1 Global Economic Impacts Associated with Artificial Intelligence Nicholas Chen, Lau Christensen, Kevin Gallagher, Rosamond Mate, Greg Rafert 1 I. Introduction Artificial intelligence (“AI”), a term first coined in 1956, is a branch of computer science that aims to create intelligent machines that work and react like humans. 2 At the beginning of the study of AI, British mathematician Alan Turing proposed the true indication of computer intelligence to be when a question‐asker could not distinguish between answers from a human and those from a computer. 3 In contrast, today, 60 years later, AI is characterized by a number of applications, including computers playing games against humans and understanding human languages, virtual personal assistants, and robotics which involve computers seeing, hearing, and reacting to sensory stimuli. 4 Looking forward to the next decade, technologists have offered a wide array of predictions for AI, ranging from AI being used as a tool to aid relatively simple processes (which some refer to as weak AI) to robots with human‐like mental capabilities (which is sometimes referred to as strong AI). 5 According to AI expert Sir Nigel Shadbolt: “What we really have in AI is a whole spectrum of abilities, from programs that are smart but they are not smart like us, to programs that are super clever in specific areas.” 6 For the purposes of our study, we consider AI to broadly be computational devices and systems made to act in a manner that can be deemed intelligent. 7 In other words, AI is technology that appears to emulate human performance by learning, coming to its own conclusions, understanding complex content, engaging in dialog with people, enhancing human cognitive performance, or replacing humans in executing both routine and non‐ routine tasks. Because of the uncertainty of the future development and diffusion of AI, this view is intentionally broad, covering the AI existing today, such as targeted advertising and virtual personal assistants, as well as the AI that may exist in the future, such as robots with human‐like processing capabilities. The range of AI’s progress in the future will determine 1 The authors are all employed by Analysis Group, Inc. Funding for this study was provided by Facebook, Inc. 2 Miller, Stephen. "Computer Scientist Coined 'Artificial Intelligence'." WSJ, 26 Oct. 2011. 3 Turing, A. “Computing machinery and intelligence.” Mind 1950. Reprinted in Computation and Intelligence by G. Luger, ed. MIT Press, Cambridge, 1995, pp.23‐46. 4 Pareek, Rahul. "Web Intelligence‐An Emerging Vertical of Artificial Intelligence." International Journal Of Engineering And Computer Science 3.12 (2012): 9430‐436. 5 For example, “However, the characteristic of [AI’s current] success is what I call the combination of brute force and a little insight” (Shadbolt, Nigel, “Why We Should Not Fear AI. Yet.” Wired. May 8, 2015) and “If you think Siri is useful now, the next decade's generation of Siri will be much more like JARVIS from Iron Man, with expanded capabilities to understand and answer… In a decade, it will be normal for you to give your AI access to listen to all of your conversations, read your emails and scan your biometric data because the upside and convenience will be so immense. (Diamandis, Peter, “The World in 2025: 8 Predictions for the Next 10 Years.” Singularity Hub, May 11, 2015. 6 “The Life Scientific‐Nigel Shadbolt.” Interview by Jim Al‐Khalili. BBC Radio 4. 14 Apr. 2015. Radio. 7 McCarthy, John. "What Is Artificial Intelligence?" Computer Science Department Stanford University, 28 Sept. 2001, Revised 12 Nov. 2007; Thomason, Richmond. "Logic and Artificial Intelligence." Department of Philosophy University of Michigan, 27 Aug. 2003.
Artificial intelligence (“AI”), a term first coined in1956, is a branch of computer sciencethataimstocreateintelligentmachinesthatworkandreactlikehumans.2Atthebeginningof the study of AI, British mathematician Alan Turing proposed the true indication ofcomputerintelligencetobewhenaquestion‐askercouldnotdistinguishbetweenanswersfrom a human and those from a computer.3 In contrast, today, 60 years later, AI ischaracterized by a number of applications, including computers playing games againsthumans and understanding human languages, virtual personal assistants, and roboticswhichinvolvecomputersseeing,hearing,andreactingtosensorystimuli.4Lookingforwardto thenextdecade, technologistshaveofferedawidearrayofpredictions forAI, rangingfromAIbeingusedasatooltoaidrelativelysimpleprocesses(whichsomerefertoasweakAI)torobotswithhuman‐likementalcapabilities(whichissometimesreferredtoasstrongAI).5 According to AI expert Sir Nigel Shadbolt: “What we really have in AI is a wholespectrum of abilities, from programs that are smart but they are not smart like us, toprogramsthataresupercleverinspecificareas.”6For the purposes of our study,we consider AI to broadly be computational devices andsystemsmade to act in amanner that can be deemed intelligent.7 In otherwords, AI istechnology that appears to emulate human performance by learning, coming to its ownconclusions, understanding complex content, engaging in dialog with people, enhancinghuman cognitive performance, or replacing humans in executing both routine and non‐routinetasks.BecauseoftheuncertaintyofthefuturedevelopmentanddiffusionofAI,thisviewisintentionallybroad,coveringtheAIexistingtoday,suchastargetedadvertisingandvirtualpersonalassistants,aswellastheAIthatmayexistinthefuture,suchasrobotswithhuman‐likeprocessingcapabilities.TherangeofAI’sprogressinthefuturewilldetermine
the economic impact of AI on the global economy, with more limited advances andapplications (i.e., weak AI only) corresponding to more limited economic impacts, andmore substantial progress (i.e., strong AI) corresponding to more significant economicimpacts.This study estimates the projected global economic impacts associated with the use,development,andadoptionofAIoverthenexttenyears,andfindsareasonablerangeofAI’seconomicimpactoverthenext10yearstobebetween$1.49trillionand$2.95trillion.During this timeperiod,AI ispredicted tohavewide‐rangingapplications, including,butnotlimitedto:
Machinelearningthatautomatesanalyticalmodelbuildingbyusingalgorithmsthatallow machines to operate without human assistance.8 Potential applicationsinclude predicting cause‐and‐effect relationships from biological data, identifyingnewdrugs,9self‐drivingcars,10andprotectingagainstfraud.11
Virtual personal assistants that help users by providing reminders, schedulingappointments,organizing theirpersonal finances,and findingprovidersofvariousservices.14
Machine vision that allows computers to identify objects, scenes and activities inimages. Current applications of machine vision include providing objectdescriptionsfortheblind,15realisticfacialreconstructions,16car‐safetysystemsthatdetectpedestriansandbicyclists,17andstreet‐viewmaps.18
WeexpecttheeconomiceffectsofAItoincludebothdirectGDPgrowthfromsectorsthatdevelop or manufacture AI technology, and indirect GDP growth through increasedproductivity in existing sectors that employ some form of AI. Growth in AI producingsectorscould leadto increasedrevenues,andemploymentwithintheseexisting firms,aswellasthepotentialcreationofentirelyneweconomicactivity.Productivityimprovementsin existing sectors could be realized through faster and more efficient processes anddecisionmakingaswellasincreasedknowledgeandaccesstoinformation.The extent of these economic gains will be driven in large part by the rate of theadvancement and diffusion of AI. If AI is an increasingly critical component of moreproducts, itwillbecomean integralpartofmanypeople’s lives.AsourdefinitionofAI isintentionally broad, our definition of the diffusion ofAI is necessarily broad aswell.WegenerallyconceiveofthelevelofAIdiffusionastheportionofanindividual’sactivitiesinanaveragedaythatareproducedorshapedAIinanymanifestation.AsAIisincreasinglyincorporatedintomoreapplications,orAIisreliedonmoreheavilyinexistingapplications,AIwillbecomeamoreintegralpartofdailylife,increasingAI’slevelofdiffusion.TheextentofAI’seconomiceffectisalsolikelytovaryfromregiontoregion,thoughthisvariationmaybemoredependenton thepredominateeconomicactivityofa region,andthusAI’sabilitytoinfluenceeconomicactivity,ratherthantheeconomicordevelopmentalstatusof