https://doi.org/10.1080/23808985.2017.1374201
—— >IATFK
—/—>KA—3FEOF>I>
—3——|%IL?>IFP>QFLK—J>V—PLLK—>@@BIBO>QB—>D>FK—x—QFJB—QL—DBQ—ALJBPQF@—MLIF@FBP—OFDEQ}—
https://voxeu.org/article/globalisation-may-soon-accelerate-again-time-get-domestic-policies-right
 References 126

—— >OQLK
—"
—4LBQWBI
—(
—0BLKD
—(
—>KA—1F>K
—.—–PRGĻAG?J–'LRCJJGECLAC–'KNJGA?RGMLQ–DMP–!FGL?–+@)FKPBV—%IL?>I—'KPQFQRQB—
3LI—"FP@RPPFLK—/BQOFBSBA—COLJ www.mckinsey.com/mgi.
>PI>KAWB
—0——|1EB—/LIB—LC—QEB—'1—/BSLIRQFLK—FK—)KLTIBADB—"FȴRPFLK
—'KKLS>QFLK—>KA—/B>IIL@>QFLK}—'K——+CCRGLE–
– .?NCPQ ,L——0L@FBQV—CLO—#@LKLJF@—"VK>JF@P
—— BIM>BJB
—1
—)BKKBAV
—(
—/>J>@E>KAO>K
—
—0@>PPBII>QF
— 
—>KA—1>K>H>
—$——|0L@F>I—OL?LQP—CLO—BAR@>QFLK——OBSFBT}—
– 1AGCLAC–0M@MRGAQ
–
—B>>Q
BODJ>KK
—"—>KA—+LOOFP
—0——'LDMPK?RGML–"CQGEL–?–3LGĻCB–.CPQNCARGTC—!LTIBP—$LRKA>QFLK—"FP@RPPFLK—Ç>MBO—
FRRNQ AMUJCQW?JCCBS QGRCQ BCD?SJR ĺJCQ ĺJCQ NS@ B BNBD
—— FDDFL
— 
—,BIPLK
— 
—>KA—*>PHLS
—Ç——.MGQMLGLE–RR?AIQ–?E?GLQR–1SNNMPR–4CARMP–+?AFGLCQ—https://arxiv.org/
abs/1206.6389
FGHBO
—4—BQ—>I——#AP—2FC–1MAG?J–!MLQRPSARGML–MD–2CAFLMJMEGA?J–1WQRCKQ–,CU–"GPCARGMLQ–GL–RFC–1MAGMJMEW–?LB–&GQRMPW–
– MD–2CAFLMJMEW–!?K@PGBEC—+'1—ÇOBPP
LODJ>KK
———2CAFLMJMEW–?LB–RFC–!F?P?ARCP–MD–!MLRCKNMP?PW–*GDC––.FGJMQMNFGA?J–'LOSGPW—!EF@>DL—2KFSBOPFQV—
— LC—!EF@>DL—ÇOBPP
— LQPJ>K
—/——| FD—A>Q>—JBBQP— FD— OLQEBO—>P—!EFK>—JLSBP—QL—O>QB—FQP—@FQFWBKP}—4FOBA—-@QL?BO——/BQOFBSBA—COLJ
https://www.wired.co.uk/article/chinese-government-social-credit-score-privacy-invasion
—— LR@EBO
—Ç
—,>P@FJBKQL
—0
—3BPKF@IRGBSF@
—*—>KA—%RFJ>O>BP—ÇBOBFO>
——–#RFGAQ–"G?JMESCQ–#VNCPGCLAGLE–#RFGAQ–
– 2FPMSEF–{2FGLEQ|–̨–-NCL–'M2
–!GTGJ–"PMLCQ–?LB–5C?P?@JC–1CLQMPQ–*RUBJ?LROD—ÇR?IF@>QFLKP—-ȷ@B—LC—QEB—#ROLMB>K—2KFLK—
—— ORKA>DB
—+
—!I>OH
—(
—IIBK
—%—!
—$IVKK
—!
—$>ONRE>O
—0
—!OLLQLC
—/
—>KA— OVPLK
—(——|1EB—+>IF@FLRP—2PB—LC—OQFȵ@F>I—
— 'KQBIIFDBK@B—$LOB@>PQFKD
—ÇOBSBKQFLK
—>KA—+FQFD>QFLK}—'KCLOJ>QFLK—0L@FBQV—ÇOLGB@Q—$RQROB—LC—&RJ>K—'KPQFQRQB—/BQOFBSBA—
— COLJ https://www.repository.cam.ac.uk/bitstream/handle/1810/275332/1802.07228.pdf?sequence=1
—— OVKGLICPPLK
—#
—+FQ@EBII
—1
—>KA—/L@H
—"——|4E>Q—!>K—+>@EFKBP—*B>OK
—>KA—4E>Q—"LBP—'Q—+B>K—CLO—-@@RM>QFLKP—
— >KA—QEB—#@LKLJV}—#–.?NCPQ–?LB–.PMACCBGLEQ—3LI—. https://www.aeaweb.org/articles?id=10.1257/pandp.20181019
—— OVKGLICPPLK
—#
—/L@H
—"—>KA—0VSBOPLK
—!——2FC–NPMBSARGTGRW–(ASPTC–FMU–GLR?LEG@JCQ–AMKNJCKCLR–ECLCP?J–NSPNMQC
technologies
—, #/—TLOHFKD—M>MBO—KL—
—-@QL?BO—
—— OVKGLICPPLK
—#
—/L@H
—"
—0VSBOPLK
—!
—BQ—>I——PRGĻAG?J–'LRCJJGECLAC–?LB–RFC–+MBCPL–.PMBSARGTGRW–.?P?BMV––!J?QF–
– MD–#VNCAR?RGMLQ–?LB–1R?RGQRGAQ—http://www.nber.org/papers/w24001
——!>QE
—!
—4>@EQBO
—0
—+FQQBIPQ>AQ
— —BQ—>I——|OQFȵ@F>I—'KQBIIFDBK@B—>KA—QEB—|%LLA—0L@FBQV}—QEB—20
—#2
—>KA—2)—>MMOL>@E}
1AGCLAC–?LB–#LEGLCCPGLE–#RFGAQ
————
—!E>OFPF
—3
—*FBJ
—!
—%ŃJBW
—#——|,LSBIQV?>PBA—@LDKFQFSB—MOL@BPPBP—FK—RKPQOR@QROBA—JRPF@J>HFKD—MI>V—PBQQFKDP—
— FK—B>OIV—@EFIAELLA}–.PMACCBGLEQ–th–(MGLR–'###–'LRCPL?RGML?J–!MLDCPCLAC–MD–"CTCJMNKCLR–?LB–*C?PLGLE–?LB–#NGECLCRGA
– 0M@MRGAQ–'!"*#MF/L?
—MM—
!EF>@@EFL
—$
—ÇBQOLMLRILP—%
—>KA—ÇF@EIBO—"——2FC–'KN?AR–MD–'LBSQRPG?J–0M@MRQ–ML–#3–#KNJMWKCLR–?LB–5?ECQ–
– –*MA?J–*?@MSP–+?PICR–NNPM?AF— ORBDBI—4LOHFKD—Ç>MBOP— ORBDBI
!ELIIBQ
—$—–"CCN–*C?PLGLE–UGRF–.WRFML–+?LLGLE–
!ELRIAB@ELS>
———|$>FO—MOBAF@QFLK—TFQE—AFPM>O>QB—FJM>@Q——PQRAV—LC—?F>P—FK—OB@FAFSFPJ—MOBAF@QFLK—FKPQORJBKQP}—
GE–"?R?
————
!L@H?ROK
—'—+
—&BKABOPLK
—/
—>KA—0QBOK
—0——2FC–'KN?AR–MD–PRGĻAG?J–'LRCJJGECLAC–ML–'LLMT?RGML——, #/—
— 4LOHFKD—Ç>MBO—0BOFBP
——!LB@HBI?BODE
—+——|1B@EKLILDV—>KA—QEB—DLLA—PL@FBQV——MLIBJF@>I—BPP>V—LK—PL@F>I—LKQLILDV
—MLIFQF@>I—MOFK@FMIBP
—
— >KA—OBPMLKPF?FIFQV—CLO—QB@EKLILDV}–2CAFLMJMEW–GL–1MAGCRW—
——
—!LILJ?L
—$
—+ROOR
—+—$
—1LPLKF
—0——|1EB—MLPQFKQBOJBAF>QFLK—LC—QORQE—,BTPJ>HFKD—COLJ—+BAF>—!LJM>KFBP—
— QL—ÇI>QCLOJ}–!MKSLGA?XGMLG–QMAG?JG
–—
—!LJJFPPFLK—,>QFLK>IB—'KCLOJ>QFNRB—BQ—*F?BOQB—!,'*——How can humans keep the upper hand? The ethical matters
P?GQCB–@W–?JEMPGRFKQ–?LB–?PRGĻAG?J–GLRCJJGECLAC–/BMLOQ—LK—QEB—MR?IF@—AB?>QB—IBA—?V—QEB—$OBK@E—A>Q>—MOLQB@QFLK—>RQELOFQV
!,'*—>P—M>OQ—LC—QEB—BQEF@>I—AFP@RPPFLK—>PPFDKJBKQ—PBQ—?V—QEB—AFDFQ>I—OBMR?IF@—?FII—/BQOFBSBA—COLJ https://www.cnil.fr/sites/
– BCD?SJR ĺJCQ ?RMKQ ĺJCQ ALGJ=P?NNMPR=?G=E@=UC@NBD
 127 References

