Vous êtes sur la page 1sur 14

Are We Living in a Computer Simulation?

Author(s): Nick Bostrom


Reviewed work(s):
Source: The Philosophical Quarterly, Vol. 53, No. 211 (Apr., 2003), pp. 243-255
Published by: Wiley-Blackwell for The Philosophical Quarterly
Stable URL: http://www.jstor.org/stable/3542867 .
Accessed: 04/09/2012 20:35

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp

.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact support@jstor.org.

Wiley-Blackwell and The Philosophical Quarterly are collaborating with JSTOR to digitize, preserve and
extend access to The Philosophical Quarterly.

http://www.jstor.org
ThePhilosophical Vol.53,No.2
Quarterly, April
2003
ISSNoo3-8g94

ARE WE LIVING IN A COMPUTER SIMULATION?

ByNICK BOSTROM

I arguethatat leastoneofthe
following (i) thehuman
is true:
propositions is very
species to
likely
become extinctbefore a 'posthuman'
reaching stage;(2) anyposthuman is extremely
civilization
toruna sgnficant
unlikely number ofsimulations
ofitsevolutionary (orvariations
history thereof);
(3) wearealmost certainly ina computer
living Itfollows
simulation. thatthere
thatthebelief is a
chance
signficant thatweshallonedaybecome posthumans whorunancestor-simulationsisfalse,
livingina simulation.
unlesswearecurrently I discusssomeconsequences
ofthisresult.

I. INTRODUCTION

Many worksof sciencefictionas well as some forecastsby serioustechno-


logistsand futurologistspredictthatenormousamountsofcomputingpower
will be available in the future.Let us suppose for a momentthat these
predictionsare correct.One thingthatlatergenerationsmightdo withtheir
super-powerful computersis rundetailedsimulations oftheirforebearsor of
people like theirforebears.Because theircomputerswould be so powerful,
theycould runa greatmanysuch simulations. Suppose thatthesesimulated
people are conscious(as theywould be if the simulationswere sufficiently
fine-grained and ifa certainquitewidelyacceptedpositionin thephilosophy
of mindis correct).Then itcould be thecase thatthevastmajorityofminds
likeoursdo notbelongto theoriginalrace but ratherto people simulatedby
the advanced descendantsof an originalrace. It is thenpossibleto argue
thatifthiswerethecase, we wouldbe rationalto thinkthatwe are likelyto
be among the simulatedmindsratherthan among the originalbiological
ones. Thereforeifwe do notthinkthatwe are currently livingin a computer
simulation,we are not entitledto believe thatwe shall have descendants
who will run lots of simulationsof theirforebears.That is the basic idea.
The restofthispaperwillspellitout morecarefully.
Apartfromtheinterest thisthesismayholdforthoseengagedin futuristic
speculation, thereare also morepurelytheoretical
rewards.The argumentis
a stimulusforformulating some methodological
and metaphysical questions,
? The Editorsof ThePhilosophical 2003. Publishedby BlackwellPublishing,
Quarterly, Road, Oxfordox4 2DQUK,
9600 Garsington
and 350 Main Street,Maiden,MA02148,USA.
244 NICK BOSTROM

and it suggestsnaturalisticanalogiesof certaintraditionalreligiousconcep-


tions,whichsomemayfindamusingor thought-provoking.
The structureof thepaper is as follows.First,I formulatean assumption
which I need to importfromthe philosophyof mind in order to get the
argumentstarted.Secondly,I considersome empiricalreasonsforthinking
thatrunningvastlymanysimulationsof humanmindswould be withinthe
capabilityof a futurecivilizationthathas developedmanyof thosetechno-
logiesthatcan alreadybe shownto be compatiblewithknownphysicallaws
and engineering constraints.This partis notphilosophicallynecessary,butit
provides an incentiveforpaying attentionto the Then
rest. followsthe core
of the argument,whichmakesuse of some simpleprobability and
theory, a
sectionprovidingsupportfora weak indifference principlethe argument
employs.Lastly,I discusssome interpretations of thedisjunctionmentioned
in theabstract,whichformstheconclusionofthesimulationargument.