theregion. Infact,withthecurrentmovetowardsaccessibilityandopensourcedevelopment, AI has the potential to transcend income classes and to bring significantgains to both developed and developing countries. For example, AI has the potential tooptimize food production around the world by analyzing agricultural regions andidentifyingwhatisnecessarytoimprovecropyields.Intotal,thebroaderanddeepertheapplicationsofAI inagivenregionoreconomicsector, thegreatereconomic impact it isexpectedtohave.In estimating the future economic effects associatedwith an innovation such as AI, it isimportanttonotethat it ischallengingtoaccuratelypredictwhichapplicationsofAIwillultimately be commercially successful. Further, even if one could predict successfulcommercial applications, it is difficult to predict towhat extentAIwill be adopted in itssuccessfulapplications,theprecisewaysinwhichtheywillbedeployed,andtheresultingeconomiceffects.Given these challenges, calculating a single accurate estimate of AI’s economic effect isdifficult (if not impossible). Instead, we utilize several approaches to construct areasonable range of estimates of the potential economic effects associated with AI,concluding that this reasonable range is between $1.49 trillion to $2.95 trillion over thenexttenyears.However,andaswediscussinmoredetailbelow,ifAIisultimatelynotassuccessful as someare currentlypredicting,or ifAIdevelopsasquicklyand is aswidelyadoptedasitsstrongestproponentssuggest,theeconomicimpactscouldbeeithersmalleror larger than our initial estimated range of $1.49 trillion to $2.95 trillion. Our analysisaccounts for the uncertainty associated with AI’s economic impacts and yields results 18Miller,Greg."TheHuge,UnseenOperationBehindtheAccuracyofGoogleMaps."Wired.com,12Aug.2014.
4
whichweconsidersupplementallowerandupperboundestimatestoourrange.Ourfirstapproach for estimating the economic impact of AI is a “bottom up” approach, which isdescribed in the paragraph immediately below, and provides a lower bound estimate of$359.6billionforAI’seconomiceffectifAIultimatelydoesnotdevelopattheratesomearecurrently predicting. Our second approach, as described in the subsequent paragraph, ismore of a “top down” approach. This approach focuses on the global impacts of pasttechnologicalinnovationsaspotentialbenchmarksforAIandprovidesbothourreasonablerange of $1.49 trillion to $2.95 trillion in economic impact as well as an upper boundestimate, should the adoption and development of AI follow the more optimisticpredictions,of$5.89 trillion.Asdetailed in the followingsections,a carefulevaluationofthese methodologies leads us to conclude that our bottom up approach is likely tooconservative, while some of our benchmarks in the top down approach are likely toooptimistic. Therefore, we conclude that a reasonable range for AI’s economic impact isunlikely to include either one of these extremes, but may instead range between $1.49trillionand$2.95trillionoverthenextdecade.Inourfirstbottomupapproachweapplymethodologiestoestimatetheeconomiceffectsof investments in firmsdevelopingAI technologysince investment levels ina technologyare a telling sign of the future potential of that technology. We use two measures ofinvestment:privatesectorandventurecapital.Forourprivatesectorinvestmentanalysis,we rely on historical investment in AI to forecast future investment and economicliterature toestimate theeconomiceffectsof these investments.Using thisapproach,weestimatethatAIwillleadtoanincreaseofbetween$296.5billionand$657.7billionintheGDPofhigh‐incomecountriesinthenexttenyears.Underasimilarmethodologyfocusingonventurecapital investment,weestimatethatAIwill leadtobetween$63.1billionand$115.5billioninGDPofhigh‐incomecountriesinthenexttenyears.19Acknowledgingthatthere may be some overlap between these two approaches, we estimate that the totaleconomicimpactofinvestmentsbythesetwosectorsalone(i.e.,notincludingotherformsinvestmentinAIsuchascapitalinvestment)wouldimply$359.6billionto$773.2billionineconomicgrowthoverthenexttenyears.Weconclude,however,thatthesetwosectorsareunlikely to sufficiently represent the future potential ofAI.We therefore turn to the topdownapproachtoallowforamorecomprehensiveanalysisofAI’spotential.Ourtopdownapproachappliesthehistoryandestimatedimpactsofpriortechnologiesasbenchmarks forhowAI’sdevelopmentanddiffusionmayaffect theglobaleconomyoverthenexttenyears.AIhasthepotentialtoaffectbusinessacrosstheglobeinawiderangeofindustriesinwaysonlyanumberoftechnologieshavedoneinthepast.Forexample,AIisexpected to be a useful tool for enhancing human capabilities and in some instancesreplacing functions such as driving a car. Similarly, past technological innovations andinvestmentssuchas the investment in informationtechnology(IT), thedevelopmentandadoptionofbroadband internet, thedevelopmentandadoptionofmobile telephony,andindustrial robotic automation have served to enhance human capabilities and in somecases, replace humans. In this approach,we rely on academic research on the economicimpact of IT investment, broadband internet, mobile phones, and industrial robotics to
establish benchmarks for the potential impacts of AI. Themost reasonable benchmarkssuggestaboosttoglobaleconomicoutputofbetween$1.49trillionand$2.95trillion.Giventhe potential range of AI’s development over the next decade, we also include thepossibilityofmoreoptimisticscenarioswhichresultinanupperboundestimateashighas$5.89trillionoverthenexttenyears.This report focuses on the potential net economic effects of AI and not on the specificmechanisms that lead to economic outcomes. While AI is likely to affect both theproductivity and employment components of economic growth inmany sectors, parsingtheseeffects independently isbeyondthescopeof thisanalysis.Significantpublicdebatehas focusedonprojectionsofAI’s effect on the labor force, however. For instance, someresearchers have argued that the rise of AI and automation will lead to significantunemploymentascapitalissubstitutedforlabor.Indoingso,theypointtotheconcernthatthe increasing sophistication of AImay jeopardize skilled and semi‐skilled workers andreducethesizeofthemiddleclass.20Thisisnotanewargument,asthefearoftechnologynegatively affecting the labor force and leading to mass unemployment has beenarticulatedasearlyastheIndustrialRevolutionwheneconomistDavidRicardowrotethatthe“substitutionofmachineryforhuman labor, isoftenvery injurioustothe interestsofthe class of laborers.”21 The alternative view, however, is supported by the economichistoryofprevious “disruptive” technologies.Although employment in certain industrieshas been reduced in the past due to technological advancements, the net effect oftechnological advancement has not appeared to lead to a reduction in long‐term totalemployment.22 To date, the labor market has adapted to the introduction of newtechnologies, giving rise to new jobs in new areas. If the labor market continues todemonstrateitshistoricalresilience,thentheadvancementofAImayalsobeaccomplishedwithoutareductiontototal‐employmentinthelong‐run.Thispaperproceedsas follows. Section IIprovidesadditionaldetail on themethodologybehindourfirst,bottom‐upapproach,reviewingtheacademicliteratureandavailabledataon current investment levels in AI. Section III then discusses our second, top‐downapproachtakingeachofthefourbenchmarktechnologiesinturn,reviewingtheacademicliteratureoneach,andweighingtheirrelativestrengthsandweaknessesasanappropriatebenchmark for AI. Section IV then summarizes these findings and concludes with theestimatedreasonablerangeofeconomicimpactforAIoverthenexttenyears.II. BenchmarkingAIusingSectorInvestments
A prerequisite for the widespread adoption of any technology is the large amount ofresearch,development,andinvestmentthat it takestobringthattechnologytomarket.23Therefore, the high levels of current investment by industry players and venture capital 20Frey,CarlBenediktandMichaelA.Osborne.“TheFutureofEmployment:HowSusceptibleareJobstoComputerisation.”OxfordMartinSchool,September17,2013.21Ricardo,David.“OnthePrinciplesofPoliticalEconomyandTaxation”JohnMurray,3rd.ed.,Chapter31.3.22Atkinson,Robert.“StopSayingRobotsareDestroyingJobs‐TheyAren’t.”TechnologyReview,Sept.3,2013.23Forexample,BloombergreportsthatApplespent$2.7billiondollarstobringthefirstiPhonetomarket.“HowMuchDidAppleSpendonR&DfortheiPhone?”Bloomberg,Feb.7,2016.