——!LRK@FI—LC—#ROLMB—!L#—–JEMPGRFKQ–?LB–&SK?L–0GEFRQ 0QRAV—LK—QEB—ERJ>K—OFDEQP—AFJBKPFLKP—LC—>RQLJ>QBA—A>Q>

—— MOL@BPPFKD—QB@EKFNRBP—>KA—MLPPF?IB—OBDRI>QLOV—FJMIF@>QFLKP—/BQOFBSBA—COLJ—https://rm.coe.int/study-hr-dimension-
of-automated-data-processing-incl-algorithms/168075b94a
—!LRK@FI—LC—#ROLMB—!L#——2CAFLMJMEGA?J–AMLTCPECLAC
–?PRGĻAG?J–GLRCJJGECLAC–?LB–FSK?L–PGEFRQ/BQOFBSBA—COLJ
http://www.assembly.coe.int/nw/xml/XRef/Xref-DocDetails-EN.asp?FileID=24236&lang=EN
—!O>DIF>
—+
—&O>AB@
—(—>KA—1OLRPP>OA
—5——| FD—">Q>
—'
—>KA—ÇR?IF@—ÇLIF@V}—FK—0R@E>
—3
—+>FO
—"
—0FBKHFBTFQW
—+—
— BQ—>I—#AP——1AGCLAC–?LB–#TGBCLAC–GL–RFC–.MJGAW–#AMQWQRCK–&?LB@MMI
—#IPBSFBO—CLOQE@LJFKD
——!O>TCLOA
—+
—)ELPEDLCQ>>O
—1—+
—ÇORP>
—"—(—BQ—>I——|0ROSBV—LC—OBSFBT—PM>J—ABQB@QFLK—RPFKD—J>@EFKB—IB>OKFKD—
— QB@EKFNRBP}–(MSPL?J–MD– GE–"?R?—0MOFKDBO—,>QROB——ALF P
——">K>EBO
—(——JEMAP?AW–?Q–&WNCPLSBEGLE––,CU–5?W–RM–3LBCPQR?LB–RFC–2FPC?R–MD–JEMAP?AW—/BQOFBSBA—COLJ
https://ieet.org/index.php/IEET2/more/Danaher20170117
">OS>P
—7
—>KA—4LIȴ
—%— ——AL–L?RMKW–MD–'LAJSQGTC–%PMURF–GL–#SPMNC– ORBDBI— IRBMOFKQ—
—-@QL?BO—
">RQE
—4
—$FKABFPBK
—0
—0RBABHRJ
—(
—>KA—4LBPPKBO
—,——BHSQRGLE–RM–0M@MRQ–5MPICP*CTCJ–#TGBCLAC–$BABO>I—/BPBOSB
— >KH—LC—+FKKB>MLIFP
—4LOHFKD—Ç>MBO—KL——https://www.minneapolisfed.org/institute/institute-working-papers/
adjusting-to-robots-worker-level-evidence
——"B—%OLBSB
—1
—)>OJBK
—Ç
—>KA—3BOK>@@FKF
—*——'LBCV–DMP–0GQI–+?L?ECKCLR––'LDM0+–!MLACNR–?LB–+CRFMBMJMEW
–
3BOPFLK——#ROLMB>K—!LJJFPPFLK
—(LFKQ—/BPB>O@E—!BKQOB
"B—ÇO>QL
—%
—*LMBW—!L?L
—+—BQ—>I——2FC–'–2CAFLM#AMLMKGA–1CEKCLR–L?JWQGQ–1CJCARCB–'LBGA?RMPQ–#ROLMB>K—
— !LJJFPPFLK
—(LFKQ—/BPB>O@E—!BKQOB
—0BSFIIB
—0M>FK
"B>OK
—(—Ç>QQBOPLK
—"—>KA—6LRKD
—!——|—,BT—%LIABK—DB—FK—!LJMRQBO—O@EFQB@QROB—#JMLTBOFKD—QEB—+>@EFKB*B>OKFKD—
— /BSLIRQFLK}–'###–+GAPM–—
"BI>DB
—$——|!EFK>—AFMILJ>@F>—B@LKŃJF@>
—@LKPB@RBK@F>P—DBLMLIľQF@>P}—!S?BCPLMQ–BC–#QRP?RCEG?
–
——
"BICLODB
—
—%BO>OA
—*
—|,LQOB—SFB—MOFSĺB—BPQBIIB—OĺBIIBJBKQ—JFPB—BK—A>KDBO—M>O—IBP—OL?LQP——ğQRAB—ABP—OFPNRBP—BQ—>K>IVPB
— ABP—PLIRQFLKP—>MMLOQĺBP—M>O—IB—%"Ç/}
—FK—*|GLRCJJGECLAC–?PRGĻAGCJJC–CR–JC–BPMGR
–*>O@FBO
— ORPPBIP——
"BILFQQB——1RSBW–ML–CKCPEGLE–GQQSCQ–MD–B?R?–MULCPQFGN
–GLRCPMNCP?@GJGRW
–PCSQ?@GJGRW–?LB–?AACQQ–RM–B?R?
–?LB–JG?@GJGRW

MOBM>OBA—CLO—QEB—#ROLMB>K—!LJJFPPFLK
—>S>FI>?IB—>Q—http://ec.europa.eu/newsroom/dae/document.cfm?doc_id=51486
——"BJFKD
—"—(——|1EB—%OLTFKD—'JMLOQ>K@B—LC—0L@F>I—0HFIIP—FK—QEB—*>?LO—+>OHBQ}—/S?PRCPJW–(MSPL?J–MD–#AMLMKGAQ
–
KL—(>KR>OV— https://doi.org/10.1093/qje/qjx022
"FU!>OKBFOL
—/——|1O>AB—*F?BO>IFW>QFLK—>KA—*>?LO—+>OHBQ—"VK>JF@P}—#AMLMKCRPGA?
–
—
——"OBUI
—(
—>—|"BPFDKFKD—!LJMBQFQFSB—+>OHBQP—CLO—'KARPQOF>I—">Q>}x— CRUCCL–.PMNCPRGQ?RGML–?LB–AACQQ
—
—('Ç'1#!
—_
— 
—
— "OBUI
—(
—|-K—QEB—$RQROB—#2—*BD>I—$O>JBTLOH—CLO—QEB—"FDFQ>I—#@LKLJV——@LJMBQFQFLK?>PBA—/BPMLKPB—QL—QEB—-TKBOPEFM
@@BPP—"B?>QB}
—FK—*LEPPB
—0
—0@ERIWB
—/—>KA—0Q>RABKJ>VBO
—"
—BAP
—2P?BGLE–"?R?–GL–RFC–"GEGR?J–#AMLMKW–*CE?J–!MLACNRQ
and Tools
—,LJP
— >ABK >ABK
—
—M
——"OFD>P
——>KA—'L>KKFALR
—/—#——|0MB@F>I—#AR@>QFLK—>KA—'!1P} 'LRCPL?RGML?J–(MSPL?J–MD–#KCPEGLE–2CAFLMJMEGCQ–
– GL–*C?PLGLE–'(#2———x
——"R@E OLTK
—,
—+>OQBKP
— 
—+RIIBO*>KDBO
—$
——2FC–CAMLMKGAQ–MD–MULCPQFGN
–?AACQQ–?LB–RP?BC–GL–BGEGR?J–B?R?

— (/!—1B@EKF@>I—OBMLOQ
—(/!—"FDFQ>I—#@LKLJV—4LOHFKD—Ç>MBO—
—#ROLMB>K—!LJJFPPFLK
—(LFKQ—/BPB>O@E—!BKQOB
—
— 0BSFIIB
—0M>FK
"RQQLK
—1——'—PQO>QBDFBP https://medium.com/politics-ai/an-overview-of-national-ai-strategies-2a70ec6edfd
#AMLMKGQR
——|1EB—"B>QE—LC—"FPQ>K@B
—PROSBV—LC—1BIB@LJJRKF@>QFLKP}
#AMLMKGQR
—0BMQBJ?BO
—SLI——KL—
——#AT>OAP
—*—>KA—3B>IB
—+——|0I>SB—QL—QEB—IDLOFQEJ—4EV—>—|/FDEQ—QL—>K—#UMI>K>QFLK}—FP—ÇOL?>?IV—KLQ—QEB—/BJBAV—
— 6LR—>OB—*LLHFKD—$LO}—"SIC–*?U––2CAFLMJMEW–0CTGCU
——x—
——#ROLMB>K—!LJJFPPFLK—#!
——"FOB@QFSB—  #!—LC—QEB—#ROLMB>K—Ç>OIF>JBKQ—>KA—LC—QEB—!LRK@FI—LC——+>O@E——
— LK—QEB—IBD>I—MOLQB@QFLK—LC—A>Q>?>PBP. https://eur-lex.europa.eu/lexUriServ/LexUriServ.do?uri=CELEX:31996L0009:EN:HTML
——#ROLMB>K—!LJJFPPFLK—#!
——"FOB@QFSB—  #!—LC—QEB—#ROLMB>K—Ç>OIF>JBKQ—>KA—LC—QEB—!LRK@FI—LK—QEB—OBRPB
 References 128

— LC—MR?IF@—PB@QLO—FKCLOJ>QFLK—LC—  
—OBSFPBA—?V—"FOB@QFSB—  #!—LC—   https://ec.europa.eu/
digital-single-market/en/european-legislation-reuse-public-sector-information
——#ROLMB>K—!LJJFPPFLK—#!
—>—/BDRI>QFLK—#2— —LC—QEB—#ROLMB>K—Ç>OIF>JBKQ—>KA—LC—QEB—!LRK@FI—
— LC——MOFI——LK—QEB—MOLQB@QFLK—LC—K>QRO>I—MBOPLKP—TFQE—OBD>OA—QL—QEB—MOL@BPPFKD—LC—MBOPLK>I—A>Q>—>KA—
— LK—QEB—COBB—JLSBJBKQ—LC—PR@E—A>Q>
—>KA—OBMB>IFKD—"FOB@QFSB—  #!—%BKBO>I—">Q>—ÇOLQB@QFLK—/BDRI>QFLK
—-(—*—
——#ROLMB>K—!LJJFPPFLK—#!
—?
—ÇOLMLP>I—CLO—>—"FOB@QFSB—LC—QEB—#ROLMB>K—Ç>OIF>JBKQ—>KA—LC—QEB—!LRK@FI—LK—@LMVOFDEQ—
— FK—QEB—"FDFQ>I—0FKDIB—+>OHBQ
—!-+—ȵK>I
— ORPPBIP
——#ROLMB>K—!LJJFPPFLK—#!
—@—"FOB@QFSB—#2— —LC—QEB—#ROLMB>K—Ç>OIF>JBKQ—>KA—LC—QEB—!LRK@FI—
— LC——(RKB——LK—QEB—MOLQB@QFLK—LC—RKAFP@ILPBA—HKLTELT—>KA—?RPFKBPP—FKCLOJ>QFLK—QO>AB—PB@OBQP—>D>FKPQ—QEBFO—
— RKI>TCRI—>@NRFPFQFLK
—RPB—>KA—AFP@ILPROB
—-(—*——
——#ROLMB>K—!LJJFPPFLK—#!
—>
—!LJJRKF@>QFLK—COLJ—QEB—!LJJFPPFLK—QL—QEB—#ROLMB>K—Ç>OIF>JBKQ
—QEB—!LRK@FI
—
— QEB—#ROLMB>K—#@LKLJF@—>KA—0L@F>I—!LJJFQQBB—>KA—QEB—!LJJFQQBB—LC—QEB—/BDFLKP—| RFIAFKD——#ROLMB>K—">Q>—#@LKLJV}—
— !-+——ȵK>I— ORPPBIP
——#ROLMB>K—!LJJFPPFLK—#!
—?—(LFKQ—!LJJRKF@>QFLK—QL—QEB—#ROLMB>K—Ç>OIF>JBKQ—>KA—QEB—!LRK@FI—/BPFIFBK@B
—
— "BQBOOBK@B—>KA—"BCBK@B— RFIAFKD—PQOLKD—@V?BOPB@ROFQV—CLO—QEB—#2—1B@E—OBM
—#2. https://eur-lex.europa.eu/legal-content/
EN/TXT/PDF/?uri=CELEX:52017JC0450&from=ES
FRRNQ CACSPMN?CS BGEGR?JQGLEJCK?PICR CL LCUQ QR?ĶUMPIGLEBMASKCLR?LBCVCASRGTCQSKK?PWCT?JS?RGML
directive-969ec-legal-protection
——#ROLMB>K—!LJJFPPFLK—#!
—>—!LJJRKF@>QFLK—COLJ—QEB—!LJJFPPFLK—QL—QEB—#ROLMB>K—Ç>OIF>JBKQ
—QEB—!LRK@FI
—
— QEB—#ROLMB>K—#@LKLJF@—>KA—0L@F>I—!LJJFQQBB—>KA—QEB—!LJJFQQBB—LC—QEB—/BDFLKP
—OQFȵ@F>I—'KQBIIFDBK@B—CLO—#ROLMB
— !-+——ȵK>I— ORPPBIP
——#ROLMB>K—!LJJFPPFLK—#!
—?—'—$>@QPEBBQ FRRNQ CACSPMN?CS BGEGR?JQGLEJCK?PICR CL LCUQ D?ARQFCCR?PRGĺAG?J
intelligence-europe
——#ROLMB>K—!LJJFPPFLK—#!
—@—!LJJRKF@>QFLK—COLJ—QEB—!LJJFPPFLK—QL—QEB—#ROLMB>K—Ç>OIF>JBKQ
—QEB—!LRK@FI
—
— QEB—#ROLMB>K—#@LKLJF@—>KA—0L@F>I—!LJJFQQBB—>KA—QEB—!LJJFQQBB—LC—QEB—/BDFLKP
—|1LT>OAP—>—@LJJLK—#ROLMB>K—A>Q>
— PM>@B}—!-+——ȵK>I— ORPPBIP
——#ROLMB>K—!LJJFPPFLK—#!
—A—ÇOLMLP>I—CLO—>—"FOB@QFSB—LC—QEB—#ROLMB>K—Ç>OIF>JBKQ—>KA—LC—QEB—!LRK@FI—
— LK—QEB—OBRPB—LC—MR?IF@—PB@QLO—FKCLOJ>QFLK—OB@>PQ—!-+  —ȵK>I
— ORPPBIP
——#ROLMB>K—!LJJFPPFLK—#!
—C—!LJJFPPFLK—0Q>ȴ—4LOHFKD—"L@RJBKQ—%RFA>K@B—LK—PE>OFKD—MOFS>QB—PB@QLO—A>Q>—
— FK—QEB—#ROLMB>K—A>Q>—B@LKLJV—@@LJM>KVFKD—QEB—AL@RJBKQ—!LJJRKF@>QFLK—COLJ—QEB—!LJJFPPFLK—QL—QEB—#ROLMB>K
— Ç>OIF>JBKQ
—QEB—!LRK@FI
—QEB—#ROLMB>K—B@LKLJF@—>KA—PL@F>I—!LJJFQQBB—>KA—QEB—!LJJFQQBB—LC—QEB—/BDFLKP—|1LT>OAP—
— >—@LJJLK—#ROLMB>K—A>Q>—PM>@B}
—04"—ȵK>I
— ORPPBIP
——#ROLMB>K—!LJJFPPFLK—#!
—D—!LJJFPPFLK—0Q>ȴ—4LOHFKD—"L@RJBKQ—#S>IR>QFLK—LC—"FOB@QFSB—  #!—LK—QEB—IBD>I—
— MOLQB@QFLK—LC—A>Q>?>PBP
—04"——ȵK>I
— ORPPBIP—
#ROLMB>K—!LJJFPPFLK—#!—C—!LRK@FI—OB@LJJBKA>QFLK—LC——+>V——LK—HBV—@LJMBQBK@BP—CLO—IFCBILKD—IB>OKFKD