II. THE ASSUMPTION OF SUBSTRATE-INDEPENDENCE

A common assumptionin the philosophyof mind is that of substrate-


The idea is thatmentalstatescan superveneon any ofa broad
independence.
class of physicalsubstrates.Provideda systemimplementsthe rightsortof
computationalstructures and processes,it can be associatedwithconscious
experiences.It is not an essentialpropertyof consciousnessthatit is imple-
mented on carbon-basedbiological neural networksinside a cranium:
silicon-basedprocessorsin a computercould in principledo thetricktoo.
Arguments forthisthesishave been givenin theliterature, and although
itis notentirelyuncontroversial, I shallheretakeitas given.
The argumentI shall presentdoes not, however,depend on any very
strongversionof functionalism or computationalism. For example,I need
not assumethatthethesisof substrate-independence is necessarily
true(either
or -
analytically metaphysically)merely that a computerrunninga suitable
program would in fact be conscious.Moreover, I need not assume thatin
orderto createa mindon a computeritwouldbe necessaryto programitin
such a way thatit behaveslike a human in all situations, includingpassing
theTuringtest,etc. I need onlytheweakerassumptionthatit would suffice
forthegenerationofsubjectiveexperiencesthatthecomputational processes
of a human brain are structurally replicatedin suitablyfine-grained detail,
such as on the level of individualsynapses.This attenuatedversion of
substrate-independence is quitewidelyaccepted.
Neurotransmitters, nerve growthfactorsand other chemicalsthat are
smallerthana synapseclearlyplaya role in humancognitionand learning.
? The EditorsofThePhilosophical
Qjuarterly,
200oo3
ARE WE LIVING IN A COMPUTER SIMULATION? 245

The substrate-independence thesisis not thatthe effectsof thesechemicals


but
are smallor irrelevant, rather that they affectsubjectiveexperienceonly
via theirdirector indirectinfluenceon computationalactivities.For ex-
ample,iftherecan be no difference in subjectiveexperiencewithoutthere
also being a differencein synapticdischarges,then the requisitedetail of
simulationis at thesynapticlevel(or higher).

III. THE TECHNOLOGICAL LIMITS OF COMPUTATION

At our currentstage of technologicaldevelopment,we have neither


sufficientlypowerfulhardwarenor therequisitesoftware to createconscious
mindsin computers.But persuasivearguments have been givento theeffect
thatif technologicalprogresscontinuesunabated,thenthesetechnological
shortcomings will eventuallybe overcome.Some authorsargue that this
stage may be only a fewdecades away.' Yet presentpurposesrequireno
assumptionsabout the time-scale.The simulationargumentworksequally
well forthosewho thinkthatit willtake hundredsof thousandsofyearsto
reach a 'posthuman'stage of civilization,where humankindhas acquired
mostofthetechnologicalcapabilitiesthatone can currently showto be con-
sistentwithphysicallaws and withmaterialand energyconstraints.
Such a maturestageoftechnologicaldevelopment willmakeitpossibleto
convertplanetsand otherastronomicalresourcesintoenormously powerful
computers.It is currently hard to be confidentin any upperbound on the
computingpower thatmay be available to posthumancivilizations. As we
are stilllackinga 'theoryof everything',we cannot rule out the possibility
thatnovelphysicalphenomena,not allowedforin currentphysicaltheories,
may be utilizedto transcendthose constraints thatin our currentunder-
standingimpose theoreticallimitson the informationprocessingattainable
in a givenlump of matter.2 We can withmuch greaterconfidenceestablish
lowerbounds on posthumancomputation,by assumingonly mechanisms
that are already understood.For example, Eric Drexler has outlineda
theComing
ofCreation:
1 See, e.g., K.E. Drexler,Engines Era ofNanotechnology
(London: Fourth
Estate,1985);N. Bostrom,'How Long BeforeSuperintelligence?', International
JournalofFutures
Studies,2 (1998); R. Kurzweil, The Age of Spiritual
Machines:WhenComputers ExceedHuman
Intelligence(New York: Viking,1999); H. Moravec, Robot:MereMachineto Transcendent Mind
(OxfordUP, 1999).
2 I.e., constraintssuch as the Bremermann-Bekenstein bound and the black hole limit:
HJ. Bremermann,'Minimum Energy Requirementsof InformationTransferand Com-
puting',International ofTheoretical
Journal 21 (1982),pp. 203-17;J.D. Bekenstein,'Entropy
Physics,
Contentand Information Flow in SystemswithLimitedEnergy',Physical Review,D 30 (1984),
pp. 1669-79; A. Sandberg,'The Physicsof InformationProcessingSuperobjects:the Daily
LifeamongtheJupiterBrains',Journal ofEvolution
andTechnology,5 (1999)-
C The EditorsofThePhilosophical 2003
Quarterly,
246 NICK BOSTROM