6
firmsinAIprovideatellingsignofthecurrentstateof,andimpendingadvancesin,AI.In2014and2015alone,eightmajorglobaltechfirmsmadeatleast26acquisitions(totalingover$5billion)ofcompaniesdevelopingAItechnology.24PrivateinvestmentinAIhasalsotaken the form of in‐house spending in addition to startup acquisitions. For example,Facebook’s AI Research lab, Google’s Machine Intelligence lab, and Microsoft’s MachineLearning and Artificial Intelligence research division are all making advances in AItechnology and investing in the industry’s top talent.25 Additionally, between 2010 and2015,nearly $5billion inventure capital fundingwas invested in firmsacross theglobedevelopingandemployingAItechnology.26Below,weuseamethodologybasedinacademicliteraturetoestimatethepotentialglobaleconomicimpactsofprivateindustryandventurecapitalinvestmentinAI.Indoingso,wenote thatwe have thoroughly reviewed the limited economic literature surrounding themacroeconomic effects of R&D and venture capital investment. Based on the availableresearch,we estimate that industry investment in AI over the next ten years could leaddirectlytoaneteconomicimpactof$296.5billionto$657.7billion,whileventurecapitalinvestment could lead to an additional net economic impact of $63.1 billion to $115.5billionoverthenexttenyears.Takentogether,weestimatethatthetotaleconomicimpactofinvestmentsbythesetwosectorsalonecouldleadtobetween$359.6billionand$773.2billionineconomicgrowthoverthenexttenyears.TheseestimatesreflecttheimpactsofAIduetoinvestmentbyonlyaportionoftheglobaleconomy,excludinginvestmentsandresearchconductedbygovernmentandhighereducation, forexample.Nonetheless, theyprovideasignalofthepotentialofAItohavelargeglobalimpactsoverthenexttenyears.ImpliedImpactofPrivateIndustryInvestmentAsnotedabove,privateinvestmentinAIhasbeengrowingrapidlyinrecentyears.Romainand van Pottelsberghe (2004) provide a methodology that we use to estimate the neteconomic impacts of that investment.27 The researchers find that private R&D, venturecapital, and public R&D investment all have strongnet effects on economic growthwithventurecapitalfundinghavingthestrongestsucheffect.Theresearchershypothesizethatventure capital investment contributes to economic growth through innovation and bybolsteringthecapacityofaneconomytouseexistingknowledgetoincreaseproductivity.Using a panel regression framework to analyze historical data from 16 OECD countries,theyestimatetheimpactsofventurecapital,businessR&D,andpublicR&Dintensitiesonmulti factor productivity. The paper finds that the elasticities of output in relation to 24TheeightfirmsincludedareGoogle,Microsoft,Apple,Amazon,IBM,Yahoo,Facebook,andTwitter.2014and2015acquisitionsspannedservicesincludingspeechandimagerecognition,healthcareanalytics,homeautomation,datasecurity,cognitivecomputing,andmachinelearning.$5billionisaconservativeestimateasdealvalueswerenotpublishedformorethanhalfoftheseacquisitions.25FacebookAIResearchavailableathttps://research.facebook.com/ai;ResearchatGoogle–MachineIntelligenceavailableathttp://research.google.com/pubs/MachineIntelligence.html;MicrosoftResearch–MachineLearningandArtificialIntelligenceavailableathttp://research.microsoft.com/en‐us/research‐areas/machine‐learning‐ai.aspx.26AccordingtoThomsonOnePrivateEquityscreenersearchingforfirmswiththekeywords‘ArtificialIntelligence,’‘MachineLearning,’‘NaturalLanguageProcessing,’‘Self‐Driving,’and‘ImageRecognition’intheirbusinessdescriptions.27Romain,Astrid,andBrunoVanPottelsberghe."TheEconomicImpactofVentureCapital."UniversitéLibreDeBruxelles,SolvayBusinessSchool,CentreEmileBernheim,Apr.2004.