— —-(—!—
—
—M—x—https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX%3A32018H0604%2801%29
——#ROLMB>K—!LJJFPPFLK—#!
—F—!LJJFPPFLK—/B@LJJBKA>QFLK—#2— —LC——MOFI——LK—>@@BPP—
— QL—>KA—MOBPBOS>QFLK—LC—P@FBKQFȵ@—FKCLOJ>QFLK
—-ȷ@F>I—(LROK>I—LC—QEB—#ROLMB>K—2KFLK
—*— 
——#ROLMB>K—">Q>—ÇOLQB@QFLK—0RMBOSFPLO—#"Ç0— 2MU?PBQ–?–BGEGR?J–CRFGAQ
–0CNMPR–DPMK–#".1–#RFGAQ–BTGQMPW–%PMSN–
– 0CRPGCTCB–DPMK FRRNQ CBNQCSPMN?CS QGRCQ CBN ĺJCQ NS@JGA?RGML =C?E=PCNMPR=CLNBD
——#ROLMB>K—#@LKLJF@—>KA—0L@F>I—!LJJFQQBB—##0!——##0!—-MFKFLK—LK—OQFȵ@F>I—'KQBIIFDBK@B https://www.eesc.europa.
– CS CL MSPUMPI MNGLGMLQGLDMPK?RGMLPCNMPRQ MNGLGMLQ ?PRGĺAG?JGLRCJJGECLAC
——#ROLMB>K—%OLRM—LK—#QEF@P—#%#——|0Q>QBJBKQ—LK—OQFȵ@F>I—'KQBIIFDBK@B
—/L?LQF@P—>KA—|RQLKLJLRP}—0VPQBJP—
— #ROLMB>K—%OLRM—LK—#QEF@P—FK—0@FBK@B—>KA—,BT—1B@EKLILDFBP}— ORPPBIP—#ROLMB>K—!LJJFPPFLK https://ec.europa.eu/
research/ege/pdf/ege_ai_statement_2018.pdf
—#ROLMB>K—Ç>OIF>JBKQ—#Ç——#ROLMB>K—!FSFI—*>T—/RIBP—FK—/L?LQF@P. http://www.europarl.europa.eu/RegData/etudes/
STUD/2016/571379/IPOL_STU(2016)571379_EN.pdf
 129 References

——#ROLMB>K—Ç>OIF>JBKQ—#Ç——/BPLIRQFLK—LC——$B?OR>OV——TFQE—OB@LJJBKA>QFLKP—QL—QEB—!LJJFPPFLK—LK—!FSFI—
— *>T—/RIBP—LK—/L?LQF@P— ',*—
——$>IF>DH>
—#
—BQ—>I——|MMIF@>QFLK—LC—J>@EFKB—IB>OKFKD—>IDLOFQEJP—QL—>K—LKIFKB—OB@ORFQJBKQ—PVPQBJ}—'K—Çroceedings.
– 'LRCPL?RGML?J–!MLDCPCLAC–ML–'LRCPLCR–?LB–5C@–NNJGA?RGMLQ–?LB–1CPTGACQ—
——$B>OLK
—(——4E>Q—FP—FABKQFQV—>P—TB—KLT—RPB—QEB—TLOA—/BQOFBSBA—COLJ—https://web.stanford.edu/group/fearon-
research/cgi-bin/wordpress/wp-content/uploads/2013/10/What-is-Identity-as-we-now-use-the-word-.pdf
——$BBK?BOD
——–!PGRGA?J–2FCMPW–MD–2CAFLMJMEW–,BT—6LOH—-UCLOA—2KFSBOPFQV—ÇOBPP—
——$BBK?BOD
———/SCQRGMLGLE–2CAFLMJMEW—,BT—6LOH—/LRQIBADB—
$BFGLL
—!—BQ—>I——2FC–'LBSQRPG?J–'LLMT?RGML–#AMQWQRCK–MD–PRGĻAG?J–'LRCJJGECLAC–GL–!FGL?–!SPPCLR–QR?RSQ–?LB–NPMQNCARQ

#ROLMB>K—!LJJFPPFLK
—(LFKQ—/BPB>O@E—!BKQOB
—0BSFIIB
—0M>FK—CLOQE@LJFKD
——$BOKijKABW+>@ľ>P
—#——|(L?—MLI>OFW>QFLK—FK—#ROLMB—!E>KDBP—FK—QEB—BJMILVJBKQ—PQOR@QROB—>KA—GL?—NR>IFQV}
—
— —5MPI–?LB–-AASN?RGMLQ
–
—
$BOKijKABW+>@ľ>P
—#
—>KA—&ROIBV
—(——/LRQFKB?F>PBA—QB@EKF@>I—@E>KDB—>KA—GL?—MLI>OFW>QFLK—FK—#ROLMB—1MAGM#AMLMKGA
– 0CTGCU
–
—
——$BOKijKABW+>@ľ>P
—#
—>KA— FPBIIL
—+——+C?QSPGLE–2FC–!MLRCLR–?LB–+CRFMBQ–MD–5MPI––!MKNPCFCLQGTC–2?QI–
– $P?KCUMPI–http://ibs.org.pl/app/uploads/2017/10/D1.-Martina-Bisello.pdf
——$BOKijKABW+>@ľ>P
—#
—%ŃJBW
—#
—&BOKijKABW-O>IIL
—(
—*LB
— —0
—+>OQBKP
— 
—+>OQľKBWÇIRJBA
—$
—>KA—1LI>K
—0——
— |—JRIQFAFP@FMIFK>OV—Q>PH?>PBA—MBOPMB@QFSB—CLO—BS>IR>QFKD—QEB—FJM>@Q—LC—'—>RQLKLJV—>KA—DBKBO>IFQV—LK—QEB—CRQROB—
— LC—TLOH}—?P6GT–NPCNPGLR–?P6GT
——$FKKFPE—%LSBOKJBKQ—
—ÇR?IF@>QFLKP—LC—QEB—+FKFPQOV—LC—#@LKLJF@—ȴ>FOP—>KA—#JMILVJBKQ—+FKFPQOV
— 
—
$GLJ?LB|Q–EC–MD–PRGĻAG?J–'LRCJJGECLAC–2SPLGLE–$GLJ?LB–GLRM–?–JC?BGLE–AMSLRPW–GL–RFC–?NNJGA?RGML–MD–?PRGĻAG?J–GLRCJJGECLAC–
– -@HCARGTC–?LB–PCAMKKCLB?RGMLQ–DMP–KC?QSPCQ–FRRN HSJI?GQSRT?JRGMLCSTMQRMĺ @GRQRPC?K F?LBJC  
TEMrap_47_2017_verkkojulkaisu.pdf?sequence=1&isAllowed=y
$ILOFAF
—*—BA——2FC–-LJGDC–+?LGDCQRM– CGLE–&SK?L–GL–?–&WNCPAMLLCARCB–#P?—*LKALK—0MOFKDBO—-MBK
——$ILOFAF
—*—–1MDR–CRFGAQ–?LB–RFC–EMTCPL?LAC–MD–RFC–BGEGR?J ÇEFILPLMEV—>KA—1B@EKLILDV
——. https://doi.org/10.1007/
s13347-018-0303-9
$P?LAC–'LRCJJGECLAC–PRGĻAGCJJC–$'––0?NNMPR–BC–QWLRFƁQC FRRNQ UUUCAMLMKGCEMSTDP ĺJCQ ĺJCQ ."$ 
Rapport_synthese_France_IA_.pdf
$P?LAC–'LRCJJGECLAC–PRGĻAGCJJC–$'––0?NNMPR–BC–QWLRFƁQC–%PMSNC–BC–RP?T?GJ FRRNQ UUUCAMLMKGCEMSTDP ĺJCQ ĺJCQ
PDF/2017/Conclusions_Groupes_Travail_France_IA.pdf
$OBV
—!— 
—>KA—-P?LOKB
—+—— 2FC–$SRSPC–MD–#KNJMWKCLR–5MPIGLE–.?NCP–-VDMPB–-VDMPB–+?PRGL–.PMEP?KKC–
– ML–2CAFLMJMEW–?LB–#KNJMWKCLR–https://www.oxfordmartin.ox.ac.uk/downloads/academic/future-of-employment.pdf
——$OBV
—!— 
—>KA—-P?LOKB
—+———|1EB—$RQROB—LC—#JMILVJBKQ—&LT—0RP@BMQF?IB—OB—(L?P—QL—!LJMRQBOFP>QFLK}
– 2CAFLMJMEGA?J–$MPCA?QRGLE–?LB–1MAG?J–!F?LEC––. https://doi.org/10.1016/j.techfore.2016.08.019
——$ROJ>K
—(
—>KA—0B>J>KP
—/——'–?LB–RFC–#AMLMKW–, #/—4LOHFKD—Ç>MBO—KL—
——%>OȵKHBI
—0
—+>QQEBTP
—(
—0E>MFOL
—0—0
—>KA—0JFQE
—(—+——|1LT>OA—>IDLOFQEJF@—QO>KPM>OBK@V—>KA—>@@LRKQ>?FIFQV}—
!MKKSLGA?RGMLQ–MD–RFC–!+
—
— https://doi.org/10.1145/3125780
%BFDBO
—!
—$OLPFL
—%—>KA— RI>VBKHL
—-
——|!O>CQFKD—>—QBUQ—>KA—A>Q>—JFKFKD—BU@BMQFLK—CLO—J>@EFKB—IB>OKFKD—>KA—?FD—A>Q>
— FK—QEB—"FDFQ>I—0FKDIB—+>OHBQ}—FK—0BR?>
—5
—%BFDBO
—!—>KA—ÇBKFK
—(
—#AP— 'LRCJJCARS?J–.PMNCPRW–?LB–"GEGR?J–2P?BC–
– GL–RFC–EC–MD–PRGĻAG?J–'LRCJJGECLAC–?LB– GE–"?R?
—M
—>S>FI>?IB—>Q—https://www.ictsd.org/themes/innovation-and-ip/
– PCQC?PAF GLRCJJCARS?JNPMNCPRW?LBBGEGR?JRP?BCGLRFC?ECMD?PRGĺAG?J
%FIBP
—+
——|'Q}P—QFJB—QL—OBFK—FK—QEB—A>Q>—?>OLKP}–+'2–2CAFLMJMEW–0CTGCU https://www.technologyreview.com/s/611425/
its-time-to-rein-in-the-data-barons/
——%LIAC>O?
—
—1OBȶBO
—"
—>—|&LT—>OQFȵ@F>I—FKQBIIFDBK@B—FJM>@QP—FKQBOK>QFLK>I—QO>AB}—5MPJB–2P?BC–PCNMPR–
–
The future of world trade: how digital technologies are transforming global commerce
—M——
——%LIAC>O?
—
—>KA—1OBȶBO
—"—?—|'—>KA—'KQBOK>QFLK>I—1O>AB}—FK—DO>T>I
—
—%>KP
—(—0
—>KA—%LIAC>O?
——#AP——
— 1EB—B@LKLJF@P—LC—>OQFȵ@F>I—FKQBIIFDBK@B—2KFSBOPFQV—LC—!EF@>DL—ÇOBPP
 References 130

