ResearchReport

TOBIASSYTSMA

QuantifyingAI’s

EconomicPotential

GrowthDifferentialsBetweenAssistiveandAutonomous

DevelopmentScenarios

RR-A4220-1

Formoreinformationonthispublication,visit

/t/RRA4220-1

.

AboutRAND

RANDisaresearchorganizationthatdevelopssolutionstopublicpolicychallengestohelpmakecommunitiesthroughouttheworldsaferandmoresecure,healthierandmoreprosperous.RANDisnonprofit,nonpartisan,andcommittedtothepublicinterest.TolearnmoreaboutRAND,visit

.

ResearchIntegrity

Ourmissiontohelpimprovepolicyanddecisionmakingthroughresearchandanalysisisenabledthroughourcorevaluesofqualityandobjectivityandourunwaveringcommitmenttothehighestlevelofintegrityandethicalbehavior.Tohelpensureourresearchandanalysisarerigorous,objective,andnonpartisan,wesubjectourresearchpublicationstoarobustandexactingquality-assuranceprocess;avoidboththeappearanceandrealityoffinancialandotherconflictsofinterestthroughstafftraining,projectscreening,andapolicyofmandatorydisclosure;andpursuetransparencyinourresearchengagementsthroughourcommitmenttotheopenpublicationofourresearchfindingsandrecommendations,disclosureofthesourceoffundingofpublishedresearch,andpoliciestoensureintellectualindependence.Formoreinformation,visit

/about/research-integrity

.

RAND’spublicationsdonotnecessarilyreflecttheopinionsofitsresearchclientsandsponsors.

PublishedbytheRANDCorporation,SantaMonica,Calif.

?2025RANDCorporation

RANDeisaregisteredtrademark.

LimitedPrintandElectronicDistributionRights

Thispublicationandtrademark(s)containedhereinareprotectedbylaw.ThisrepresentationofRANDintellectualpropertyisprovidedfornoncommercialuseonly.Unauthorizedpostingofthispublicationonlineisprohibited;linkingdirectlytoitswebpageonisencouraged.PermissionisrequiredfromRANDtoreproduce,orreuseinanotherform,anyofitsresearchproductsforcommercialpurposes.Forinformationonreprintandreusepermissions,visit

/about/publishing/permissions

.

iii

AboutThisReport

Theanalysisinthisreportmodelstheeconomicimplicationsoftwocontrastingboundary

scenariosforartificialintelligence(AI)development:restrictingAItopurelyassistivetoolsthat

augmenthumanproductivityversusenablingagentsthatcanautonomouslyperformtasksatleastaswellashumansandreplicatethemselves+TheAgentWorldscenariorepresentsthepotential

emergenceofartificialgeneralintelligence(AGI)capabilities,whereastheToolWorldscenario

representsAIprogressasavarietyofnarrow,specializedsystemsthatenhancehumanproductivitybutdonotachievegeneralintelligence.Thesestylizedscenariosarenotmeanttobeinterpretedaspredictions;rather,theyaremeanttoestablishavarietyofpotentialeconomicoutcomesfor

policymakers,economicadvisors,andresearchersworkingattheintersectionoftechnologypolicyandeconomicstrategy+UsingacalibratedendogenousgrowthmodelandMonteCarlosimulationsfrom2025to2045,theanalysissuggeststhatembracingAIagentscouldresultintheeconomy

growing3+8percentagepointsfasterannually,onaverage,thanlimitingAItotoolswould+ThisdifferencecompoundstomakeAgentWorldeconomiesnearlyfourtimeslargerby2045+

AlthoughtheAgentWorldscenarioassumessuccessfulresolutionofAIsafetyandalignmentchallenges,theresultshighlightthesubstantialeconomicincentivesdrivingtowardautonomousAIdevelopmentandillustratetheeconomictrade-offsthatareinherenttodifferentAI

developmentstrategies+

TechnologyandSecurityPolicyCenter

RANDGlobalandEmergingRisksisadivisionofRANDthatdeliversrigorousandobjectivepublicpolicyresearchonthemostconsequentialchallengestocivilizationandglobalsecurity+Thisworkwasundertakenbythedivision’sTechnologyandSecurityPolicyCenter,whichexplores

howhigh-consequence,dual-usetechnologieschangetheglobalcompetitionandthreat

environment,thendevelopspolicyandtechnologyoptionstoadvancethesecurityoftheUnitedStates,itsalliesandpartners,andtheworld+Formoreinformation,contact