7
business R&D, venture capital, and public R&D intensities are 19.9%, 0.9%, and 13.6%,respectively.Theythencalculatethattheeffectonoutput,oneyearlater,ofonedollarofbusiness R&D, venture capital, and public R&D spending to be $1.99, $3.33, and $2.69,respectively.28Thesefindingsrepresentnetimpactsontheeconomy,takingintoaccountthesumofanypositiveandnegativeimpactsthefundingorinvestmentmighthave.Forexample,imagineaventurecapitalfirminvestsinacompanythathelpsfarmsthatproduceacertaincroptoconverttotheproductionofamorelucrativecrop.Inthisexamplethelossoftheoutputofthe less lucrative crophas anegativeeconomic impact, but thegainof theoutputof themore lucrative crop has a greater positive effect on the economy representing a netpositive impact on economic output of the venture capital funding that led to the cropswitch. These impacts are not realized solely by the firms that receive venture capitalinvestments, but could also take the form of output growth attributable to widespreadadoptionofanewtechnologydevelopedbyaventurefundedfirm.Thisbroad,netimpactmultipliermethodology isespeciallyappropriate forAIbecauseof the largepotential forspillovereffectsofinvestmentsinthetechnologyasitgainstraction.AccordingtoindustryanalystsattheInstitutefortheFuture,privateinvestmentinAIhasgrownfrom$1.7billionin2010to$14.9billionin2014.29Toestimatetheglobaleconomicimpactsof this rapidlygrowingprivate investment inAI technology,weuseRomainandvan Pottelsberghe’s 2004 finding that the marginal impact of one dollar invested inbusiness R&D is a $1.99 increase in output.30 Specifically, we assume two differentscenarios about the trajectory of private investment inAI over the next ten years. First,taking the conservative assumption that the amountofprivate industry investmentoverthenexttenyearsremainsconstantatthe2014levelof$14.9billionperyear,thenusingRomain and van Pottelsberghe’s findings, we estimate those investments would lead toapproximately$296.5billionofeconomicgrowthoverthenexttenyears.31Alternatively,ifwe take the more optimistic view that private industry investment in AI technologyincreasesatalinearrate32equaltotheaverageincreaseinthelevelofinvestmentbetween 28ThefindingsofKortumandLerner(2000)providerobustnesssupportforthisfindingindicatingthatventurecapitalisamorepotentimpetusforpatentcreationthanR&Dspending.KortumandLerner.“AssessingtheContributionofVentureCapitaltoInnovation.”RandJournalofEconomics,Winter2000.29Trabulsi,Andrew.“TheFutureofArtificialIntelligence.”InstitutefortheFuture,viaQuid,June2015.ThesefiguresarenotlimitedtoprivateR&DspendinginAI,butaimtoincludeallprivatesectorinvestmentinAItechnology,notincludingmergersandacquisitions.30RomainandvanPottelsberghe’sstudyspecificallyestimatestheneteconomicimpactsofprivateR&Dspending.ThelevelsofprivateinvestmentwehavecitedabovearenotstrictlylimitedtoR&Dspending,butcouldalsoincludecapitalexpendituresorotherinvestments.Therefore,ourestimatesoftheneteconomicimpactsofprivateinvestmentAItechnologyoverthenexttenyearsrelyupontheassumptionthatallprivateinvestmentwilleffectsimilarneteconomicimpactsasprivateR&Dspendingspecifically.(Romain,Astrid,andBrunoVanPottelsberghe."TheEconomicImpactofVentureCapital."UniversitéLibreDeBruxelles,SolvayBusinessSchool,CentreEmileBernheim,Apr.2004.)31Assumingthe$14.9billionofindustryinvestmentinAItechnologyobservedin2014continuesthrough2024.RomainandvanPottelsberghefindthatthemarginalimpactofonedollarinvestedinbusinessR&Dinyeartisa$1.99increaseineconomicoutputinyeart+1.Calculatedas($14.9billioninvested*1.99marginalimpactofadollarinvestedinbusinessR&D)*10yearsfrom2015‐2024=$296.5billion.32ReferencingR&DdataofU.S.firmscompiledbytheNationalScienceFoundationaspartofitsannualScienceandEngineeringIndicators,weobservelineartypeincreasesinR&DspendingbypublicU.S.firms
8
2010and2014(anincreaseof$3.3billionperyear)forthenexttenyears,thenbasedonRomainandvanPottelsberghe’sfindings,andassumingthatprivateindustryinvestmentinAI specifically results in the same increase in economic output as general private sectorR&D,weestimatethatprivateinvestmentinAIwillleadtoapproximately$657.7billionineconomicgrowthoverthenexttenyears.33ImpliedImpactofVentureCapitalInvestmentAsanalternativetoestimatingtheeconomicimpactofAIbywayofprivateinvestment,wecanalso examine levelsof venture capital investment in the sector.Within the academicliterature, there is general consensus that increases in venture capital funding lead toincreases in macroeconomic indicators such as the number of firms, employment, andwages. The theory behind this line of research is that venture capitalists contribute toeconomic growth in several important ways. First, venture funding spurs innovation.34Second, venture capitalists alleviate capital constraints, thereby allowing individualsotherwise lacking sufficient capital to engage in entrepreneurship. Third, venturecapitalists can encourage further entrepreneurship simply by being active in themarketandincreasingentrepreneurs’expectationsofreceivingfunding.Finally,venturecapitalistscan affect the economy by increasing the probability of spinoff companies as venturefundedfirmsinspireoutsiderstobegintheirowncompanies,andpotentiallypreparetheiremployees to start their own firmsby exposing them to the rigors of running a start‐upfirm.35Forexample,inastudyofthe329metropolitanstatisticalareasintheU.S.,Samilaand Sorensen (2011) found that doubling the number of firms receiving venture capitalfundingacrossallindustriesinaregionincreasedthenumberoffirmsbybetween0.48%and 2.21%, increased the number of jobs in that area by up to 1.24%,36 and increasedaggregateincomebybetween0.48%and3.78%fiveyearsfollowingtheinvestment.37 from1994‐2000ondevelopingtechnologies(similartoAI’scurrentstate)likemobilephonesandcomputerstoragedevices.Alternatively,duringthesametimeframe,weobserveflatR&Dspendinginmatureandstableindustriessuchaschemicalmanufactureandaerospace.ThesehistoricalobservationsmakeflatR&DspendinginAIoverthenexttenyearsaconservativeassumptionasthefieldisstilldeveloping.“U.S.andInternationalResearchandDevelopment:FundsandTechnologyLinkages–NationalR&DTrends”NationalScienceBoardScienceandEngineeringIndicators2004,May2004atChapter4,Appendixtable4‐22.33From2010through2014,theamountofventurecapitalinvestmentinfirmsdevelopingAItechnologyincreasedattheaveragelinearrateof$3.3billionperyear,calculatedas($14.9billioninvestedin2014‐$1.7billioninvestedin2010)/(4yearsfrom2010–2014).Letusassumethe$14.9billionofprivateinvestmentobservedin2014increasesatthisrateof$3.3billioneachyearthrough2024,RomainandvanPottelsberghefindthatthemarginalimpactofonedollarofbusinessR&Dinyeartisa$1.99increaseineconomicoutputinyeart+1.ThetotaleconomicimpactofventurecapitalinvestmentinfirmsdevelopingAItechnologyoverthenexttenyearsiscalculatedasthesumofprojectedfutureinvestmentsinthesectorincreasingatthelinearrateof$3.3billionperyearovertheyears2015–2024multipliedby$1.99ingrowthforeachdollarinvested,yieldinganestimatedtotaleconomicimpactof$657.7billion.34ThefindingsofKortumandLerner(2000)providerobustnesssupportforthisfindingindicatingthatventurecapitalisamorepotentimpetusforpatentcreationthanR&Dspending.35Samila,Sampsa,andOlavSorenson."VentureCapital,EntrepreneurshipandEconomicGrowth."TheReviewofEconomicsandStatistics,February2011.36ItisimportanttonoteherethatVCfundingforfirmsdevelopingAItechnologyhasthepotentialtoinduceanetnegativeemploymentimpactasAIoftentakesontasksthatotherwisewouldbeconductedbyahuman,potentiallydisplacingworkers.Todate,noacademicstudieshavebeenconductedestimatingthebroadereconomiceffectsofVCinvestmentinAIspecifically.37Samila,Sampsa,andOlavSorenson."VentureCapital,EntrepreneurshipandEconomicGrowth."TheReviewofEconomicsandStatistics,February2011.