——%ŃJBW
—#—BA——|PPBPPFKD—QEB—FJM>@Q—LC—J>@EFKB—FKQBIIFDBK@B—LK—ERJ>K—?BE>SFLRO—>K—FKQBOAFP@FMIFK>OV—BKAB>SLRO}
—
— PQ—&2+',1—TLOHPELM
— >O@BILK> FRRNQ CACSPMN?CS HPA AMKKSLGRGCQ ĺJC  BMULJM?BRMICL=0"2N'D
%ŃJBW
—#
— I>>RT
—+
— LK>A>
—(—BQ—>I——"CCN–*C?PLGLE–DMP–1GLEGLE–.PMACQQGLE–AFGCTCKCLRQ
–!F?JJCLECQ–?LB–'KN?AR–
– ML–1GLECPQ–?LB–*GQRCLCPQ https://arxiv.org/abs/1807.03046
%LLIP?BB
———.S@JGA–.MJGAW–GL–?L–'–#AMLMKW–,L—T—,>QFLK>I— ROB>R—LC—#@LKLJF@—/BPB>O@E
——%LLP
—+
—+>KKFKD
——>KA—0>ILJLKP
———|#UMI>FKFKD—GL?—MLI>OFW>QFLK—OLRQFKB?F>PBA—QB@EKLILDF@>I—@E>KDB—
— >KA—LȴPELOFKD}—KCPGA?L–#AMLMKGA–0CTGCU
——
——%O>BPPBO
——!
—Ç—!EFMJ>K
— —&>VKBP—!
—>KA—-IKBV
———|RQL1RQLO—K—'KQBIIFDBKQ—1RQLOFKD—0VPQBJ—TFQE—+FUBA
— 'KFQF>QFSB—"F>ILDRB}–'###–2P?LQ?ARGMLQ–ML–#BSA?RGML–– https://doi.org/10.1109/TE.2005.856149
——%OBBKMB>@B——How Clean is Your Cloud? https://www.greenpeace.org/archive-international/Global/international/
publications/climate/2012/iCoal/HowCleanisYourCloud.pdf
——%OBDLOV
—1
—0>ILJLKP
——>KA—7FBO>EK
—2——0?AGLE–UGRF–MP–?E?GLQR–RFC–K?AFGLC
–CTGBCLAC–DPMK–#SPMNC
—!#0FCL—TLOHFKD
— M>MBO—KL—
—0BMQBJ?BO—
——%R>A>JRW
———|"L—>KAOLFAP—AOB>J—LC—BIB@QOF@—@LMVOFDEQ—!LJM>O>QFSB—>K>IVPFP—LC—LOFDFK>IFQV—FK—>OQFȵ@F>I—FKQBIIFDBK@B
— DBKBO>QBA—TLOHP}
–'LRCJJCARS?J–.PMNCPRW–/S?PRCPJW
– https://ssrn.com/abstract=2981304
——%RFIIĺK
—+—>KA—/BAAV
—0——|4B—HKLT—BQEF@P—PELRIA—FKCLOJ—'— RQ—TEF@E—BQEF@P}–5MPJB–#AMLMKGA–$MPSK–LMRC
https://www.weforum.org/agenda/2018/07/we-know-ethics-should-inform-ai-but-which-ethics-robotics/
——&>II
—4
—>KA—ÇBPBKQF
—(—–%PMUGLE–RFC–PRGĻAG?J–'LRCJJGECLAC–'LBSQRPW–GL–RFC–3)–https://www.gov.uk/government/
– SNJM?BQ QWQRCK SNJM?BQ ?RR?AFKCLR=B?R? ĺJC  %PMUGLE=RFC=?PRGĺAG?J=GLRCJJGECLAC=GLBSQRPW=GL=RFC=3)NBD
&>O>OF
—6—,
— Homo Deus.—&>OSFI—0B@HBO
&FDDP
—#
—*FDEQ
——>KA—0QOLKD
—"—#AP—. 2CAFLMJMEW–?LB–RFC–%MMB–*GDC–!EF@>DL—1EB—2KFSBOPFQV—LC—!EF@>DL—ÇOBPP
——&LBHPQO>
—&—
—>KA—3>K—0IRGFP
—#——+?L?EGLE–!MKNCRCLACQ–'KNJCKCLRGLE–&SK?L–0CQMSPAC–+?L?ECKCLR—PPBK—
— )LKFKHIFGHB—3>K—%LO@RJ
——&LOKLH
—!
—>KA—)LOBK
—+——|4FKKBOP—>KA—ILPBOP—LC—DIL?>IFP>QFLK—0FUQBBK—@E>IIBKDBP—CLO—JB>PROBJBKQ—>KA—QEBLOV}
– #AMLMKGAQ–UGRFMSR–@MPBCPQ–CAMLMKGA–PCQC?PAF–DMP–#SPMNC?L–NMJGAW–AF?JJCLECQ–!>J?OFADB—2KFSBOPFQV—ÇOBPP
—!>J?OFADB
—
2)
——https://www.cambridge.org/core/books/economics-without-borders/winners-and-losers-of-globalization-six
teen-challenges-for-measurement-and-theory/E5BFE7843A887EB875AB8AC8CC179C5A
&R>KD
—*
—(LPBME
—
—,BIPLK
— —BQ—>I——|ASBOP>OF>I—+>@EFKB—*B>OKFKD}–'L–.PMACCBGLEQ–MD–th–!+–5MPIQFMN–ML–PRGĻAG?J–
– 'LRCJJGECLAC–?LB–1CASPGRW
—-@QL?BO—
—MM——https://people.eecs.berkeley.edu/~tygar/papers/SML2/Adversarial_
AISEC.pdf
——&R>KD
—0
—Ç>MBOKLQ
—,
—%LLACBIILT
—'—BQ—>I——BTCPQ?PG?J–RR?AIQ–ML–,CSP?J–,CRUMPI–.MJGAGCQ https://arxiv.org/
abs/1702.02284
——&RDBKELIQW
— ——|D>FKPQ—">Q>—ÇOLMBOQV}—FK—&>KKP
—2
—"O>ELP
—Ç—>KA—%EFAFKF
—%—#AP——)PGRGI?
–#QQ?WQ–
– ML–'LRCJJCARS?J–.PMNCPRW
–3LI—'''
—#ID>O—https://www.e-elgar.com/shop/kritika-essays-on-intellectual-property-15154
#>OIFBO—SBOPFLK—>Q https://www.ivir.nl/publicaties/download/Data_property_Muenster.pdf
——'###——#RFGA?JJW–JGELCB–"CQGEL—/BQOFBSBA—COLJ http://standards.ieee.org/develop/indconn/ec/autonomous
systems.html
——'###——1NCAG?J–'QQSC–&?AIGLE–UGRFMSR–&SK?LQ—'###—0B@ROFQV——ÇOFS>@V—https://publications.computer.org/
security-and-privacy/tag/hacking-without-humans/
'KPQFQRQB—CLO—#KBODV—#ȷ@FBK@V—'##——1A?J?@JC
–#LCPEW#ĸAGCLR–"?R?–!CLRCPQ–?LB–!JMSBQ https://iee.ucsb.edu/sites/
– GCCSAQ@CBS ĺJCQ B?R?=ACLRCP=PCNMPRNBD
——(>HL?PPLK
—(——|,V>—JFIGLKBO—PH>—Q>—ABO>P—AVPIBUFPQ>OQRM—QFII—20}—"F—"FDFQ>I
—+>O@E—
——https://digital.di.se/
artikel/nya-miljoner-ska-ta-deras-dyslexi-startup-till-usa—$LO—#KDIFPE—SBOPFLK
—PBB http://www.lexplore.com/
(>P>KLȴ
—0——1B@EKLILDFBP—LC—ERJFIFQV—!FQFWBK—M>OQF@FM>QFLK—FK—DLSBOKFKD—P@FBK@B—+GLCPT?
——
——(''Ç
—1,-——1B@EKLMLIFP——1RSBW–GL–QSNNMPR–MD–RFC–CT?JS?RGML–MD–"GPCARGTC–  #!–ML–RFC–JCE?J–NPMRCARGML–MD–B?R?
bases
—#ROLMB>K—!LJJFPPFLK https://publications.europa.eu/en/publication-detail/-/publication/f10f8ee0-597e-11e8-
ab41-01aa75ed71a1/language-en
 131 References