designfora systemthe size of a sugar cube (excludingcoolingand power


supply)thatwould performIo21instructions per second.3Anotherauthor
givesa roughestimateof I042operationsper second fora computerwitha
massoftheorderofa largeplanet.4(Ifwe could createquantumcomputers,
or learnto buildcomputersout ofnuclearmatteror plasma,we could push
closerto the theoreticallimits.Seth Lloyd calculatesan upper bound fora
I kgcomputerof5 x 1050 logicaloperationsper secondcarriedout on -I031
bits.5However, it sufficesfor my purposesto use the more conservative
estimatethatpresupposesonlycurrently knowndesignprinciples.)
The amountof computingpowerneeded to emulatea humanmindcan
likewisebe roughlyestimated.One estimate,based on how computationally
expensiveitis to replicatethefunctionality
ofa piece ofnervoustissuewhich
we have alreadyunderstoodand whose functionality has been replicatedin
silice,namely,contrastenhancementin the retina,yieldsa figureof - IO14
operationsper secondforthe entirehumanbrain.6An alternative estimate,
based on the numberof synapsesin the brain and theirfiringfrequency,
givesa figureof-IO16-IO17 operationsper second.7Conceivably,evenmore
could be requiredifwe want to simulatein detailthe internalworkingsof
synapsesand dendritictrees.However,it is likelythat the human central
nervoussystemhas a highdegreeof redundancyon themicroscaleto com-
pensateforthe unreliability and noisinessof itsneuronalcomponents.One
would thereforeexpect a substantialefficiencygain when using more
reliableand versatilenon-biologicalprocessors.
Memory seems to be no more stringenta constraintthan processing
power.8Moreover,since the maximumhuman sensorybandwidthis -Io8
bitsper second,simulatingall sensoryeventsincursa negligiblecost com-
pared to simulating thecorticalactivity.
We can therefore use theprocessing
powerrequired to simulatethe centralnervous system an estimateofthe
as
totalcomputationalcostofsimulating a humanmind.
If the environmentis included in the simulation,this will require
additionalcomputingpower- how much,dependson thescope and granu-
larityofthesimulation.Simulatingtheentireuniversedownto thequantum
levelis obviouslyinfeasible,
unlessradicallynewphysicsis discovered.But in
orderto geta realisticsimulationofhumanexperience,muchless is needed
- onlywhateveris requiredto ensurethatthesimulatedhumans,interacting
3 K.E. Drexler,Nanosystems(New York:JohnWiley& Sons, 1992).
4 RJ. Bradbury,'MatrioshkaBrains',working (2002), http://www.aeiveos.com/
manuscript
-bradbury/MatrioshkaBrains/MatrioshkaBrains.html.
5 S. Lloyd, 'Ultimate Physical Limits to Computation',Nature,406 (31 August 2000),
pp. 1047-54.
6 H. Moravec,MindChildren (HarvardUP, 1989).
7 See my'How Long beforeSuperintelligence?'.
8 See referencesin foregoingfootnotes.