9
LikeprivateindustryinvestmentinAI,anincreasinglevelofventurecapitalinvestmentinAIisanindicatorofthefuturepotentialofthetechnology.Table1below,assembledfromthe ThomsonONE private equity database, shows the amount of venture capital fundingthat has been invested since 2010 across the globe in firms that are employing anddeveloping AI technology.38 In compiling this estimate, we identified this set of firmsreceivingventurecapitalfundingbysearchingfirmdescriptionsforthetargetedkeywordsinthetable.ThesekeywordsarerepresentativeofthebranchesofAIthatareconsideredby technologists as likely to have an effect on the global economy in thenear future.Asevident inTable1, theamountof venture capital funding flowing to firmsdevelopingAItechnology has increased dramatically since 2010, with the largest annual amount, $1.9billion, being invested in2015.AI’s relative shareof venture capital investmenthas alsoincreased substantially; according to Ernst andYoung, global venture capital investmenttotaled $46.6 billion in 2010 and had nearly doubled to $86.7 billion in 2014.39 Bycomparison,duringthattimeventurecapitalinvestmentinfirmsdevelopingAItechnologyhasincreasedovereighttimes.
SimilartoourmethodologyforestimatingtheeconomicimpactofprivatespendingonAI,we use the results fromRomain and vanPottelsberghe (2004) to estimate the potentialglobaleconomicimpactofventurecapital fundingforfirmsdevelopingandemployingAItechnology.Using the samemethodology of projectingprivate investment inAI over thenexttenyears,weestimatetheglobaleconomiceffectsofventurecapitalinvestmentinAI.
TheRomainandvanPottelsberghe(2004)studyestimatesthatthemarginalimpactofonedollarinvestedbyventurecapitalistsisa$3.33increaseineconomicoutput.Assumingthattheamountofventurecapitalinvestmentoverthenexttenyearsremainsconstantatthe2015 level of $1.9 billion per year, then using Romain and van Pottelsberghe’s (2004)findings, we estimate those investments would lead to approximately $63.1 billion ofeconomic growthover thenext ten years.40Alternatively, ifwe take themoreoptimisticassumptionthatventurecapitalinvestmentinfirmsdevelopingAItechnologyincreasesata linearrate41equaltotheaverageincreaseinthelevelof investmentbetween2010and2015(anincreaseof$349.4millionperyear)forthenexttenyears,thenassumingthatVCinvestmentinAIspecificallyresultinthesameincreaseineconomicoutputasgeneralVCinvestment,RomainandvanPottelsberghe’s findingssuggest that investment inAI firmswillleadtoapproximately$115.5billionineconomicgrowthoverthenexttenyears.42AcceptingthatourestimationsoftheeconomicimpactofAIoverthenexttenyearsbasedon private sector and venture capital investments in the technology may have someoverlap,weestimate that thesumtotaleconomic impactofAIasadirect resultof thesetwosourcesofinvestmenttobebetween$359.6billionand$773.2billionoverthenexttenyears. In interpreting these estimates of the economic impact of AI, it is important toconsider that private industry and venture capital investment are only a portion of theglobal economy. Thus, our estimates of the economic impacts of these two sources ofinvestment inAI technologyrepresentonlyaportionof the totaleconomiccontributionsthatAIislikelytogenerate.Furthermore, the methodology used here to estimate the economic impact of AI onlyreflects the social return on investment dollars one year after the investment has beenmade.Itisverylikelythatadollarofinvestmentwillgeneratereturnsoveralongertimeperiodthanoneyearsuggesting largereconomic impacts thanwehavecalculatedabove.Nonetheless,currentlevelsofprivateindustryexpenditureandventurecapitalinvestmentinAIsendaveryclearmessageaboutAI;inparticular,AIisalreadyacontributingfactorof
Duringthelasthalfcentury,technologicalinnovationshavebeenlargedriversofgrowthineconomic output as well as in labor and capital productivity. Innovations such as thepersonalcomputer,broadbandinternet,mobilephones,andindustrialrobotshaveallhadmeasurable and significant effects on the global economy. Applications of thesetechnological innovations are wide‐ranging, and so are their economic benefits. Thesetechnological toolshavereducedcosts, increasedproductivity, expandedoutput, spurredfurther innovation, and given rise to new and significant sectors of the current globaleconomywhich,inturn,affectedtheallocationoflaborinthemoderneconomy.AIhasthepotentialtoaffecttheglobaleconomyinthesesameways.Asanalternativetothebottom‐upapproachtakeninSectionII,thissectionestimatesthepotentialeconomicimpactbyusingprevioustechnologicalinnovationswhoseeffectsmaybesimilartothoseAIwillhaveoverthenexttenyears.Specifically, inthefollowingsub‐sections,weexaminefourbenchmarkstoinformAI’spotentialeconomicimpact:generalITinvestment,broadbandinternet,mobilephones,andindustrialrobotics.Thesebenchmarkseachrepresentrecentandsignificanttechnologicalinnovationthatexperiencedrapidandextensive adoption, and had significant global economic effects through themechanismsdiscussedabove.Aspre‐cursorstoAI,thebenchmarktechnologiesareappropriateinpartbecausetheirmethodandrateofdiffusionmayberepresentativeofthefuturediffusionofAI.43A baseline of computing power, internet access, mobile phone penetration, and basicrobotic integration is likelynecessarytoallowthesubsequentadoptionofAItechnology,and the changes that industriesunderwentas theseprevious technologieswereadoptedand leveraged serve as a good indication of the industries’ ability to respond flexibly tofuture technological advances such as AI. Furthermore, the mechanisms through whichthesebenchmarks influenced theglobaleconomyaresimilar to themechanisms throughwhich AI is anticipated to affect productivity and growth. Therefore, our benchmarksrepresentarangeofeconomiceffectstoappropriatelyreflecttheuncertaintyofAI inthefuture.Inundertakingthisapproach,wesummarizetheacademicliteraturethathasestimatedthecontributionofeachtechnologytoeconomicgrowth,evaluatethesebenchmarks’relativestrengths and weaknesses, and use these studies to inform a potential range for the
economiceffectofAIthrough2025.AssummarizedinTable2anddiscussedinmoredetailbelow, these benchmarks imply a full range of economic effects forAI of between $1.49trillion and $5.89 trillion (0.2% to 1.0% of GDP), with themore likely estimates, whichexcludes the more highly optimistic scenarios relying instead on more conservativemethodologies and closer parallels between AI and the benchmark technology, rangingbetween$1.49trillionand$2.95trillion.