tasp@rand+org

+

Funding

TheGeopoliticsofAGIInitiativeisindependentlyinitiatedandconductedwithinthe

TechnologyandSecurityPolicyCenter+FundingforthisresearchwasprovidedbyincomefromoperationsandgiftsfromRANDsupporters,includingphilanthropicgiftsmadeorrecommendedbyDALHAPInvestmentsLtd+,ErgoImpact,FoundersPledge,CharlottesochFredriksStiftelse,GoodVentures,JaanTallinn,Longview,andOpenPhilanthropy+Formoreinformation,contact

tasp@rand+org

+RANDdonorsandgrantorshavenoinfluenceoverresearchfindingsor

recommendations+

iv

Acknowledgments

IwouldliketothankJoelPredd,JimMitre,LisaAbraham,JessieWang,DanielMarkus,JeffAlstott,andthebroaderRANDGeopoliticsofAGIInitiativeforcommentsonanearlierversionofthisreport+IwouldalsoliketothankVegardNygaardandJimBakerfortheirrefereereports+

v

Summary

Theanalysisinthisreportmodelstheeconomicimplicationsoftwocontrastingboundary

scenariosforartificialintelligence(AI)development:restrictingAItopurelyassistivetoolsthat

augmenthumanproductivityversusenablingagentsthatcanautonomouslyperformtasksatleastaswellashumansandreplicatethemselves+TheAgentWorldscenariorepresentsthepotential

emergenceofartificialgeneralintelligence(AGI)capabilities,whereastheToolWorldscenario

representsAIprogressashavingavarietyofnarrow,specializedsystemsthatenhancehuman

productivitybutdonotachievegeneralintelligence+Thesestylizedscenariosarenotmeanttobeinterpretedaspredictions;rather,theyaremeanttoestablishavarietyofpotentialeconomic

outcomesforpolicymakers,economicadvisors,andresearchersworkingattheintersectionof

technologypolicyandeconomicstrategy+Usinganendogenousgrowthmodelcalibratedwith

MonteCarlosimulationsfrom2025to2045,theanalysissuggeststhatembracingAIagentscouldresultintheeconomygrowing3+8percentagepointsfasterannually,onaverage,thanlimitingAItotoolswould+ThisdifferencecompoundstomaketheAgentWorldeconomythreetimeslargerby2045+AlthoughtheAgentWorldscenarioassumessuccessfulresolutionofAIsafetyand

alignmentchallenges,theresultshighlightthesubstantialeconomicincentivesdrivingtoward

autonomousAIdevelopmentandillustratetheeconomictrade-offsthatareinherenttodifferentAIdevelopmentstrategies+

KeyFindings

ThisanalysiscontraststwostylizedboundaryscenariosthatspanexistingAIcapabilitiestopotentialAGIandartificialsuperintelligence:aToolWorldinwhichAIremainsstrictlyassistiveunderhumansandanAgentWorldinwhichAIisfullyautonomousandcancompletelyreplacehumanworkers+AgentWorldassumessafe,aligned,andsuccessfullydeployedadvancedAI

systems,whichareunproven+Theresults,asfollows,establishanupperboundonpotentialeconomicupsideswithoutimplyingthatsuchsolutionsareguaranteed:

?BuildingonestablishedeconomicgrowththeoryandusingMonteCarlosimulation

methods,theresultsindicatethatAgentWorldexceedsToolWorldbyanaverageof3+8percentagepointsinannualgrossdomesticproduct(GDP)growth,withamedian

differenceof2+6percentagepoints+

?By2045,themodelsuggeststhatAgentWorld’sGDPisapproximately3+6timeslargerthanToolWorld’sGDP,assumingsuccessfulAIdeploymentwithoutmajortransitioncostsorpolicyconstraints+SomesimulationssuggestthatAgentWorldGDPisseveralordersofmagnitudelarger+

?RegressiontreeanalysisofmodelsimulationsidentifiestheconversionratefromcomputetoAIagentsandthecompute-specificinvestmentrateasthemostimportantpredictivefactorsforgrowthdifferentials+

?Usingthemodel,AgentWorldeconomieswith90percentormoreAI-agentshareoftheresearchanddevelopment(R&D)workforceachieve25ormorepercentagepointgrowth

vi

advantagesoverToolWorldby2045,whilepartialintegrationcreatessmaller

improvements,suggestingthatcompletetransformationunlocksqualitativelydifferentgrowthregimes+