? The Editorsof ThePhilosophical 2003


Quarterly,
ARE WE LIVING IN A COMPUTER SIMULATION? 247

in normalhumanwayswiththeirsimulatedenvironment, do notnoticeany
irregularities.The microscopicstructure of the insideof the Earth can be
safelyomitted.Distant astronomicalobjects can have highlycompressed
representations: verisimilitude need onlyextendto thenarrowband ofpro-
pertiesthatwe can observefromour planetor solar systemspacecraft.On
thesurfaceof Earth,macroscopicobjectsin inhabitedareas mayneed to be
continuouslysimulated,but microscopicphenomena could probablybe
filledin ad hoc.What you see throughan electronmicroscopeneeds to look
unsuspicious, but you usuallyhave no way of confirming itscoherencewith
unobservedpartsofthemicroscopicworld.Exceptionsarisewhenwe delib-
eratelydesignsystemsto harnessunobservedmicroscopicphenomenathat
operatein accordancewithknownprinciplesto get resultswe are able to
verifyindependently. The paradigmcase of thisis a computer.The simula-
tionmay therefore need to includecontinuousrepresentation of computers
down to the level of individuallogic elements.This presentsno problem,
sinceour currentcomputingpoweris negligiblebyposthumanstandards.
Moreover,a posthumansimulatorwould have enoughcomputingpower
to keep trackof the detailedbelief-states in all human brainsat all times.
Therefore, when it saw that a human was about to make an observationof
themicroscopicworld,itcould fillin sufficient detailin thesimulationin the
appropriate domain as and where needed. Should any erroroccur, the
directorcould edit the statesof any brainsthathave become aware of an
anomalybeforethisspoilsthe simulation.Alternatively, the directorcould
skipback a fewsecondsand rerunthesimulationso as to avoid theproblem.
It thus seems plausible that the main computationalcost in creating
simulations thatare indistinguishable fromphysicalrealityforhumanminds
in the simulationresidesin simulating organicbrainsdown to the neuronal
or sub-neuronallevel. As we build more and fastercomputers,the cost of
simulating our machinesmighteventually come to dominatethecostofsim-
ulatingnervoussystems. While it is notpossibleto get a veryexact estimate
of thecostof a realisticsimulationof humanhistory, we can use
as a estimate.9 As we -oI33-IO36
operations rough gain more experiencewithvirtual
reality,we shall get a bettergrasp of the computationalrequirements for
making such worlds appear realistic to their visitors.
But in any case, even if
the estimateis inaccurateby several ordersof magnitude,this does not
mattermuch formy argument.I notedthata roughapproximationof the
computationalpower of a planetary-mass computeris lo42operationsper
second, and that assumesonly alreadyknownnanotechnologicaldesigns,
whichare probablyfarfromoptimal.A singlesuchcomputercould simulate
9 IOO billionhumansx 50 years/humanx 30 millionsecs/yearx [o014, Iol7 operationsin
each humanbrainper second= [1033,Io36]operations.
C The EditorsofThePhilosophical
Quarterly,
2003
248 NICK BOSTROM

theentirementalhistoryofhumankind(I shallcall thisan ancestor-simulation)


by usingless than one millionthof itsprocessingpower forone second.A
posthumancivilizationmay eventuallybuild an astronomicalnumberof
such computers.I can conclude that the computingpower available to a
posthumancivilizationis sufficient to run a huge number of ancestor-
simulationseven ifit allocatesonlya veryminutefractionofitsresourcesto
that purpose. I can draw this conclusion even while leaving a huge
marginoferrorin all our estimates:
substantial
Post-humancivilizationswould have enough computingpower to run
hugelymanyancestor-simulations even whileusingonlya tinyfraction
oftheirresourcesforthatpurpose.

IV. THE CORE OF THE SIMULATION ARGUMENT

The basic idea of thispaper can be expressedroughlyas follows:if there


were a substantialchance that our civilizationwill get to the posthuman
stageand runmanyancestor-simulations, thenhow come we are notliving
in sucha simulation?
I shalldevelopthisidea intoa rigorousargument.I need to introducethe
following notation:

fp: Fractionof all human-leveltechnologicalcivilizationsthat surviveto


reacha posthumanstage
N: Average number of ancestor-simulations run by a posthuman
civilization
H: Averagenumberof individualsthathave livedin a civilizationbeforeit
reachesa posthumanstage.
The actualfractionofall observerswithhuman-type
experiencesthatlivein
simulations
is then

sim=
+H
(f)VH)
Writingfiforthe fractionof posthumancivilizationsthatare interestedin
runningancestor-simulations(or thatcontainat least some individualswho
are interestedin them and have sufficient resourcesto run a significant
numberof such simulations), and NI forthe average numberof ancestor-
simulations
runbysuchinterested we have
civilizations,

?Th
Editors
of Th Philosophicalarter,

C The Editors of The PhilosophicalQuarterly,


2oo3
ARE WE LIVING IN A COMPUTER SIMULATION? 249
and thus

F. fm=
(fif N) +i
Because of the immensecomputingpower ofposthumancivilizations, XNis
extremely large,as I pointedout in the previoussection.What (F) showsis
thatat leastoneofthefollowing threepropositionsmustbe true:

I. fp=O
2. fj-O
3. fsizr I

V. A BLAND INDIFFERENCE PRINCIPLE

I can takea further


stepand concludethatgiventhetruthof(3),one'scredence
in the hypothesisthatone is in a simulationshouldbe close to unity.More
generally,if we knew that a fractionx of all observerswith human-type
experienceslivein simulations,and we have no information to indicatethat
our own particularexperiencesare any more or less likelythan other
human-typeexperiencesto have been implementedin vivoratherthan in
machina, thenour credencethatwe are in a simulationshouldequal x:
S. Cr(sIMI fm = x) = x.
This step is sanctionedby a veryweak indifference principle.Two cases
need to be distinguished.The firstcase, whichis theeasiest,is whereall the
minds in question are like our own in the sense that they are exactly
qualitativelyidenticalwithours:theyhave exactlythesame information and
the same experiencesas we have. The second case is wherethe mindsare
'like' each otheronly in the loose sense of being the sort of minds that
are typicalof human creatures,but where theyare qualitativelydistinct
fromone anotherand each has a distinctsetof experiences.I maintainthat
even in the lattercase, where the minds are qualitativelydifferent, the
simulationargumentstillworks,providedthatwe have no information bear-
ing on the questionofwhichof thevariousmindsare simulatedand which
are implemented biologically.
A detaileddefenceofa stronger principle,whichimpliestheabove stance
forboth cases as trivialspecialinstances,has been givenin the literature.10
Space does notpermita recapitulation ofthatdefencehere,but I can bring
10
In, e.g., N. Bostrom,'The Doomsday Argument,Adam and Eve, UN++, and Quantum
Joe', Synthese, 127 (2001), pp. 359-87; and most fullyin my book Anthropic
Bias: Observation
Selection
EfectsinScience andPhilosophy(New York: Routledge,2002).

? The Editorsof ThePhilosophical


Quarterly,
2003
250 NICK BOSTROM

out one of the underlyingideas by rehearsingan analogous situationof a


morefamiliarkind.Suppose thatx% ofthepopulationhas a certaingenetic
sequence S withinthe part of theirDNA commonlydesignatedas 'junk
DNA'. Suppose further thatthereare no manifestations of S (shortofwhat
would turnup in a gene assay)and that there are no known correlationsbe-
tweenhavingS and any observablecharacteristic. Then quiteclearly,unless
one has had one's DNA sequenced,it is rationalto assigna credenceofx%
to the hypothesis thatone has S. And thisis so quite irrespective of thefact
thatthe people who have S have qualitatively different mindsand experi-
ences fromthe people who do not have S. (They are different simplybe-
cause all humanshave different experiencesfromone another,not because
ofanyknownlinkbetweenS and whatkindofexperiencesone has.)
The same reasoningholds if S is not the propertyof havinga certain
geneticsequencebut insteadtheproperty ofbeingin a simulation, assuming
only thatwe have no information thatenables us to predictany differences
between the experiencesof simulatedminds and those of the original
biologicalminds.
It shouldbe stressedthatthebland indifference principleexpressedby (S)
prescribesindifference onlybetweenhypothesesabout whichobserverone
is, when one has no information about whichof theseobserversone is. It
does not in generalprescribeindifference betweenhypotheseswhen one
lacksspecificinformation about whichof the hypothesesis true.In contrast
with Laplacean and othermore ambitiousprinciplesof indifference, it is
therefore immuneto Bertrand'sparadox and similarpredicaments thattend
to plagueindifference principlesofunrestricted scope.
Readers familiar with the doomsday argument"may worrythat the
bland principleof indifference invokedhere is the same assumptionas is
responsibleforgettingthedoomsdayargumentofftheground,and thatthe
counter-intuitivenatureofsome oftheimplications ofthelatterincriminates
or castsdoubt on the validityof the former.This is not so. The doomsday
argumentrestson a muchstronger and morecontroversial premise,namely,
thatone shouldreason as ifone were a randomsamplefromthe set of all
people who willeverhave lived(past,present,and future)eventhough weknow
thatweareliving in theearlytwenty-first ratherthanat some pointin the
century
distantpast or the future.The bland indifference principle,by contrast,
appliesonly to cases where we have no information about whichgroupof
we
people belong to.
If bettingodds providesome guidanceto rationalbelief,it may also be
worthponderingthatifeverybody were to place a bet on whethertheyare
" See, e.g.,J. Leslie, 'Is the End of the World Nigh?', ThePhilosophical
Quarter{y,
40 (I990),
pp. 65-72.

? The EditorsofThePhilosophical
Quarterly,
2003
ARE WE LIVING IN A COMPUTER SIMULATION? 251

in a simulationor not,thenifpeople use thebland principleofindifference,


and consequentlyplace theirmoneyon being in a simulationiftheyknow
thatthatis wherealmostall people are, thenalmosteveryonewillwin their
bets.If theybet on notbeingin a simulation, thenalmosteveryonewilllose.
It seemsbetterthatthebland indifference principleshouldbe heeded.
Further,one can considera sequence of possiblesituationsin whichan
increasingfractionof all people live in simulations:98%, 99%, 99-9%,
99-9999%,and so on. As one approachesthelimitingcase in whicheverybody
is in a simulation(fromwhich one can deductively inferthat one is in a
simulationoneself),it is plausibleto requirethatthecredenceone assignsto
being in a simulationshouldgraduallyapproach the limitingcase of com-
pletecertaintyin a matchingmanner.