For each benchmark in the following sub‐sections, we incorporate the most relevantavailable academic literature on the topic, but note several limitations of this approach.First,whilewehave conducteda thorough reviewof researchon thebenchmarks, somebenchmarkshavebeenmore comprehensively studied thanothers.For instance, both ITinvestment and broadband have been well studied. As discussed in the following sub‐sections,while the resultsof the IT andbroadband studiesdonot all agreeon theexactmagnitudeoftheeconomiceffectgeneratedbythesetechnologies,thevolumeofresearchavailable on these topics does serve to generally confirm the presence and range of thepotential economic effects. In contrast, there are relatively fewer papers that attempt toassesstheeconomiceffectofroboticsormobilephones.Therefore,wehavelesscertaintyabout the range of economic effects created by robotics andmobile phones, and futureresearchmayrevealadifferentmagnitudethanthebasisforourcurrentestimates.Second, as is the case with any benchmarking exercise, benchmarks can serve asinformativeapproximationsbutarenot intendedtorepresentanidenticalscenariotoAI.AIisunlikelytomanifestidenticallytoanyparticularbenchmark.Ourintentionis,instead,to employ benchmarks that share certain characteristics of AI, and from these potential
similarities assess a range of effects that are possible for AI should AImanifest inwaysmorecomparabletooneoranotherbenchmark.
A. EconomicEffectsUsingInformationTechnologyInvestmentasaBenchmarkThe first benchmarkwe consider is the rapid investment in IT that occurred during the1990s and early 2000s. During these years, IT investment was largely composed ofinvestments in computing hardware and software which caused a dramatic increase incomputing prevalence in businesses.44 AI is already pushing electronics growth intomarketsnotpreviouslypenetratedbyadvancedcomputingwithinnovationssuchassmartappliancesandrobo‐advisors.45ThesecurrentachievementsareastrongindicationthatAIhasthepotentialtosubstantiallychangehowentiresectorsoftheeconomyusetechnology,justasITinvestmentdidduringthe1990sandearly2000s.Asabenchmark,ITalsohastheadvantage of several similarities with our functional definition of AI.For instance, ITinvestment resulted in a wide range of economic benefits including reduced costs,increased productivity, expanded output, and spurring further innovation, all of whichmirror AI’s potential.Furthermore, IT is a bundled term that refers to a wide array oftechnology, some of which were individually very substantial contributors to economicgrowth(suchaspersonalcomputers)whilethemajorityofthetechnologyencompassedinthe IT category were likely contributing to economic output at a much lower rate. TheeffectofthisblendisessentiallyaweightedaverageassessmentofthepotentialimpactsofIT.ThisblendmayappropriatelyreflectAI’srangeofapplications,assomearelikelytobesubstantialcontributorstooutputwhileothersmayhaveasmallereconomiceffect.The effects of IT on increasingproductivity andoutput havebeen studied extensively. Awidely‐citedmeta‐analysisconductedbyDedricketal.(2003)examinedover50papersinthe preceding decade and a half and concluded that “[a] number of major studies havedocumentedthesignificantimpactofITinvestmentontheproductivityoffirms,industries,andcountries,showingthatcomputersdo,infact,showupintheproductivitystatistics.”46Furthermore, theywrite: “IT investments actually have been increasing productivity for
morethanthreedecades.”47ITinvestmentshavebeentiedtosignificantgainsineconomicoutputinAsia,Europe,andNorthAmerica.48Forthepurposesofthisbenchmarkingexercise,wefocusonseveralmorerecentstudiesthat estimated IT’s net contribution to economic growth. For example, O’Mahony andTimmer(2009)estimatedthatfrom1995‐2005,investmentsinICTcontributedtoEUandU.S. annual output growth by 0.6 and 1.0 percentage points respectively.49 While otherstudiessupportthemagnitudeofthesefindings,someindicatethatgainsfromtechnologymayhaveslowed.Forexample,Olineretal.(2007)foundthatfrom1995‐2000,ITcapitalinvestmentscontributed1.09percentagepointstoannualU.S.productivitygrowth,butitscontributionsubsequentlydroppedto0.61percentagepointsfrom2000‐2006.50Similarly,Jorgensonetal.(2011)foundthatITcapitalwasresponsiblefor1.02percentagepointsofthe annual output growth experienced from 1995‐2000 in the U.S., and then only 0.49percentagepointsofannualoutputgrowthfrom2000‐2007.51Applying theresultsof this literature,weestimate thepotentialeconomiceffect fromAI,shouldAIperformasITdidduringthe1990sand2000s.WechoosetorelyontheresultsofO’MahonyandTimmer(2009)because(1)theirstudyhadabroaderscopebycoveringmore countries, (2) their results for theU.S.were consistentwith other studies, and (3)theiranalyticalapproachyieldedmoreconservativeestimates.52However,thestudydoesnotassessIT’simpactonemploymentorproductivityindependentlyfromIT’sneteffectoneconomicgrowth.53To calculate a single economic effect we blend O’Mahony and Timmer’s U.S. and EUestimates (weighted by GDP), yielding an estimated IT effect on annual GDP growth of0.79%.54 To convert this estimated annual contribution to GDP growth into dollars, we
combinedata fromtheConferenceBoardandtheWorldBanktoestimateGDP for2015‐2025.55Becausethebenchmarkpapersfocusonhigh‐income,developednations,wefocusontheimpliedimpactsofAIforhigh‐incomecountries.AsdiscussedinSectionI,AIislikelytogenerateeconomicbenefitsforbothdevelopinganddevelopednations.However,sincethebenchmarkpapersdonotincludedevelopingcountriesintheirestimatesweareunabletoestimateAI’spotentialeffectforthesecountriessoweconservativelyassumeAIhasnoimpact in these regions. Multiplying the projected GDP estimates through 2025 by the0.79%estimatedeconomiccontributionofIT,wecalculatethatifAIperformsequivalentlytohistorical IT, thenAIcouldhaveacumulativeeconomiceffectof$4.78trillionthrough2025,or0.8%ofGDP.56WhiletherearetechnologistswhobelievethatAIislikelytomimicpastITrevolutionsanddrive a corresponding increase in IT investment such as the investment analyzed in thebenchmarkstudies,57aneconomiceffectduetoAIonthescaleofthe1990’sITexperienceis likely too optimistic, as this would imply AI singularly attaining the same level ofsignificanceas theentire ITsectorof the1990swithinthenext tenyears. Inaddition,asindicatedbytheacademicliterature,thecontributionofITmayhavedeclinedovertimeasthegainsfromITdecreased.GiventhislikelydeclineinIT’seffect,thedifferenceintimingbetweentheanticipatedadoptionofAIandthepreviousdiffusionofIT(2016‐2025versus1997‐2007) is likely to reduce the effects of investment below those documented in thebenchmarkstudies.Thusthe$4.78trillionineconomicimpactimpliedbythisbenchmarkis likely only reasonable if the development and diffusion of AI over the next decadematchesorexceedsexpectationsofAI’sstrongestproponents.
B. EconomicEffectsUsingBroadbandInternetasaBenchmarkThe second benchmark we consider is broadband internet which realized a rapid andpervasivediffusioninthelate1990sandearly2000s.Asabenchmark,broadbandhastheadvantage of several similarities with our functional definition of AI. In particular,broadbandinternetincreasedthespeedandavailabilityofinformation,resultinginawiderange of economic benefits including reduced costs, increased productivity, expandedoutput, and spurring further innovation, all ofwhichmirrorAI’spotential.Relative to ITinvestment,broadbandinternethastheadvantageofbeingonlyacomponentof ITwhilestillbeingbroadlyapplicabletoessentiallyallsectorsoftheeconomy.ThisislikelyaclosercorollaryforAIwithinthenexttenyears,asAIis likelytobebroadlyapplicabletomanyeconomic sectors, but less likely to grow rapidly enough in the next decade to rival themagnitudeof the full ITsector. Inaddition, theeconomiceffectsofbroadbandhavebeenrelativelywellstudiedwithintheacademicliterature.