?Onaverage,theeconomicopportunitycostofforgoingAgentWorlddevelopment

through2045wouldrequirenonmodeledbenefits(suchassafety,stability,or

distributionaladvantages)thatareequivalentto1+2–2+8timestheexistingannualGDP,dependingonriskpreference+

vii

Contents

AboutThisReport++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++iiiSummary+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++vFiguresandTables+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++viiiPutaNumberonIt:ValuingAIAgencyintheTrade-OffBetweenAIBenefitsandRisk+++++++++++++++++++++++++++++1 Model++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++4 EmpiricalCalibrationandSimulationMethods+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++14 Results+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++15 Limitations+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++25 Conclusion++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++28Appendix+ParameterCalibrationandJustification+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++30Abbreviations++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++35References+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++36

viii

FiguresandTables

Figures

Figure1+ModelSchematic++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++6Figure2+TrajectoriesofAgentWorldGDPRelativetoToolWorldGDP+++++++++++++++++++++++++++++++++++++++++++++++++++15Figure3+DistributionofGrowthDifferentialsBetweenAgentandToolWorlds+++++++++++++++++++++++++++++++++++++++++16Figure4+GrowthRateDifferentialsandNumberofAIAgents+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++18Figure5+FactorImportanceinExplainingGrowthDifferentials++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++19Figure6+GrowthDynamicsandAIAgentWorkforce+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++20Figure7+AgentWorldTFPGrowth++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++21

Tables

Table1+KeyModelParameters,Interpretation,andCalibration+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++13Table2+GrowthResultsbyScenario++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++16Table3+RequiredBoostforToolWorldRelativetoAgentWorldThrough2045+++++++++++++++++++++++++++++++++++++++23Table4+PresentValueofPullingtheTakeoffYearForward++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++24

1

QuantifyingAI’sEconomicPotential:GrowthDifferentialsBetween

AssistiveandAutonomousDevelopmentScenarios

Recentbreakthroughsinartificialintelligence(AI)havesparkedadebateoverhowAIshouldbeintegratedintoeconomiesandsociety+SomeexpertsadvocateforAIasapowerfulbutstrictlyassistivetoolthatshouldremainundertighthumansupervision+Othersforesee(orwarnof)AIasanautonomousagentthatiscapableofmakingdecisionsandperformingtaskswithoutdirect

humanoversight+Thisdividedefinesthefollowingpolicychallenge:ShouldweaccelerateagenticAItocaptureitsunprecedentedupsidesorlimitAItobeingonlyatooltoguardagainst

potentiallycatastrophicorevenexistentialrisks?

ThedegreeofautonomythatispermittedinAIsystemsrepresentsapolicydimensionwithsignificantimplicationsforeconomicgrowth,labormarkets,globalcompetitivedynamics,and

humanity’slong-termsafety+Asofthiswriting,thisdebatehaslimitedempiricalgrounding,whileseeminglyunboundedpotentialforscientificadvancementandgrowthisweighedagainst

potentiallycatastrophiccosts+ThisreportcontributestothisdiscoursebymodelingthepotentialeconomicimplicationsofdifferentlevelsofconstraintonAI+Init,Iestablishquantitative

referencepointsthatmayhelppolicymakersunderstandthemagnitudeoftheeconomictrade-offsthatareinvolvedinvariousapproachestoAIgovernance,alongsidesafetyconcerns,transition

costs,anddistributionaleffectsthatarenotcapturedinthisanalysis+

Economistshavelongrecognizedthattechnologyisaprimarydriverofeconomicgrowth+Inclassicalgrowththeory,advancesintotalfactorproductivity(TFP)(whichisofteninterpretedastechnology)arenecessarytosustainlong-termgrowthbecausecapitalaccumulationfaces

diminishingreturns(Solow,1956)+EndogenousgrowthmodelsfurtherformalizehowintentionalresearchandinnovationinvestmentscanproducenewideasthatdriveTFPupward(Romer,