VI. INTERPRETATION

The possibility represented byproposition(I) is fairlystraightforward.


If (I) is
true, then humankind will almostcertainlyfail to reach a posthumanlevel;
forvirtually no speciesat our levelofdevelopmentbecome posthuman,and
it is hard to see anyjustificationforthinkingthatour own specieswill be
or
especiallyprivileged protected from futuredisasters.Conditionallyon (i),
therefore, we must a
give high credence to DOOM,the hypothesisthat
humankindwillgo extinctbeforereachinga posthumanlevel:

Cr(DOOM fp= I)= I


One can imaginehypothetical situations
wherewe have such evidenceas
would trumpknowledgeoffp.For example,ifwe discoveredthatwe were
about to be hit by a giant asteroid,thismightsuggestthatwe had been
exceptionallyunlucky.We could thenassigna credenceto DOOMlargerthan
our expectationofthefractionofhuman-levelcivilizations thatfailto reach
posthumanity. In the actual case, however,we seem to lack evidencefor
thinking thatwe are specialin thisregard,forbetteror worse.
Proposition(I) does not by itselfimplythatwe are likelyto go extinct
soon,onlythatwe are unlikelyto reach a posthumanstage.This possibility
is compatiblewithour remainingat, or somewhatabove, our currentlevel
of technologicaldevelopmentfora long timebeforegoingextinct.Another
way for (I) to be true is if it is likelythat technologicalcivilizationwill
collapse.Primitivehumansocietiesmightthenremainon Earthindefinitely.
There are manyways in which humanitycould become extinctbefore
reachingposthumanity. Perhapsthemostnaturalinterpretation of (I) is that
we are likelyto go extinctas a resultof the developmentof some powerful
? The Editors of The PhilosophicalQuarterly,
200oo3
252 NICK BOSTROM

but dangeroustechnology.'2One candidateis molecularnanotechnology,


whichin its maturestagewould enable the construction of self-replicating
nanobots capable of feedingon dirt and organic matter- a kind of
mechanicalbacteria. Such nanobots,designedfor malicious ends, could
cause theextinction ofall lifeon our planet.13
The secondalternative in thesimulationargument'sconclusionis thatthe
fractionof posthumancivilizationsthatare interested in runningancestor-
simulationsis negligiblysmall.In orderfor(2) to be true,theremustbe a
strongconvergence amongthecoursesofadvancedcivilizations. If the number
of ancestor-simulations createdby the interestedcivilizationsis extremely
large, the rarityof such civilizationsmust be correspondingly extreme.
no
Virtually posthuman civilizations decide to use their resources to run
large numbers ofancestor-simulations. Furthermore, all
virtually posthuman
civilizationslack individualswho have sufficient resourcesand interestto
run ancestor-simulations; or else they have reliablyenforcedlaws that
prevent such individuals from actingon theirdesires.
What forcecould bringabout such a convergence?One mightspeculate
thatadvancedcivilizations all developalong a trajectory thatleads to recog-
nitionofan ethicalprohibition againstrunning ancestor-simulations because
of the suffering thatis inflicted on the inhabitantsof the simulation.How-
ever,fromour presentpointof view,it is not clear thatcreatinga human
race is immoral.On thecontrary, we tendto view theexistenceof our race
as constituting a greatethicalvalue. Moreover,convergenceon an ethical
view of the immoralityof runningancestor-simulations is not enough: it
mustbe combinedwithconvergenceon a civilization-wide social structure
thatenablesactivities consideredimmoralto be effectively banned.
Anotherpossible convergencepoint is that almost all individualpost-
humansin virtually all posthumancivilizations developin a directionwhere
theylose theirdesireto run ancestor-simulations. This would requiresigni-
ficantchangesto themotivesdrivingtheirhumanpredecessors, forthereare
certainlymanyhumanswho would like to run ancestor-simulations ifthey
could affordto do so. But perhaps many of our human desireswill be
regarded as silly by anyone who becomes a posthuman.Maybe the
scientificvalue of ancestor-simulations to a posthuman civilizationis
negligible(which is not too implausiblegivenits unfathomableintellectual
and
superiority), maybeposthumansregardrecreationalactivitiesas merely
12 See
my'ExistentialRisks:AnalyzingHuman ExtinctionScenariosand Related Hazards',
Journalof Evolution and 9 (2001), for a surveyand analysis of the presentand
Technology,
anticipatedfuturethreatsto humansurvival.
13 See, e.g., Drexler;and R.A. FreitasJr,'Some Limitsto Global
Ecophagyby Biovorous
Nanoreplicators, with Public Policy Recommendations', Zyvexpreprint, April (2000),
http://www.foresight.org/NanoRev/Ecophagy.html.