——)>EK
—(——|(RPQ—&LT—0E>IILT—'P—QEB—OQFȵ@F>I—'KQBIIFDBK@B—1>IBKQ—ÇLLI}– JMMK@CPE!MK
$B?OR>OV—
——https://www.
– @JMMK@CPEAMK LCUQ ?PRGAJCQ  HSQRFMUQF?JJMUGQRFC?PRGĺAG?JGLRCJJGECLACR?JCLRNMMJ
——)>J?>QI>
—)
—)LIIF>P
—%
—)RJ>O
—3
—%O>J>
———|1OBKAP—FK—?FD—A>Q>—>K>IVQF@P}
–(MSPL?J–MD–.?P?JJCJ–?LB–"GQRPG@SRCB–
– !MKNSRGLE
–——
——)>KF>
—#——|!EFK>—'P—-K—>—4ELIBLC,>QFLK—ÇRPE—CLO—'—1EB—20—+RPQ—+>Q@E—'Q}—!W@CP–5?PD?PC
— https://doi.
org/10.1016/B978-0-12-803843-7.00083-1
——)>O>?>O?LRKFP
—*—>KA—,BFJ>K
— ——2FC–%JM@?J–"CAJGLC–MD–RFC–*?@MP–1F?PC
—, #/—4LOHFKD—Ç>MBOP—
—,>QFLK>I—
— ROB>R—LC—#@LKLJF@—/BPB>O@E
—'K@—https://ideas.repec.org/p/nbr/nberwo/19136.html
——)>OID>>OA
—/—>KA—+>ILKB
—+——|!FQV—SP—@LRKQOV—1LJ—ÇBQBOP——%BLODB—%FIABO—AB?>QB—QEB—FJM>@Q—LC—QB@EKLILDV—
— LK—IL@>QFLK}—$MP@CQ–1.–2CAFLMJMEW–'QQSC
—$B?OR>OV
—
—MM—
——)BO?BO
—4— –LCU–GLRCJJCARS?J–NPMNCPRW–PGEFR–DMP–B?R?–?L–CAMLMKGA–?L?JWQGQ–%/2/——
——)BO?BO
—4——|%LSBOK>K@B—LC—">Q>—#U@IRPFSB—ÇOLMBOQV—SP—@@BPP}—FK–'LRCPL?RGML?J–0CTGCU–MD–'LRCJJCARS?J–.PMNCPRW–
– ?LB–!MKNCRGRGML–*?U
——
—
—>S>FI>?IB—>Q https://link.springer.com/article/10.1007/s40319-016-0517-2
)BO?BO
—4—>KA—$O>KH
—(0——"?R?–%MTCPL?LAC–PCEGKCQ–GL–RFC–"GEGR?J–#AMLMKW–RFC–#V?KNJC–MD–!MLLCARCB–!?PQ
FRRNQ UUURGJ@SPESLGTCPQGRWCBS SNJM?B AC@DACCD?CBAC=)CP@CP=$P?LI=
01102017_Data%20Governance%20and%20Connected%20Cars.pdf
——)FI?BOQRP
—,
—!>ORII>
—+—/
—Ç>O>P@>KALIL
—BQ—>I——TMGBGLE–"GQAPGKGL?RGML–RFPMSEF–!?SQ?J–0C?QMLGLE
—,'Ç0—
https://papers.nips.cc/paper/6668-avoiding-discrimination-through-causal-reasoning.pdf
——)KFDEQ
—4——|ÇOLDOBPP—FK—'—FPK}Q—>P—'JMOBPPFSB—>P—6LR—+FDEQ—1EFKH}
–+'2–2CAFLMJMEW–0CTGCU–
https://www.technologyreview.com/s/609611/progress-in-ai-isnt-as-impressive-as-you-might-think/
?utm_source=MIT+Technology+Review&utm_campaign=ab120fee85-weekly_roundup_2017-12-07_edit&utm
_medium=mail&utm_term=0_997ed6f472-ab120fee85-156491921&goal=0_997ed6f472-ab120fee85
-156491921&mc_cid=ab120fee85&mc_eid=dafbd41dd1
——)KFDEQ
—4——|'KPFAB—QEB—!EFKBPB—I>?—QE>Q—MI>KP—QL—OBTFOB—QEB—TLOIA—TFQE—'}–+'2–2CAFLMJMEW–0CTGCU–
https://www.technologyreview.com/s/610219/inside-the-chinese-lab-that-plans-to-rewire-the-world-with-ai/
——)LOȴ
—"—4>DKBO
— —>KA—ÇLTIBP
—(—BQ—>I—— MSLB?PGCQ–MD–*?U–#VNJMPGLE–2P?LQN?PCLAW
–AAMSLR?@GJGRW
–?LB–-TCPQGEFR–
– MD–%MTCPLKCLR–1SPTCGJJ?LAC–0CEGKCQ–2KFSBOPFQV—LC—!>J?OFADB—$>@RIQV—LC—*>T—/BPB>O@E—Ç>MBO—KL— —/BQOFBSBA
— COLJ—https://ssrn.com/abstract=2894490
——)LOFKBH
—
—>KA—0QFDIFQW
—(——PRGĻAG?J–'LRCJJGECLAC–?LB–'RQ–'KNJGA?RGMLQ–DMP–'LAMKC–"GQRPG@SRGML–?LB–3LCKNJMWKCLR
https://doi.org/10.3386/w24174.
——*BB
—)—$——|1EB—OB>I—QEOB>Q—LC—'}—,7–2GKCQ FRRNQ UUULWRGKCQAMK    MNGLGML QSLB?W ?PRGĺAG?J
-intelligence-economic-inequality.html
+>K@>
——/
— BK@WRO
—Ç—>KA—%FLS>KKFKF
—#—— SGJBGLE–?–1AGCLRGĻA–,?PP?RGTC–2MU?PBQ–?–+MPC–0CQGJGCLR–#3–QMAGCRW–
(/!—0@FBK@B—CLO—ÇLIF@V—/BMLOQ—http://publications.jrc.ec.europa.eu/repository/bitstream/JRC106265/jrc106265_100417
_resilience_scienceforpolicyreport.pdf
——+>OHE>J
—
—1FFABK?BOD
—)—>KA—&BOJ>K
———|#QEF@P—>P—+BQELAP—"LFKD—#QEF@P—FK—QEB—#O>—LC— FD—">Q>—/BPB>O@E—
— x—'KQOLAR@QFLK}—1MAG?J–+CBG?–?LB–1MAGCRW
—(RIV0BMQBJ?BO——
——+>ONR>OQ
—0——1MSRF–)MPC?L–%MTCPLKCLR–LLMSLACQ–,C?PJW–– GJJGML–GL–'–DSLBGLE—https://futurism.com/south
-korean-government-announces-nearly-1-billion-ai-funding/
——+>OQBKP
— ——L–CAMLMKGA–NMJGAW–NCPQNCARGTC–ML–MLJGLC–NJ?RDMPKQ (/!—"FDFQ>I—#@LKLJV—TLOHFKD—M>MBO——
— 0BSFIIB—(/!
——+>OQBKP
— —>KA—+RIIBO*>KDBO
—$——Access to car data and competition in aftersales services markets
—(/!—AFDFQ>I—
— —B@LKLJV—TLOHFKD—M>MBO
—#ROLMB>K—!LJJFPPFLK
—(LFKQ—/BPB>O@E—!BKQOB
—0BSFIIB
—0M>FK—CLOQE@LJFKD——
+>VBO
—+——'LLMT?RGML–?R–%MMEJC–2FC–NFWQGAQ–MD–B?R?
—>S>FI>?IB—>Q—FRRNQ FSĶBSĶCPAMK K?PQF?JJIGPIN?RPGAI 
——+@)FKPBV—%IL?>I—'KPQFQRQB—+%'——,MRCQ–DPMK–RFC–'–DPMLRGCP–+MBCJGLE–RFC–GKN?AR–MD–'–GL–RFC–UMPJB–CAMLMKW–
FRRNQ UUUKAIGLQCWAMK DC?RSPCBGLQGEFRQ ?PRGĺAG?JGLRCJJGECLAC LMRCQDPMKRFC?GDPMLRGCPKMBCJGLERFCGKN?AR
-of-ai-on-the-world-economy
 References 132

——+BFA>K
—6
— LE>A>K>
—+
—+>QELS
—6—BQ—>I——|, >'L1—,BQTLOH?>PBA—"BQB@QFLK—LC—'L1— LQKBQ—QQ>@HP—2PFKD—"BBM—
— RQLBK@LABOP}–'###–.CPT?QGTC–!MKNSRGLE
–(SJW 1CN––
——+BWW>KLQQB
—$——|@@BPP—QL—">Q>—QEB—/LIB—LC—!LKPBKQ—>KA—QEB—*F@BKPFKD—0@EBJB}
—FK—*LEPPB
—0
—0@ERIWB
—
— /—>KA—0Q>RABKJ>VBO
—"—#AP——2P?BGLE–"?R?–GL–RFC–"GEGR?J–#AMLMKW–*CE?J–!MLACNRQ–?LB–2MMJQ
—,-+-0—M
+F@E>RU
— ——|0FKDRI>OFQĺ—QB@EKLILDFNRB
—PFKDRI>OFQĺ—ERJ>FKB—BQ—AOLFQ—A}>RQBRO}
—FK—"B—1BOT>KDKB
—!
—"BDO>SB
—
— #—>KA—"RPLIIFBO
—0—#AP——*?UQ
–,MPKQ–?LB–$PCCBMKQ–GL–!W@CPQN?AC– –"PMGRQ
–LMPKQ–CR–JG@CPRƂQ–B?LQ–JC–AW@CPKMLBC

— *F?BO—JF@LORJ—6SBP—ÇLRIIBQ
—*>O@FBO
—M
+FIIBO
—!
—-EOSFH0QLQQ
—(
—!LIAF@RQQ—/
——0CESJ?RGLE–DMP–0CQNMLQG@JC–2CAFLMJMEW–!?N?AGRW
–#TGBCLAC–?LB–0CBPCQQ–
– ?–LCU–QWQRCK–DMP–?–D?GPCP–DSRSPC *LKALK—"LQBSBOVLKB https://doteveryone.org.uk/project/regulating-for
-responsible-technology
+FPPFLK—3FII>KF—PRO—I}FKQBIIFDBK@B—>OQFȵ@FBIIB——$MP–?–KC?LGLEDSJ–?PRGĻAG?J–GLRCJJGECLAC–2MU?PBQ–?–$PCLAF–?LB–#SPMNC?L
– 1RP?RCEW https://www.aiforhumanity.fr/pdfs/MissionVillani_Report_ENG-VF.pdf
——+FQ@E>J
—!——|ÇEFILPLMEV—LC—1B@EKLILDV}—'K—"RO?FK
—Ç1—BA——–%SGBC–RM–RFC–!SJRSPC–MD–1AGCLAC
–2CAFLMJMEW–
– ?LB–+CBGAGLC—1EB—$OBB—ÇOBPP
——+LKQOB>I—"B@I>O>QFLK——$MP–?–0CQNMLQG@JC–"CTCJMNKCLR–MD–PRGĻAG?J–'LRCJJGECLAC–/BQOFBSBA—COLJ
https://www.montrealdeclaration-responsibleai.com/the-declaration
——+LOFH>T>
—+——QQCQQGLE–RFC–GKN?AR–MD–'–?LB–PM@MRGAQ–ML–HM@–CVNCAR?RGMLQ–SQGLE–(?N?LCQC–QSPTCW–B?R?
https://voxeu.org/article/who-fears-losing-their-job-ai-and-robots-japanese-survey-data
——+R?>VF
—Ç
—!EBKD
—#
—1BOOV
—&—Ç
—1FIQLK
—
—>KA—&LR
—1——!FGL?|Q–0GQC–GL–PRGĻAG?J–'LRCJJGECLAC—%LIAJ>K—0>@EP—/BQOFBSBA
— COLJ—https://www.cnbc.com/2017/09/01/goldman-says-china-has-talent-data-and-infrastructure-to-embrace-ai.html
,>P@FJBKQL
—0
—ÇŃISLO>
—
—Ç>FL
—
—-IFSBFO>
—0—BQ—>I——|0RPQ>FK>?IB—1B@EKLILDFBP—>KA—1O>KPAFP@FMIFK>OV—$RQROBP—$OLJ—
— !LII>?LO>QFSB—"BPFDK—QL—"FDFQ>I—$>?OF@>QFLK}
1AGCLAC–?Q–!SJRSPC
3LI——,L—
—MM——
——,>QFLK>I—0@FBK@B—>KA—1B@EKLILDV—!LRK@FI—,01!—@—,?RGML?J–PRGĻAG?J–GLRCJJGECLAC–PCQC?PAF–?LB–BCTCJMNKCLR–
strategic plan.–FRRNQ M@?K?UFGRCFMSQC?PAFGTCQEMT QGRCQ BCD?SJR ĺJCQ UFGRCFMSQC=ĺJCQ KGAPMQGRCQ MQRN ,12!
national_ai_rd_strategic_plan.pdf
——,>QFLK>I—0@FBK@B—>KA—1B@EKLILDV—!LRK@FI—,01!—? PRGĻAG?J–GLRCJJGECLAC
–?SRMK?RGML–?LB–RFC–CAMLMKW
FRRNQ UUUUFGRCFMSQCEMT QGRCQ UFGRCFMSQCEMT ĺJCQ GK?ECQ #+ 0%-#"͕'͕#AMLMKW͕0CNMPRNBD–
——,>QFLK>I—0@FBK@B—>KA—1B@EKLILDV—!LRK@FI
—,01!—>—.PCN?PGLE–DMP–RFC–DSRSPC–MD–?PRGĻAG?J–GLRCJJGECLAC
—#UB@RQFSB—-ȷ@B
— LC—QEB—ÇOBPFABKQ
—>–FRRNQ GLDMNS@JGAGLRCJJGECLACLCR 5FGRC&MSQCPRGĺAG?J'LRCJJGECLAC.PCN?P?RGMLQNBD–
——,BABIHLPH>
—*—>KA—.RFKQFKF
—%——SRMK?RGML
–1IGJJQ–3QC–?LB–2P?GLGLE–-#!"—0L@F>I
—#JMILVJBKQ—>KA—+FDO>QFLK—
— 4LOHFKD—Ç>MBOP—Ç>OFP—-#!"
——,H>J?LR
—/
—WBSBAL
—/—>KA—3>PPFIBS>
—(—#AP——'LRCJJGECLR–2SRMPGLE–1WQRCKQ–th–'LRCPL?RGML?J–!MLDCPCLAC
–
– '21–
–+MLRPC?J
–/!
–!?L?B?
–(SLC–̨
–
–.PMACCBGLEQ–ÇOLDO>JJFKD—>KA—0LCQT>OB—#KDFKBBOFKD—0MOFKDBO—
— 'KQBOK>QFLK>I—ÇR?IFPEFKD
——-},BFI
—!——5C?NMLQ–MD–K?RF–BCQRPSARGML–&MU–@GE–B?R?–GLAPC?QCQ–GLCOS?JGRW–?LB–RFPC?RCLQ–BCKMAP?AW–,BT—6LOH
—— !OLTK—ÇR?IFPEBOP
——-#!"——(M@–!PC?RGML–?LB–*MA?J–#AMLMKGA–"CTCJMNKCLR–.PCN?PGLE–DMP–RFC–$SRSPC–MD–5MPI–-#!"—ÇR?IFPEFKD
—Ç>OFP