1990)+Akeyinsightfromthisliteratureisthataneffectivesupplyofresearchers(ormore

generally,problem-solvers)iscriticalforinnovation+Ifmoremindsaredevotedtoresearchanddevelopment(R&D),moreideascanbedeveloped,althoughwithdiminishingreturnsasexistingknowledgegrows(Bloometal+,2020;Jones,1995)+1However,intheexistinghuman-centric

paradigm,thestockofresearchersisultimatelyconstrainedbypopulation,demographics,andtrainingpipelines+Thisobservationhasledtoconcernsthatideaproductionisslowingdown+

1AdvancedAIsystemsmightpotentiallyalterthisdynamicbyexcellingatknowledgesynthesisacrossincreasingly

fragmentedandspecializedresearchfrontiers,whichwouldchangethequantityofresearchinputsandtheefficiencyoftheresearchprocessitself+However,evensuperintelligentAIresearchersmightencounterdiminishingreturns,whethertothefundamentalcomplexityofremainingunsolvedproblems,thetimerequiredforreal-worldexperimentationandvalidation,orevencoordinationcostsasresearchteams’scaleup+Theanalysistreatssuchstructuralresearchrelationshipsasfixed

parametersthatarecalibratedtohistoricalexperience,but,asdiscussedinthelimitationssection,transformativeAImightendogenouslyaltertheserelationshipsinwaysthatcouldamplifythegrowtheffectsestimatedhere+

2

Evidencesuggeststhat,despiterisingR&Deffort(orspending),researchproductivityper

researcherisdeclininginmanyfields(Bloometal+,2020)+2Inotherwords,ideasaregettinghardertoadvance,andtheexponentialgrowthofknowledgemaybedeceleratingunderhuman-only

research+

AIhasthepotentialtoalterthisequation+IfAIsystemscantakeonresearchtasksorevenimproveAIitself,thenumberofeffective“researchers”couldincreasefarbeyondthehumanworkforce+Historically,eachmajortechnologicalera(e+g+,steam,electricity,computing)

acceleratedgrowthforatimebuteventuallyfaceddiminishingreturnsinpartbecausehumansremainedthecentralactorsininnovation+AdvancedAIthatcaneithersignificantlyaugmenthumanresearchers(intheToolparadigm)orreplicateandreplacehumanresearchers(intheAgentparadigm)mightbreakthroughsomeoftheselimits+

Recentresearchhasbegunexploringthesepossibilities+Aghion,Jones,andJones(2019)

considersAIasaformofautomationinR&Dandaskwhetheritcouldleadtoever-acceleratinggrowthorevena“singularity”scenario+Theauthors’analysissuggeststhatAI-drivengrowthcouldsharplyincreaseinnovation+However,variousbottlenecks(includingtheneedforhumaninputinsomestagesofinnovation)mightstillcapgrowthrates(Aghion,Jones,andJones,2019)+

Similarly,Nordhaus(2021)examineswhetherrapidprogressinAIandcomputingcouldleadtoaneconomicsingularity(essentiallyunboundedgrowth)andconcludesthatsucharegimeshiftisnotimminentunderbase-caseassumptions,althoughitremainsalonger-termpossibility+Morerecently,Jones(2024)framesthesituationasthe“AIdilemma”:Ononehand,removingthe

humanconstraintsoninnovationviaAIcouldusherinaperiodofunprecedentedeconomic

growth;ontheotherhand,suchadevelopmentraisesconcernsofexistentialriskifAIsystemsbecometoopowerfulormisaligned+JonesarguesthatifhumanlaborisabottleneckforAI

progress,growthwillremainsomewhatconstrainedandmanageable;ifAIgainstheabilityto

effectivelyself-replicate(intermsofresearchefforts),thenexplosivegrowthandlossofhumancontrolbecomereal+Similarly,AcemogluandLensman(2024)examinesoptimaladoptionpathsfortransformativetechnologiesthatofferproductivitygainsbutposepotentialdisasterrisks,

findingthatgradualadoptioncanbeoptimaltoenablelearningaboutrisksovertime+

ThisreportcontributestothisdiscoursebyprovidingaquantitativemodelingexercisefocusedontheeconomicgrowthimplicationsofAItoolsversusAIagents+Iconstructanendogenous

growthmodelinthespiritofRomer(1990),Jones(1995),andBloometal+(2020),withthefollowingtwomodifications:

1+AIcancontributetotheproductionofgoodsandserviceseitherindirectlyasatoolraisinghumanproductivityordirectlyasrobotlaborreplacinghumans(AIinproduction)+