? The Editorsof ThePhilosophical


Quarterly,
2003
ARE WE LIVING IN A COMPUTER SIMULATION? 253

a veryinefficient way of gettingpleasure- which can be obtainedmuch


more cheaply by direct stimulationof the brain's reward centres.One
conclusionthat followsfrom(2) is that posthumansocietieswill be very
different fromhuman societies:they will not contain relativelywealthy
independentagentswho have the fullgamutof human-likedesiresand are
freeto act on them.
The possibilityexpressedby alternative(3) is conceptuallythe most
intriguing one. If we are livingin a simulation,thenthecosmoswe are ob-
servingisjust a tinypiece ofthetotality ofphysicalexistence.The physicsin
the universewherethe computerrunningthe simulationis situatedmay or
may not resemblethephysicsof theworldwe observe.While theworldwe
see is in some sense'real',itis notlocatedat thefundamental levelofreality.
It maybe possibleforsimulatedcivilizations to becomeposthuman.They
may thenrun theirown ancestor-simulations on powerfulcomputersthey
build in their simulated universe. Such computerswould be 'virtual
machines',a familiarconceptin computerscience.(Javascript web-applets,
forinstance,run on a virtualmachine- a simulatedcomputer- insidea
desktop.)Virtual machinescan be stacked:it is possible to simulateone
machinesimulating anothermachine,and so on, in arbitrarily manystepsof
iteration.If we do go on to createour own ancestor-simulations, thensince
thiswouldbe strongevidenceagainst(I) and (2),we wouldtherefore have to
concludethatwe live in a simulation.Moreover,we would have to suspect
that the posthumansrunningour simulationare themselvessimulated
beings;and theircreatorsin turnmayalso be simulatedbeings.
Realitymaythuscontainmanylevels.Even ifit is necessaryforthehier-
archyto bottomout at some stage- themetaphysical statusof thisclaim is
somewhatobscure- theremay be room fora large numberof levels of
reality,and the numbercould be increasingover time.(One consideration
thatcountsagainstthe multi-level hypothesisis thatthe computationalcost
forthe basement-level simulatorswould be verygreat.Simulatingeven a
singleposthumancivilizationmightbe prohibitively expensive.If so, then
we should expect our simulationto be terminatedwhen we are about to
becomeposthuman.)
Althoughall the elementsof such a systemcan be naturalistic,even
physical,it is possibleto draw some loose analogieswithreligiousconcep-
tionsof theworld.In some ways,theposthumansrunninga simulationare
like gods in relationto the people inhabitingthe simulation:the post-
humanscreatedtheworldwe see; theyare of superiorintelligence; theyare
'omnipotent' in the sense that they can interferein the workingsof our
worldevenin waysthatviolateitsphysicallaws; and theyare 'omniscient'in
the sense thattheycan monitoreverything thathappens.However,all the
? The Editors of The PhilosophicalQuarterly,
2003
254 NICK BOSTROM