https://doi.org/10.1787/9789264305342-en
-ȷ@B—M>OIBJBKQ>FOB—A}ĺS>IR>QFLK—ABP—@ELFU—P@FBKQFȵNRBP—BQ—QB@EKLILDFNRBP—-Ç#!01— 2MU?PB–?–!MLRPMJJCB
–3QCDSJ
– ?LB–"CKWQRGĻCB–PRGĻAG?J–'LRCJJGECLAC http://www.senat.fr/rap/r16-464-1/r16-464-1-syn-en.pdf
-MEBO
——
—!ELR
—
—-KA>
——BQ—>I——2FC–0GQC–MD–RFC–"?R?–#AMLMKW–"PGTGLE–4?JSC–RFPMSEF–'LRCPLCR–MD–2FGLEQ–"?R?
+LKBQFW>QFLK—' +—!LOMLO>QFLK FRRNQ UUUG@KAMK AMKKML QQG AEG@GL QQG?JG?QFRKJĺB55531#,
-P>?>
—-—>KA—4BIPBO
—4——2FC–0GQIQ–MD–PRGĻAG?J–'LRCJJGECLAC–RM–1CASPGRW–?LB–RFC–$SRSPC–MD–5MPI—/>KA—!LOMLO>QFLK
https://www.rand.org/pubs/perspectives/PE237.html
——-P?LOKB—!I>OHB—**Ǘ—*CE?J–QRSBW–ML–-ULCPQFGN–?LB–AACQQ–RM–"?R?
#ROLMB>K—!LJJFPPFLK https://publications.
europa.eu/en/publication-detail/-/publication/d0bec895-b603-11e6-9e3c-01aa75ed71a1/language-en
 133 References

——Ç>MBOKLQ
—,
—+@">KFBI
—Ç
—(E>
—0—BQ—>I——1EB—*FJFQ>QFLKP—LC—"BBM—*B>OKFKD—FK—ASBOP>OF>I—0BQQFKDP—.PMA–'###–
– #SPMNC?L–1WKN–1CASPGRW–?LB–.PGT?AW–#SPM1–.——ALF #ROL0Ç
——Ç>R
—(
— >GBO
—(—>KA—&LRPQLK
—,——PRGĻAG?J–GLRCJJGECLAC–GL–QG?–.PCN?PCBLCQQ–?LB–0CQGJGCLAC
–PF>— RPFKBPP—!LRK@FI
http://www.asiabusinesscouncil.org/Research17_AI.html
——Ç>ULK
—3——| OL——0VPQBJ—CLO—"BQB@QFKD—,BQTLOH—'KQORABOP—FK—/B>I1FJB}—!MKNSRCP–,CRUMPIQ https://www.usenix.org/
legacy/publications/library/proceedings/sec98/full_papers/paxson/paxson.pdf
——ÇLO>VPH>ÇLJPQ>
—)——|'—FK—#AR@>QFLK—>P—>—+BQELALILDV—CLO—#K>?IFKD—#AR@>QFLK>I—#SFABK@B >PBA—ÇO>@QF@B}—
— 'K—1CTCLRCCLRF–'LRCPL?RGML?J–!MLDCPCLAC–ML–PRGĻAG?J–'LRCJJGECLAC–GL–#BSA?RGML–'#"–
x—+>AOFA
——ÇLRIIBQ
—6
—>—|'P—QEB—DBKBO>I—A>Q>—MOLQB@QFLK—OBDRI>QFLK—QEB—PLIRQFLK}
!MKNSRCP–*?U––1CASPGRW–0CTGCU–2FC–
– 'LRCPL?RGML?J–(MSPL?J–MD–2CAFLMJMEW–*?U–?LB–.P?ARGAC (CLOQE@LJFKD—
——ÇLRIIBQ
—6
—?}0E>OFKD—">Q>—0LJB—@OFQF@>I—OBȶB@QFLKP—>?LRQ—>—ÇLIF@V—>PPBOQBA—?V—QEB—#2—!LJJFPPFLK}—ÇOBPBKQ>QFLK
— ABIFSBOBA—>Q—QEB—!LKCBOBK@B—">Q>—PE>OFKD—>KA—OBRPB
—LOD>KFPBA—?V—!OFAP—>KA—!RIIBK—'KQBOK>QFLK>I—>KA—ELPQBA—FK— ORPPBIP

—— LK—OA—-@QL?BO——QL—?B—MR?IFPEBA—
——ÇLTIBP
—(—>KA—&LAPLK
—&——|%LLDIB—"BBM+FKA—>KA—EB>IQE@>OB—FK—>K—>DB—LC—>IDLOFQEJP}—&C?JRF–?LB–2CAFLMJMEW

— ——x
——ÇO>H>PE
—0
——2FC–2GKC–&?Q–!MKC–RM–3QC–PRGĻAG?J–'LRCJJGECLAC–'L–%MTCPL?LAC https://towardsdatascience.com/
– RFCRGKCF?QAMKCRMSQC?PRGĺAG?JGLRCJJGECLACGLEMTCPL?LACC@
——/>ACLOA
— —(
—MLILKFL
—*
—1OF>P
——(—>KA—0FJMPLK
—(———,CRUMPI–2P?ĸA–LMK?JW–"CRCARGML–3QGLE–0CASPPCLR–,CSP?J–
– ,CRUMPIQ https://arxiv.org/abs/1803.10769
——/>J>IEL
—
—>—|">Q>—ÇOLAR@BO}P—/FDEQ—ÇLTBO
—ÇBOFIP——ÇFQC>IIP}
— CRRCP–0CESJ?RGML–DMP–!MNWPGEFR–MOBPBKQ>QFLK
— —>Q—QEB—#ROLMB>K—Ç>OIF>JBKQ
 ORPPBIP—0BMQBJ?BO—
—
—M
—>S>FI>?IB—>Q—https://juliareda.eu/wp-content/
uploads/2017/09/2017-09-06_Better-Regulation-for-Copyright-Academics-meet-Policy-Makers_Proceedings.pdf
——/>J>IEL
——?—5GJJ–0M@MRQ–0SJC–RFC–PRGQRGA–5MPJB––.PMNMQCB–+MBCJ–DMP–RFC–*CE?J–1R?RSQ–MD–!PC?RGMLQ–@W–PRGĻAG?J
– 'LRCJJGECLAC–1WQRCKQ
—>S>FI>?IB—>Q—00/, http://dx.doi.org/10.2139/ssrn.2987757
——/>J>IEL
———.?RCLR?@GJGRW–MD–'ECLCP?RCB–GLTCLRGMLQ–GQ–?–PCDMPK–MD–RFC–N?RCLR–QWQRCK–LCCBCB–'KPQFQRQB—
— LC—'KQBIIB@QR>I—ÇOLMBOQV
—$LRKA>QFLK—CLO—'KQBIIB@QR>I—ÇOLMBOQV—LC—(>M>K
—M
—>S>FI>?IB—>Q https://ssrn.com/
abstract=3168703
——/>QEBK>R—'KPQFQRQB——Human rights in the robot age. !LRK@FI—LC—#ROLMB—/BMLOQ https://www.rathenau.nl/sites/default/
– ĺJCQ  &SK?L͕0GEFRQ͕GL͕RFC͕0M@MR͕EC0?RFCL?S͕'LQRGRSSRNBD
/FQQBO
—0
—KABOPLK
—(—/
—)LBAFKDBO
—)—/—>KA—!LO?BQQ
———|!LDKFQFSB—1RQLO—MMIFBA—/BPB>O@E—FK—+>QEBJ>QF@P—#AR@>QFLK}—
.QWAFMLMKGA– SJJCRGL––0CTGCU–——x—https://doi.org/10.3758/BF03194060
——/LAOFH
—"——|ÇLMRIFPJ—>KA—QEB—B@LKLJF@P—LC—DIL?>IFW>QFLK}–(MSPL?J–MD–'LRCPL?RGML?J– SQGLCQQ–.MJGAW
FRRNQ BPMBPGIQAFMJ?PF?PT?PBCBS ĺJCQ B?LGPMBPGI ĺJCQ NMNSJGQK=?LB=RFC=CAMLMKGAQ=MD=EJM@?JGX?RGMLNBD
/LPĺ
—!—Ç
—+>OQľKBW+>IALK>AL
—/
—&LMMB—2—BQ—>I—#AP——PRGĻAG?J–'LRCJJGECLAC–GL–#BSA?RGML–th–'LRCPL?RGML?J–
– !MLDCPCLAC
–'#"–
–*MLBML
–3)
–(SLC–̨
–
–.PMACCBGLEQ
–.?PR–'—*B@QROB—,LQBP—FK—OQFȵ@F>I—'KQBIIFDBK@B—
— 0MOFKDBO—'KQBOK>QFLK>I—ÇR?IFPEFKD
——/LV
—0—0
—0>JRF
—Ç
—"BL
—/—BQ—>I—#AP——| FD—">Q>—FK—#KDFKBBOFKD—MMIF@>QFLKP}
—1RSBGCQ–GL– GE–"?R? 3LI——0MOFKDBO—
— 0FKD>MLOB
——0>@EP
—(—"——/"
—|0QOR@QRO>I—1O>KPCLOJ>QFLK
—>KA—QEB—"FPQOF?RQFLK—LC—'K@LJB}—'K—DO>T>I
—
—%>KP
—(—0
—>KA—%LIAC>O?
—
— —#AP—–2FC–CAMLMKGAQ–MD–?PRGĻAG?J–GLRCJJGECLAC—+@)FKPBV—.R>OQBOIV
—MOFI http://conference.nber.org/
conferences/2017/AIf17/summary.html
——0>JRBIPLK
—Ç——|1EB—#2}P—!LKQOLSBOPF>I—"FDFQ>I—0FKDIB—+>OHBQ—"FOB@QFSB—x—Ç>OQ—''—4EV—QEB—ÇOLMLPBA—+>KA>QLOV
— 1BUQ—>KA—">Q>+FKFKD—#U@BMQFLK—'P—1LL—/BPQOF@QFSB}
)JSUCP–!MNWPGEFR– JME—http://copyrightblog.kluweriplawcom/
2018/07/12/eus-controversial-digital-single-market-directive-part-ii-proposed-mandatory-text-data-mining
-exception-restrictive/
——0>KLH
—"——|K—>K>IVPFP—LC—ELT—>KQFSFORP—JBQELALILDFBP—>OB—RQFIFPBA—FK—MOLQB@QFKD—@LJMRQBOP—COLJ—J>IF@FLRP—@LAB}—
.PMACCBGLEQ–MD–RFC–nd–?LLS?J–AMLDCPCLAC–ML–'LDMPK?RGML–1CASPGRW–ASPPGASJSK–BCTCJMNKCLR
 References 134