Studies of the effects of broadband penetration have found that in a wide range ofeconomies, a modest increase in broadband penetration can have significant effects onGDP.TheexactmagnitudeoftheimpactonGDP,however,variesacrossstudies.Atthehighend,Czernich et al. (2011) found that for25OECD countriesbetween1996and2007, a10%increaseinbroadbandpenetrationincreasedannualper‐capitaGDPgrowthby0.9‐1.5percentagepoints.58Weusethemoreconservative,lowerboundofthisestimatedrangetoinformourbenchmark.Similarly,QiangandRossotto(2009)foundthat from1980‐2006,high‐incomeeconomiesrealizedincreasesinaverageannualper‐capitaGDPgrowthof1.21percentage points per 10% increase in broadband penetration. However, for low andmiddle income economies Qiang and Rossotto (2009) found that broadband was aninsignificant determinant of growth. Alternatively, more conservative estimates such asKoutroumpis (2009) found that for 22 OECD countries between 2002‐2007, a 10%increase in broadband penetration increased average annual economic growth by 0.25percentage points.59 Similarly to Qiang and Rossotto (2009), Koutroumpis’ results alsoindicate that level and speed of broadbanddiffusions effect the returns on growth, suchthat broadband would have a smaller effect on countries with relatively low levels ofbroadbandpenetrationordiffusion.AswiththeITinvestmentbenchmark,thebenchmarkstudies on broadband do not explicitly address broadband’s impact on employment orproductivity(i.e.,thesecomponentsofeconomicimpactarenotmeasuredinisolation,onlyreflectedascomponentsoftheneteffectofeconomicoutput)sothesebenchmarksserveonlytoanchorneteconomiceffectsnotmorespecificemploymenteffects.Applying the results of this literature, we estimate a potential economic effect from AI,shouldAIperformasbroadband internetdidduring the1990sand2000s.Todo so,wefirstrequireanapproximationofAI’spotentialdiffusionoverthenextdecade.Broadbandinternet penetration increased rapidly during the decade 1997‐2007 covered by thisliterature,risingfromapproximately0%to20%intheOECDcountries.60Forthepurposesof thisstudyweare focusedonAIduring itsearlyyearsaswell. Importantly, the typicalrate of adoption for new technologies tends to exhibit an “S” curve shape, with slowincreases in prevalence in early years, followed by a rapid increase in diffusion speed,whichsubsequentlylevelsoutasthesaturationpointisneared.61Broadband’sexperienceinthe1990sand2000smatchesthisexpectedshapewell.WhileAIisunlikelytofollowanidenticalpatterntoanybenchmark,AIdevelopmentscouldfollowasimilardisseminationpathtobroadbandgiventheirpotentialsimilaritiesdiscussedabove.Therefore,weemploybroadband’sactualexperienceddiffusionratefrom1997‐2007asanestimateofthepathAImayfollowfrom2016‐2025.TheseresultsarepresentedbelowinTable3.Estimates of economic effect of broadband from the academic literature are presentedrelative to a 10% increase in broadband penetration. A 10% presence of AI is notapplicable for all years of our projection, therefore we use the ratio of the estimated 58Czernich,Nina,OliverFalck,TobiasKretschmer,andLudgerWoessmann."BroadbandInfrastructureandEconomicGrowth."TheEconomicJournal505‐530(2011).59Koutroumpis,Pantelis.“TheEconomicImpactofBroadbandonGrowth:ASimultaneousApproach.”TelecommunicationsPolicy,2009.471‐485.Print.60OECDHistoricalBroadbandPenetrationData.61Ineconomics,marketsaturationisusedtodescribethesituationwhereaproducthasbecomefullydiffused.
17
diffusionratesofAIover10%toadjusttheeconomiceffectofbroadbandasestimatedbythe above literature and summarized in Table 3. Finally, as with the IT benchmark, thebroadband benchmark papers focus on high‐income, developed nations. Therefore wefocus on the implied impacts of AI for high‐income countries as well, multiplying theadjusted annual contribution to economic growth by projected GDP estimates for thesecountries through2025.62Theresultingrange implies that ifAIperformsequivalently tohistorical broadband, then AI could have a cumulative economic effect between $1.49trillionand$5.89trillionthrough2025,oralternatively,between0.2%and1.0%ofGDPforhigh‐income countries during those same years.63 These calculations and results aresummarizedinTable3.
Ininterpretingtheseresults,itisimportanttounderstandwhatthebroadbandbenchmarkimpliesforAI.Today,after20years,broadbandhassignificantpenetrationinhigh‐incomecountries.AI,incontrastmaytakeamoregradualdiffusionpaththantherapidescalationexperiencebybroadband.Inaddition,broadbandissuchageneraltoolthatitwasquicklyapplicabletoessentiallyeverysectoroftheeconomy.ThoughAIcertainlyhasthepotentialtoaffectmanysectorsoftheeconomy,specificrealizationsofAIinthenexttenyearsmayhave more focused applications. Finally, the larger estimates found in the academicliteraturemayover‐attributethegrowththatoccurredduetobroadbandduringtheperiodanalyzed.Therewereseveralnearlysimultaneousinnovationsthatoccurredintheseyears,andthusitisverydifficulttoassesstheportionwhichbroadbandspecificallycontributed.Thisdifficultymayexplain,atleastinpart,thewiderangeofestimatesprovidedbythesebenchmark papers. For example, the highest implied estimates forAI’s economic impact
from the broadband internet benchmark (derived from Qiang and Rossotto (2009))exceeds even the estimate impliedby IT investment,while themiddle estimate (derivedfromCzernichetal.(2011))isnearlyequivalenttotheITbenchmark.Sinceitisourviewthat the IT benchmark is likely too optimistic, we consider these two broadbandbenchmarkestimatesasequallyoptimisticestimatesforAI’seconomicimpact.Incontrast,themoreconservativebroadbandbenchmark(derivedfromKoutroumpis(2009))impliesthat AI could have a cumulative economic effect of $1.49 trillion through 2025. As thisestimatereliesonabenchmarktechnologywhichoffersqualitativeimprovementsoverITinvestment as discussed above, and as the resulting estimate falls above our likely tooconservativeapproachinSectionII(yielding$359.6to$773.2billion)andbelowthelikelytoooptimisticITbenchmark(yielding$4.78trillion),wefindthisapotentiallyreasonablebenchmarkforAI’seconomicimpact.