2+AIcancontributetoideagenerationeitherasaresearch-augmentingtoolorasanautonomousresearcherthataddstothetotalR&Dworkforce(AIininnovation)+

Idefinetwostylizedscenariosforthefuture:oneinwhichAIisneverallowed(orneverable)tofullyreplacehumanworkersorresearchers(ToolWorld)andoneinwhichAIisfreeto

substituteforhumanlaborandcanbereplicatedasneeded(AgentWorld)+ToolWorld

representsthecontinuationofexistingAIdevelopmenttrajectories,havingpowerfulbutnarrowsystemsthataugmenthumancapabilitiesthatstillrequiresomedegreeofhumanoversightand

2SomeresearchsuggeststhatthedecliningresearchproductivitydocumentedinBloometal+(2020)isoverstatedandthatapproximatelyhalfofthestateddeclinecanbeexplainedbyselectioneffects+Forinstance,EkerdtandWu(2025)findthattheabilityoftheaverageresearcherhasdeclinedbyroughly48percentastheR&Dworkforcehasexpandedovertime+

3

control+AgentWorldrepresentsapotentiallyfundamentalshifttowardartificialgeneral

intelligence(AGI),orevenartificialsuperintelligence,inwhichAIsystemsachievehuman-levelorsuperhumancognitivecapabilitiesacrossalldomainsandcanoperatefullyautonomously+

Althoughamixofthesescenariosmayunfold,examiningextremesisusefultoboundthepotentialoutcomes+Importantly,myaimisnottoadvocateonepathortheotherbuttoilluminatethe

economicstakesinvolved+

ThisreportcontributesempiricalanalysistotheToolAIversusAgentA(G)Idebate,whichoftenproceedsonnormativegroundsalone+MaxTegmarkhasemphasizedthevalueof

controllableToolAIsystemsthatcandelivertransformativeadvancesinhealthcare,climate,anddevelopmentwhileavoidingtheexistentialrisksofunboundedAGI+3Conversely,Gwern(the

independentresearcherknownforinfluentialwritingonAIdevelopmentandcapabilities)arguesthatAgentAIsystemsinherentlyoutperformToolAIacrossvarioustasks,yieldingsuperior

economicefficiency(Gwern,2019)+Imodeltwoextremefutures(AgentWorldversusTool

World)andquantifytheireconomicgrowthdifferentialundertheassumptionthattechnical

safetyandalignmentchallengescanbesuccessfullyresolvedintheAgentWorldscenario+Ratherthanadjudicatingwhichparadigmispreferablefromawelfareorpolicyperspective,Iaimto

establishthescaleofpotentialeconomicgainsfromautonomousAIsystems,providingoneinputforpolicymakersweighingthesebenefitsagainstsafetyrisks,transitioncosts,anddistributionalconcerns+

Giventhemodelsetup,undermostparameterranges,AgentWorldgrowsatasignificantly

fasterpacethanToolWorld,whichisunsurprisingbecausethemodelexplicitlydefinesTool

Worldasbottleneckedbythehumanpopulationgrowthrate,aconstraintthatisresolvedin

AgentWorld+However,thisgrowthdifferenceisnotahardandfastlaw+Inapproximately2

percentofsimulations,ToolWorldoutgrowsAgentWorldoverthe20-yearhorizon+ThesecasesemergeonlywhenAgentWorldispopulatedbyrelativelyinefficientagentsandcompute-starved+However,inthevastmajorityofcases,AgentWorldoffersasignificantgrowthboostrelativetoToolWorld+Theaveragegrowthdifferentialbetweenthetwoworldsisnearly4percentagepointsperyear,andthemediandifferentialisslightlylower+By2045,AgentWorld’sgrossdomestic

product(GDP)isapproximatelythreetimeslargerthanToolWorld’sonaverage+

TheprojectedgrowthratesalignwiththebroaderliteratureonAGI’spotentialeconomic

impact+AgentWorld’saverageannualgrowthof4–6percentmatchesscenariosdiscussedby

leadingresearchers+Forexample,JackClarkofAnthropichassuggestedthatAGIcouldplausiblydriveapproximately5percentofannualGDPgrowth(Cowen,2025)+Therighttailofthe

distributionofmodelsimulationsreflectsscenariosofeconomicsingularitiesand

superintelligence-drivenacceleration(Davidson,2021;ErdilandBesiro?lu,2024;Korinekand