demigodsexcept those at the fundamentallevel of realityare subjectto


sanctionsbythemorepowerfulgodslivingat lowerlevels.
Furtherruminationon thesethemescould climaxin a naturalistic theogony
thatwouldstudythestructure ofthishierarchy, and theconstraints imposed
on its inhabitantsby the possibilitythat theiractionson theirown level
may affectthe treatmenttheyreceivefromdwellersof deeper levels.For
example,if nobody can be sure thattheyare at the basement-level, then
everybodywould have to considerthe possibility thattheiractionswill be
rewardedor punished,perhapsusingmoralcriteria,by theirsimulators. An
afterlifewouldbe a realpossibility. Because ofthisfundamental uncertainty,
even the basementcivilizationmay have a reasonto behave ethically.The
factthatit has such a reason formoral behaviourwould of course add to
everybodyelse's reasonforbehavingmorally,and so on, in a trulyvirtuous
circle.One mightget a kindof universalethicalimperative, whichit would
be in everybody's to obey,as itwere,'fromnowhere'.
self-interest
In additionto ancestor-simulations, one may also considerthepossibility
ofmoreselectivesimulations thatincludeonlya smallgroupofhumansor a
singleindividual.The restof humanitywould thenbe zombiesor 'shadow-
people' - humanssimulatedonlyat a level sufficient forthe fullysimulated
ones not to noticeanythingsuspicious.It is not clear how much cheaper
shadow-peoplewouldbe to simulatethanreal people. It is noteven obvious
that it is possible for an entityto behave indistinguishably froma real
human and yetlack consciousexperience.Even ifthereare such selective
simulations, we shouldnotthinkthatwe are in one of themunlesswe think
theyare much more numerousthan completesimulations.There would
have to be about Ioo billiontimesas many'me-simulations' (simulations of
the lifeof onlya singlemind)as thereare ancestor-simulations in orderfor
mostsimulatedpersonsto be in me-simulations.
There is also the possibility of simulatorsabridgingcertainpartsof the
mentallivesof simulatedbeingsand givingthemfalsememoriesof the sort
of experiencesthattheywould typically have had duringthe omittedinter-
val. If so, one can consider the following(far-fetched) solutionto the
problemof evil: thatthereis no suffering in the worldand all memoriesof
suffering are illusions.Of coursethishypothesis can be seriouslyentertained
only at those times when one is not currentlysuffering.
Supposingwe live in a simulation, whatare the implications forus? The
foregoing remarks the
notwithstanding, implications are not all that radical.
Our best guide to how our posthumancreatorshave chosen to set up our
worldis thestandardempiricalstudyoftheuniversewe see. The revisionsto
most parts of our beliefnetworkswould be ratherslightand subtle- in
proportionto our lack ofconfidencein our abilityto understandthewaysof
? The EditorsofThePhilosophical
Quarter~y,
2003
ARE WE LIVING IN A COMPUTER SIMULATION? 255

posthumans.Properlyunderstood,therefore, thetruthof (3) shouldhave no


tendencyto make us 'go crazy'or to preventus fromgoingabout our busi-
ness and makingplans and predictionsfortomorrow.The chiefempirical
importanceof (3) at the presenttimeseemsto lie in itsrole in thetripartite
conclusionestablishedabove.14We may hope that (3) is true,since that
would decreasethe probabilityof (I), althoughifcomputationalconstraints
make itlikelythatsimulators would terminate a simulationbeforeit reaches
a posthumanlevel,thenour besthope wouldbe that(2) is true.
If we learn more about posthumanmotivesand resource-constraints,
maybe as a resultof developingtowardsbecomingposthumansourselves,
thenthe hypothesis thatwe are simulatedwillcome to have a much richer
setofempiricalimplications.

VII. CONCLUSION

A technologically mature 'posthuman'civilizationwould have enormous


computingpower.Giventhisempiricalfact,thesimulationargumentshows
thatat leastoneofthefollowing propositionsis true:(I) thefractionofhuman-
levelcivilizations thatreach a posthumanstageis veryclose to zero; (2) the
fractionof posthumancivilizationsthatare interested in runningancestor-
simulations is veryclose to zero; (3) thefraction ofall people withour kindof
experienceswho are livingin a simulationis veryclose to one.
If (I) is true,then we will almostcertainlygo extinctbeforereaching
posthumanity. If (2) is true,thentheremustbe a strongconvergenceamong
the courses of advanced civilizationsso that virtuallynone containsany
relatively wealthyindividualswho desireto runancestor-simulations and are
freeto do so. If (3) is true,thenwe almostcertainlylive in a simulation.In
the darkforestofour currentignorance,it seemssensibleto apportionone's
credenceroughlyevenlybetween (2),and (3).
(I),
Unless we are now livingin a simulation,our descendantswill almost
certainly neverrunan ancestor-simulation.'5

University
Oxford

14For some reflectionsby anotherauthoron the consequencesof (3), whichwere sparked


by a privatelycirculatedearlierversionof thispaper, see R. Hanson, 'How to Live in a
Simulation',Journal andTechnology,
ofEvolution 7 (2001).
15 I am grateful
to manypeople forcomments,and especiallyto Amara Angelica,Robert
Bradbury,Milan Cirkovie,Robin Hanson, Hal Finney,Robert A. FreitasJr,John Leslie,
Mitch Porter,Keith DeRose, Mike Treder, Mark Walker, Eliezer Yudkowsky,and the
anonymousreferees.
C The Editorsof ThePhilosophical
Quarterly,
2o003

Vous aimerez peut-être aussi