0@>PPBII>QF
— 
— L@@>KCRPL
—*
—&R>KD
—!—+—BQ—>I——|'JMOLSFKD—PL@F>I—PHFIIP—FK—@EFIAOBK—TFQE—0"—RPFKD—>—ILKDQBOJ
—
— FKELJB—PL@F>I—OL?LQ}—1AGCLAC–0M@MRGAQ
—
—B>>Q
——0@EņK?BODBO
—"——|"BBM—!LMVOFDEQ—2M——KA—"LTKPQOB>J—.RBPQFLKP—/BI>QBA—QL—OQFȵ@F>I—'KQBIIFDBK@B—'—>KA—+>@EFKB
— *B>OKFKD—+*}—FK—"B—4BOO>
—(—BA——"PMGR–B|?SRCSP–– –!MNWPGEFR–
–0@ERIQEBPP—#AFQFLKP—/LJ>KABP
—M
https://ssrn.com/abstract=3098315
——0BI?PQ
——"—>KA—ÇLTIBP
—(——|+B>KFKDCRI—FKCLOJ>QFLK—>KA—QEB—OFDEQ—QL—BUMI>K>QFLK}—'LRCPL?RGML?J–"?R?–.PGT?AW–*?U
–
——x
——0BJF@LKAR@QLO—'KARPQOV—PPL@F>QFLK—0'
——'LRCPL?RGML?J–RCAFLMJMEW–PM?BK?N–DMP–QCKGAMLBSARMPQ
https://www.semiconductors.org/resources/2015-international-technology-roadmap-for-semiconductors-itrs/
——0@E>CBO
— 
—)LJRSBP
—"
—7>Q>O>FK
—(—+—,—>KA—"FSBO
—*—}—CLROQE—I>T—LC—OL?LQF@P—!LMVOFDEQ—>KA—QEB—I>T—>KA—BQEF@P—
— LC—J>@EFKB—@LMOLAR@QFLK}
—OQFȵ@F>I—'KQBIIFDBK@B—>KA—*>T

——
——0ELOQ
—(—#
— LEK
—/—#
— >OR—!——&MU–KSAF–GLDMPK?RGML–̨PCNMPR–ML–CLRCPNPGQC–QCPTCP–GLDMPK?RGML—2!0"—%IL?>I
— 'KCLOJ>QFLK—'KARPQOV—!BKQBO
——0IBBJ>K
—"
—>KA— OLTK
—(—0——'LRCJJGECLR–2SRMPGLE–1WQRCKQ—,BT—6LOH—@>ABJF@—ÇOBPP
——0QBSBKPLK
— ——|'
—'K@LJB
—#JMILVJBKQ
—>KA—+B>KFKD}—FK—DO>T>I
—
—%>KP
—(—0
—>KA—%LIAC>O?
——#AP——
2FC–CAMLMKGAQ–MD–?PRGĻAG?J–GLRCJJGECLAC–2KFSBOPFQV—LC—!EF@>DL—ÇOBPP. http://www.nber.org/chapters/c14026
——0QFIDLB
—(——|+>@EFKB—IB>OKFKD
—PL@F>I—IB>OKFKD—>KA—QEB—DLSBOK>K@B—LC—PBICAOFSFKD—@>OP}—1MAG?J–1RSBGCQ–MD–1AGCLAC
—
— ——
0WBDBAV
—!
—7>OBJ?>
—4
—0RQPHBSBO
—'—BQ—>I——'LRPGESGLE–NPMNCPRGCQ–MD–LCSP?J–LCRUMPIQ https://arxiv.orgabs/1312.6199
1>K
—4
—$>K
—6
—%ELKBFJ
——BQ—>I——|$OLJ—QEB—0BOSF@B-OFBKQBA—O@EFQB@QROB—QL—QEB—4B?—Ç'—#@LKLJV}
—'###–'LRCPLCR–
Computing. (RIRD—
——1EB—$RQROB—0L@FBQV—–+?IGLE–RFC–'–PCTMJSRGML–UMPI–DMP–CTCPWMLC—/BQOFBSBA—COLJ https://www.tuftsgloballeadership.
– MPE QGRCQ BCD?SJR ĺJCQ GK?ECQ PCQMSPACQ +G?GJFC͕0C?BGLENBD
——1LIIBK>>O
—,
—>KA—3>K—ABO—&BFGABK
—Ç—%—+——|4EF@E—JBQELA—MOBAF@QP—OB@FAFSFPJ—?BPQ——@LJM>OFPLK—LC—PQ>QFPQF@>I

— —J>@EFKB—IB>OKFKD—>KA—A>Q>—JFKFKD—MOBAF@QFSB—JLABIP}—(MSPL?J–MD–RFC–0MW?J–1R?RGQRGA?J–1MAGCRW–1CPGCQ––1R?RGQRGAQ–
– GL–1MAGCRW——
——1RLJF
—'——2FC–'KN?AR–MD–'–ML–JC?PLGLE
–RC?AFGLE
–?LB–CBSA?RGML–.MJGAGCQ–DMP–RFC–DSRSPC !>?OBO>
—+
—3RLOFH>OF
—/—
— >KA—ÇRKFB
—6—#AP—(LFKQ—/BPB>O@E—!BKQOB—0@FBK@B—CLO—ÇLIF@V—/BMLOQ
—#ROLMB>K—!LJJFPPFLK
—(LFKQ—/BPB>O@E—!BKQOB
—0BSFIIB
—
— 0M>FK—http://publications.jrc.ec.europa.eu/repository/handle/JRC113226
2)—%LSBOKJBKQ—-ȷ@B—CLO—0@FBK@B
—>—0M@MRGAQ
–?SRMK?RGML–?LB–?PRGĻAG?J–GLRCJJGECLAC–https://assets.publishing.service.
– EMTSI EMTCPLKCLR SNJM?BQ QWQRCK SNJM?BQ ?RR?AFKCLR=B?R? ĺJC  0M@MRGAQ=?SRMK?RGML=?LB=?PRGĺAG?J=
intelligence_-_cst_letter.pdf
——2)—%LSBOKJBKQ—-ȷ@B—CLO—0@FBK@B
—?—PRGĻAG?J–GLRCJJGECLAC–MNNMPRSLGRGCQ–?LB–GKNJGA?RGMLQ–DMP–RFC–DSRSPC–MD–BCAGQGML
making. FRRNQ UUUEMTSI EMTCPLKCLR SNJM?BQ QWQRCK SNJM?BQ ?RR?AFKCLR=B?R? ĺJC  EQ?PRGĺAG?J
intelligence-ai-report.pdf
— 2)—%LSBOKJBKQ
—
'LBSQRPG?J–1RP?RCEW–PRGĻAG?J–'LRCJJGECLAC–1CARMP–"C?J
–.MJGAW–.?NCP
–NPGJ–
–
FRRNQ UUUEMTSI EMTCPLKCLR SNJM?BQ QWQRCK SNJM?BQ ?RR?AFKCLR=B?R? ĺJC  = #'1='=1CARMP=
Deal__4_.pdf
——2)—&LRPB—LC—!LJJLKP—0@FBK@B—>KA—1B@EKLILDV—!LJJFQQBB
——0M@MRGAQ–?LB–PRGĻAG?J–'LRCJJGECLAC–
https://www.publications.parliament.uk/pa/cm201617/cmselect/cmsctech/145/145.pdf
2)—&LRPB—LC—*LOAP
— '–GL–RFC–3)–PC?BW
–UGJJGLE–?LB–?@JC 0BIB@Q—!LJJFQQBB—LK—OQFȵ@F>I—'KQBIIFDBK@B—/BMLOQ—
— LC—0BPPFLK—
—&*—Ç>MBO—
——2KFSBOPFQV—LC—+BI?LROKB
—— GMKCRPGA–+GPPMP–FGEFJGEFRQ–ļ?UQ–GL–?PRGĻAG?J–GLRCJJGECLAC–https://www.eurekalert.org/pub
_releases/2018-07/uom-bmh072318.php
——3>K—P?OLB@H
— 
—"B?RPP@EB
—(—>KA—!BP>O
—(
——"?R?–MULCPQFGN–GL–RFC–AMLRCVR–MD–RFC–#SPMNC?L–B?R?–CAMLMKW–
– NPMNMQ?J–DMP–?–LCU–PGEFR
–4EFQB—Ç>MBO
— FOA—— FOA
 135 References

——3>K—ABK—&LSBK
—(——|'KQOLAR@QFLK}—'K—3>K—ABK—&LSBK
—(
—+FIIBO
—0—>KA—ÇLDDB
—1—#AP——"CQGELGLE–GL–#RFGAQ
–
– NN––!>J?OFADB—2KFSBOPFQV—ÇOBPP
——3>K—"FG@H
—(——|">Q>ȵ@>QFLK
—A>Q>FPJ—>KA—A>Q>SBFII>K@B— FD—A>Q>—?BQTBBK—P@FBKQFȵ@—M>O>AFDJ—>KA—FABLILDV}
1SPTCGJJ?LAC––1MAGCRW
————
——3BO?BBH
—Ç—Ç— +MP?JGXGLE–2CAFLMJMEW–3LBCPQR?LBGLE–?LB–"CQGELGLE–RFC–+MP?JGRW–MD–2FGLEQ !EF@>DL—>KA—*LKALK—
— 2KFSBOPFQV—LC—!EF@>DL—ÇOBPP
——3BPKF@IRGBSF@
—*
— OBFQBDDBO
—+—>KA—%RFJ>O>BP—ÇBOBFO>
———|"L'Q6LROPBIC}—&B>IQE@>OB—.R>IFQV—LC—&B>IQE—
— >KA—&B>IQE@>OB—1EOLRDE—4B>O>?IB—0BKPLOP}–1AGCLAC–?LB–#LEGLCCPGLE–#RFGAQ
———
——3FA>I
—(——|1PRK>JF—LC—A>Q>}—@LRIA—@LKPRJB—LKB—ȵCQE—LC—DIL?>I—BIB@QOF@FQV—?V—}—!IFJ>QB—ELJB—KBTP—
FRRN UUUAJGK?RCAF?LECLCUQAMK    RQSL?KGB?R?AMLQSKCMLCĺDRFEJM@?JCJCARPGAGRW
——4>@EQBO
—0
—+FQQBIPQ>AQ
— 
—$ILOFAF
—*——|4EV—>—/FDEQ—QL—#UMI>K>QFLK—LC—RQLJ>QBA—"B@FPFLK+>HFKD—"LBP—,LQ—#UFPQ—
— FK—QEB—%BKBO>I—">Q>—ÇOLQB@QFLK—/BDRI>QFLK}
'LRCPL?RGML?J–"?R?–.PGT?AW–*?U
4B?BO
—/
——|'JMOLSBJBKQ—LC—">Q>—#@LKLJV—QEOLRDE—!LJMRIPLOV—*F@BK@BP}—FK—*LEPPB
—0
—0@ERIWB
—/—>KA—0Q>RABKJ>VBO
—
— "—#AP—–2P?BGLE–"?R?–GL–RFC–"GEGR?J–#AMLMKW–*CE?J–!MLACNRQ–?LB–2MMJQ
–,MKMQ
—MM
——4LBQWBI
—(
—0BLKD
—(
—4>KD
—)—BQ—>I——"GEGR?J–!FGL?–.MUCPGLE–RFC–#AMLMKW–RM–%JM@?J–!MKNCRGRGTCLCQQ /BQOFBSBA—
— COLJ—https://www.mckinsey.com/featured-insights/china/digital-china-powering-the-economy-
to-global-competitiveness
——4LLIC
— —Ç—— SGJBGLE–'LRCJJGECLR–'LRCP?ARGTC–2SRMPQ–1RSBCLR!CLRCPCB–1RP?RCEGCQ–DMP–0CTMJSRGMLGXGLE–C*C?PLGLE
— 0>K—$O>K@FP@L
—!—+LOD>K—)>RCJ>KK
——5F
—(—–2FC–EMTCPL?LAC–MD–!FGL?–QR–CB— BFGFKD—$LOBFDK—*>KDR>DBP—ÇOBPP—
——6BRKD
—)——|1EB—$LOJP—>KA—*FJFQP—LC—!ELF@B—O@EFQB@QROB—>P—>—1LLI—LC—%LSBOKJBKQ}—*?U––.MJGAW
–——
——7B@E
—&——|—IBD>I—CO>JBTLOH—CLO—>—A>Q>—B@LKLJV—FK—QEB—#ROLMB>K—"FDFQ>I—0FKDIB—+>OHBQ—OFDEQP—QL—RPB—A>Q>}—
(MSPL?J–MD–'LRCJJCARS?J–.PMNCPRW–*?U––.P?ARGAC
——
 *GQRȩMDȩ MVCQ 136