Suh,2024)+Attheconservativeend,theaverageToolWorldscenarioprojectsanadditional0+2percentagepointsofgrowthabovebaseline,whichisclosetoAcemoglu’s(2024)estimateof0+10–0+16percentagepointsoveradecade+Inthisreport,theMonteCarlosimulationsspanthe

spectrumfromconservativeestimatestotransformativeAGIscenarios,providingacomprehensiveassessmentofplausibleeconomicfutures+

IalsoevaluatewhichfactorscontributethemosttothegrowthratedifferencesbetweenToolWorldandAgentWorld,findingthatfactorsrelatingtotherateatwhichsocietyinvestsinAI-specificcapital(e+g+,compute,datacenters)andtheconversionbetweencomputeandfull-time

3MaxTegmarkhaswrittenandspokenextensivelyonthistopic,includingonsocialmedia(Tegmark,2025),ininterviews(Tegmark,2024),andinwriting(TegmarkandOmohundro,2023)+

4

equivalent(FTE)AIagentsarethemostimportantpredictivefactors+ThisfindingsuggeststhateveninaworldinwhichAIagentscanfullyreplacehumanworkersandrapidlymultiply,therearestilltangibleconstraintsongrowth+4

Ialsoassessmechanisms+Specifically,themodelsuggeststhatthebenefitsofAI-research

integrationexhibitsharpthresholdeffects,witheconomiesthatachieveover90percentoftheAI-agentresearchershareexperiencinggrowthrateadvantagesexceeding15percentannuallyby

2045,whilepartialintegration(10–50percent)deliversmarginalimprovements+Thisnonlinearityreflectsthe“burdenofknowledge”mechanismbywhichonlyAgentWorldsthatreachsufficientagenticscalecanovercomethisconstraintandachievetransformativeaccelerationafter2040+

Theseresultsareconditionalonmodelingassumptions;however,theysuggestthatpartialAIagentintegrationstrategiesmaycaptureonlyafractionofthepotentialbenefits,whilecompletetransformationcancreatequalitativelydifferentgrowthregimes+

Beyondquantifyinggrowthdifferentials,theanalysisprovidestentativeeconomicbenchmarksforweighingsafetyagainstdevelopmentspeedoverthenexttwodecades+DespitemodelingAgentWorldwithmaximumeconomicbenefitswhileassumingawayimplementationchallenges,the

analysisfindsthatsafetybenefitsworthapproximatelyonetothreetimesexistingGDP

(dependingonriskaversion)wouldjustifychoosingmore-restrictiveAIdevelopmentpolicies

through2045+Althoughthisresultishighlycontingentonmodelingassumptions,itissimilartothoseinJones(2024),whichfindsthatevenexplosiveeconomicgrowthscenariosmaynotjustifyextremerisksundercertainriskpreferences+TheanalysisalsosuggeststhatadvancingAgent

Worlddevelopmentbyoneyearcouldbeworthapproximately60percentofGDP(withsubstantialuncertainty),revealingstrongeconomicincentivesforrapiddeployment+

Inwhatfollows,Ifirstoutlinethemodel’sstructureandassumptions,includinghowI

incorporateAIintothegrowthmodel+Ithendescribethecalibrationofparametersandempiricalfoundationsofthesimulation+Next,IpresentsimulationresultsthatcompareGDPgrowthinthetwoscenarios+Idiscussthepolicyandeconomicinsightsfromthesefindingsandaddressthe

model’slimitationsandsimplifyingassumptions+IconcludebyreflectingonwhattheanalysisimpliesforeconomicpolicyinaneraofrapidAIprogress+

Model

Idevelopatwo-sectorendogenousgrowthmodelthatextendstheclassicsemi-endogenous

growthframework(Jones,1995;Romer,1990)byexplicitlyincludingAIasafactorofproductionandinnovation+ThemodeldistinguishesbetweenhumanandAIcontributionsinboththe

productionofoutput(goodsandservices)andintheproductionofnewideas(R&D)+Inessence,themodelisastandardgrowthmodelthatisaugmentedwithanAIworkforceandanAI-drivenresearchprocess+

Theanalysismodelstwostylizedextremesboundbypotentialeconomicoutcomesratherthanforecastingrealisticscenarios+TheAgentWorldscenarioassumesthattechnicalsafety,alignment,andcontrolchallengesaresuccessfullyresolved,allowingA