LIST OF BOXES
MVȩȩȩ #?PJWȩQSAACQQȩA?QCȩMDȩK?AFGLCȩJC?PLGLEȩGLȩAW@CPQCASPGRWȩQN?KȩȏJRCPGLEȩȩ 19
MVȩȩȩ Turning point in image recognition 20
MVȩȩȩ DL for singing processing 21
MVȩȩȩ 1MKCȩCV?KNJCQȩMDȩ'ȩBCTCJMNKCLRQȩGLȩRFCȩ31ȩȩ 26
MVȩȩȩ The Techno-Economic Segment (TES) approach in a nutshell 30
MVȩȩȩ 1MKCȩCV?KNJCQȩMDȩ'ȩ?NNJGA?RGMLQȩGLȩ!FGL?ȩȩ 50
MVȩȩȩ The impact of AI on working conditions 82
MVȩȩȩ AI and digital labour platforms 83
MVȩȩȩ Summary data strategy from the Mission Villani Report (2018) 109
MVȩȩȩ#V?KNJCQȩMDȩ.'QȩGLȩJMA?JȩEMTCPLKCLRȩȩ 110
 137 List of Figures

LIST OF FIGURES
Figure 1. Paradigm shift in machine learning 20
Figure 2. Percentage of AI players detected by the most representative
keywords (2000-18) 30
Figure 3. AI players in the world by geographical zones (% over world total),
2009-2018 31
Figure 4.ȩȩ 2MNȩȩ'ȩNJ?WCPQȩGLȩRFCȩUMPJBȩ?LBȩPCJ?RGMLȩUGRFȩ%".ȩ@WȩECMEP?NFGA?JȩXMLCQ
ȩ
2009-2018 32
Figure 5.ȩȩ 'ȩNJ?WCPQȩGLȩRFCȩ#3ȩ?LBȩPCJ?RGMLȩUGRFȩ%".ȩ@WȩAMSLRPW
ȩȩ 32
Figure 6. R&D versus non-R&D players in top 10 countries by number of AI players,
2009-2018 33
Figure 7. Number of AI players in selected types of activity (% over world total),
2009-2018 33
Figure 8.ȩȩ "GQRPG@SRGMLȩMDȩ#3ȩNJ?WCPQȩϤȩGLȩ#3DSLBCBȩ'ȩPCQC?PAFȩNPMHCARQ
ȩȩ 34
Figure 9.ȩȩ -AASPPCLACȩMDȩ'ȩRMNGAQȩϤȩ@WȩECMEP?NFGA?JȩXMLC
ȩȩ 35
Figure 10.ȩȩ 'ȩRFCK?RGAȩNPMȏJCQȩGLȩ?A?BCKGAȩNPMEP?KKCQȩ?QȩϤȩMDȩRMR?JȩNPMEP?KKCQȩ 74
Figure 11.ȩȩ "GQRPG@SRGMLȩMDȩSLGTCPQGRGCQȩMȎCPGLEȩ'ȩQRSBWȩRMNGAQȩMTCPȩFSK?LȩPCQMSPACQȩ
in science and technology 2017 75
Figure 12. Paradigm shift in the cybersecurity of systems following the introduction
of AI components. 89
Figure 13.ȩȩ .PMHCARGMLȩMDȩB?R?ȩ?LBȩAMKNSRGLEȩEPMURFQȩJME?PGRFKGAȩQA?JCȩ 96
Figure 14.ȩȩ "?R?ȏA?RGMLȩNPMACQQȩCL?@JGLEȩRFCȩ'ȩPCTMJSRGMLȩ 112
Figure 15. Employment in manufacturing and healthcare: EU and USA 117
Figure 16. AI preparedness in EU Member States 117
Figure 17. Human resources in science and technology and Digital Innovation Hubs 118
Figure 18. Human resources in science and technology, Digital Innovation Hubs,
and academic supply in study topics related to AI 119
!),-5*#"%#+#,21
2FCȩ K?GLȩ AMLRPG@SRMPQȩ RMȩ RFGQȩ PCNMPRȩ UCPCȩ JCQQ?LBPMȩ LLMLG
ȩ .CRCPȩ CLAXSP
ȩ .?MJMȩ CPRMJBG
ȩ +?Vȩ !P?-
EJG?
ȩ J?EMHȩ"CJGNCRPCT
ȩ%GSBGRR?ȩ"Cȩ.P?RM
ȩ#LPGOSCȩ$CPL?LBCXȩ+?AG?Q
ȩ#KGJG?ȩ%MKCX
ȩ+?PG?ȩ'EJCQG?Q
ȩ&CLPGIȩ
(SLIJCUGRX
ȩ+MLRQCPP?Rȩ*ŁNCXȩ!M@M
ȩ CPRGLȩ+?PRCLQ
ȩ1SQ?L?ȩ,?QAGKCLRM
ȩ1RCD?LMȩ,?RGTG
ȩJCV?LBPCȩ.MJTMP?
ȩ
'EL?AGMȩ1?LAFCX
ȩ1MLESJȩ2MJ?L
ȩ*SAG?ȩ4CQLGAȩJSHCTGAȩDPMKȩRFCȩ(0!
ȩ?QȩUCJJȩ?Qȩ!J?SBGMȩ$CGHŁMȩ3LGTCPQGB?Bȩ
Politécnica de Madrid), and Ilkka Tuomi (Meaning Processing).

-RFCPȩ AMLRPG@SRMPQȩ UCPCȩ K?LB?ȩ LBCP@CPE

ȩ +?PACJGLMȩ !?@PCP?ȩ %GP?JBCX
ȩ $P?LACQA?ȩ !?KNMJMLEM
ȩ
Melisande Cardona, Vicky Charisi, Nestor Duch-Brown, Serena Fatica, Elisabeta Florescu, Jiri Hradec, Ma-
AGCHȩ)PWQXRMDMUGAX
ȩ+?PGSQȩ+GPML
ȩJ@CPRMȩ+MPM
ȩLBPC?ȩ.?E?LM
ȩLBPC?ȩ.CPCEM
ȩ7TCQȩ.SLGC
ȩ0GAA?PBMȩ0GEFG
ȩ
1Mȏ?ȩ1?KMGJG
ȩ$?@G?L?ȩ1A?NMJM
ȩ*MPCLXGLMȩ4?AA?PG
ȩ*SI?XQȩ8CK@?ȩDPMKȩRFCȩ(0!ȩ?LBȩ(C?L.?SJȩ1GKMLȩ(.1ȩ
Public Policy Consulting).

We are particularly grateful to the International Editorial Board members for their comments and support:

• Dr. Bob Bishop, President and Founder of the International Centre for Earth Simulation Foundation.
• .PMDȩ4GPEGLG?ȩ"GELSK
ȩ.PMDCQQMPȩ1MAG?Jȩ?LBȩ#RFGA?JȩPRGȏAG?Jȩ'LRCJJGECLAC
ȩ3KCĵȩ3LGTCPQGRW
• Dr. Cosmina Dorobantu, Deputy Director of the Public Policy Programme and Policy Fellow at the Alan
Turing Institute for Data Science and AI.
• "Pȩ,SPG?ȩ-JGTCP
ȩ"GPCARMPȩMDȩ0CQC?PAFȩGLȩ"?R?ȩ1AGCLACȩ?Rȩ4MB?DMLCȩ?LBȩ!FGCDȩ"?R?ȩ1AGCLRGQRȩGLȩ"?R?.MNȩ
Alliance.
• .PMDȩ ?PPWȩ-1SJJGT?L
ȩ"GPCARMP
ȩ'LQGEFRȩ!CLRPCȩDMPȩ"?R?ȩL?JWRGAQ
ȩ3LGTCPQGRWȩ!MJJCECȩ!MPI
• Prof. Yves Poullet, Professor at the Faculty of Law University of Namur, and Co-Chair of the Namur
Digital Institute.
• "PȩLBPC?ȩ0CLB?
ȩ1CLGMPȩ0CQC?PAFȩ$CJJMUȩ?LBȩ&C?BȩMDȩRFCȩ%JM@?Jȩ%MTCPL?LAC
ȩ0CESJ?RGML
ȩ'LLMT?RGMLȩ
and the Digital Economy at the Centre for European Policy Studies.
GETTING IN TOUCH WITH THE EU

IN PERSON
JJȩMTCPȩRFCȩ#SPMNC?Lȩ3LGMLȩRFCPCȩ?PCȩFSLBPCBQȩMDȩ#SPMNCȩ"GPCARȩGLDMPK?RGMLȩACLRPCQȩ7MSȩA?LȩȏLBȩ
the address of the centre nearest you at: https://europa.eu/european-union/contact_en

ON THE PHONE OR BY EMAIL

Europe Direct is a service that answers your questions about the European Union. You can contact
this service:
- by freephone: 00 800 6 7 8 9 10 11 (certain operators may charge for these calls),
- at the following standard number: +32 22999696, or
- by electronic mail via: https://europa.eu/european-union/contact_en

FINDING INFORMATION ABOUT THE EU

ONLINE
'LDMPK?RGMLȩ?@MSRȩRFCȩ#SPMNC?Lȩ3LGMLȩGLȩ?JJȩRFCȩMȑAG?JȩJ?LES?ECQȩMDȩRFCȩ#3ȩGQȩ?T?GJ?@JCȩ
on the Europa website at: https://europa.eu/european-union/index_en

EU PUBLICATIONS
You can download or order free and priced EU publications from EU Bookshop at: https://publications.
europa.eu/en/publications. Multiple copies of free publications may be obtained by contacting Europe
Direct or your local information centre (see https://europa.eu/european-union/contact_en).

EU LAW AND RELATED DOCUMENTS

For access to legal information from the EU, including all EU law since 1952
GLȩ?JJȩRFCȩMȑAG?JȩJ?LES?ECȩTCPQGMLQ
ȩEMȩRMȩ#30*CVȩ?Rȩhttp://eur-lex.europa.eu

OPEN DATA FROM THE EU

The EU Open Data Portal (http://data.europa.eu/euodp/en) provides access to datasets from the EU.
Data can be downloaded and reused for free, for both commercial and non-commercial purposes.
 KJ-NA-29425-EN-C
The European Commission’s
science and knowledge service
Joint Research Centre

JRC Mission
As the science and knowledge service
of the European Commission, the Joint
Research Centre’s mission is to support
EU policies with independent evidence
throughout the whole policy cycle.

EU Science Hub
ec.europa.eu/jrc

@EU_ScienceHub
EU Science Hub - Joint Research Centre
Joint Research Centre
EU Science Hub