C. EconomicEffectsofMobilePhonesasaBenchmarkThe thirdbenchmarkwe consider forAI ismobilephone technology.Mobilephones areanother recent ICT development that experienced a rapid diffusion rate. Similar to IT,broadband,andAI’spotential,mobilephonesareconsideredageneralpurposetechnologyas therearemanyavailablemechanisms for this technology toaffecteconomicgrowth.64Mobilephoneshavechangedthewayusersinteract,andincreasedusers’accessibilityandthe speed and reliability with which contact can be made and information shared. Byimproving communication, mobile phones have resulted in an increase in productivity,efficiency, and innovation in applications from improving government tax collection tofosteringnewtypesofservicesandbusinessstructures.Similarly,currentmanifestationsof AI are already working to increase our ability to collect, share, and communicateinformation efficiently with innovations like smart medical repositories and real‐timebusiness analytics. These current achievements are a strong indication that AI has thepotentialtosubstantiallyincreasethespeedandaccessibilityofinformationjustasmobilephones have done in the last decade. In addition, mobile phones share broadband’spotentialimprovementoverITasabenchmarkforAIbyrepresentingasubstantial,butnotallencompassingportionof IT.Finally,andincontrasttothepreviousITandbroadbandbenchmarkpapers,mobilephoneshavebeen found tohave adisproportionate effect ondeveloping countriesbecause significant infrastructure investment isnotnecessary for apopulation toeffectivelyemploymobilephone technology.Forexample,by2009,mobilephonesubscribersindevelopingcountriesrepresentednearly70%ofglobalmobilephonesubscriptions (3.2 out of 4.6 billion).65 As is already being exhibited with current AIdevelopments such asApple’s Siri orMicrosoft’s Cortana, it is likely that some futureAIimprovements will be accessed through mobile phone technology. Thus, if AI were tocontinuetoenterthemarketthroughmobilephonetechnology,itispossiblethatAIcouldhavesimilarlydifferentialeffectsondevelopingcountries.
Researchershavefoundthatmobilephoneshaveapositiveimpactoneconomicgrowthinboth developing and developed economies.66 Two studies examine the effect of mobiletelecommunicationtechnologyduring itsdevelopment fromthemid‐1990s intotheearly2000s.Gruber andKoutroumpis (2011) study192 countries from1990‐2007and find apositiveandstatistically significant impactofmobilephone linesongrowth,witha10%increase in the mobile penetration rate increasing annual average economic growth by0.33‐0.89percentagepoints.67Aswiththebroadbandbenchmarkstudies,weusethemoreconservative,lowerboundofthisestimatedrangetoinformourbenchmark.Similarly,Vu(2011) studied 102 countries from 1996‐2005 and finds that a 10 percentage pointincreaseinthemobilephonepenetrationrateincreaseseconomicgrowthbyatleast0.55percentagepoints.68Applying the results of this literature, we estimate a potential economic effect from AIshould AI perform as mobile phones did during the 1990s and 2000s. Similar to thebroadbandbenchmark,inordertoestimatetheeffectofAIwerequireanapproximationofAI’spotentialdiffusionoverthenextdecade.UsingdatafromtheWorldBank,weselecttheseconddecadeofmobilephonedata as representativeofmobilephone’s first significantperiod of adoption. During the decade from 1993‐2003, subscribers of mobile phonesincreasedfrom0.6%of thepopulationto22%.WhileAI isunlikely to followan identicalpattern toanybenchmark,AIdevelopments could followa similar disseminationpath tomobile phones given the potential similarities discussed above. Therefore, we employmobile phone’s actual experienced diffusion rate from 1993‐2003 as an estimate of thepathAImayfollowfrom2016‐2025.Applying this assumed diffusion pattern, we calculate that the finding of Gruber andKoutroumpis(2011)wouldyieldanannualaverageincreaseinGDPgrowthof0.31%overthenextdecade.Similarly,weestimatetheeconomiceffectofAIusingtheresultsfromVu(2011)byapplyingthisstudy’sregressionequationtothelevelofmobilephonediffusioneach year and calculating the percentage points of GDP growth attributable to mobilephones,indicatinganaverageof0.43%overthenextdecade.69Asbothbenchmarkstudiesincluded awide range of countrieswemultiply projected global GDP estimates through2025bytherangeof0.31%to0.43%fortheestimatedeconomiceffectofmobilephones.WecalculatethatifAIperformsequivalentlytohistoricalmobilephones,AIcouldhavea
cumulativeglobaleconomiceffectof$2.95trillion–$4.24trillionthrough2025,or0.31%to0.45%ofglobalGDPduringthatperiod.As with broadband, we acknowledge that there were several nearly simultaneousinnovationsthatoccurredintheseyears,andthusitisdifficulttoassesstheportionwhichmobile phones specifically contributed. The estimates of AI’s potential economic impactproduced by mobile phone benchmarking studies fall within the range of estimatesgenerated by the broadband benchmark, thus we evaluate them similarly. Again, thehighest implied estimates for AI’s economic impact from the mobile phone benchmark(derivedfromVu(2011))isnearlyequivalenttotheITbenchmark.Aswehavedeterminedthat theITbenchmark is likely toooptimistic,weconsiderthismobilephonebenchmarkestimateasequallyoptimisticaboutAI’seconomicimpact.Thatis,aneconomicimpactof$4.24trillionislikelyonlyreasonableifthedevelopmentanddiffusionofAIoverthenextdecadematchesorexceedsexpectationsofAI’sstrongestproponents.Incontrast,themoreconservative mobile phone benchmark (derived from Gruber and Koutroumpis (2011))impliesthatAIcouldhaveacumulativeeconomiceffectof$2.95trillionthrough2025.Asthis estimate relies on a benchmark technology which offers qualitative improvementsover IT investment as discussed above, improved scope over the broadband benchmark(byalsoconsideringtheeffectsondevelopingcountries),andastheresultingestimatefallsaboveourlikelytooconservativeapproachinSectionII(yielding$359.6to$773.2billion)andbelow the likely toooptimistic ITbenchmark (yielding$4.78 trillion),we find this apotentiallyreasonablebenchmarkforAI’seconomicimpact.
D. EconomicEffectsofRoboticAutomationasaBenchmarkThefinalbenchmarkweconsiderforAIisroboticautomation.Aswithbroadband,roboticstechnology shares several components of our definition of AI. Specifically, roboticautomationisatoolthatenabledproductivitygrowthjustasAI,inthewordsofSirNigelShadbolt,is“anaidtoaugmentourintelligence.It’smakingussmarterandquickeratwhatwe do.”70 Robotic automation is potentially a more appropriate benchmark than ITinvestment,broadbandinternet,ormobilephonesforseveralreasons.First,whilebothITinvestmentandbroadband internetwereeconomy‐widephenomena, robotics,while stillapplicabletoawidesectoroftheeconomy,ismorefocusedinafewsectors.ThisislikelyaclosercorollaryforAIwithinthenexttenyears,whenrealizationsofAIarelikelytohavemoretargeted,sector‐specific,applications.Second,relativetobroadband,roboticsmaybeconsideredamore“disruptive”technologythatleadstoreconfiguringsystemsinordertoleverage the technology’s benefits in an optimal way. AI also has the potential to causedisruptionasitmayintroducenewwaystoperformatask,notjustincreasingthespeedatwhichthepreviousmethodcanbeperformed.Data,andthereforeacademicresearch,onroboticsarelimited.However,arecentpaperbyGraetz andMichaels (2015) examines the effects of increased use of robotics across 14industries in 17 developed countries from 1993‐2007.71 They find that the adoption of
robotictechnologiesoverthisperiodincreasedannualaverageGDPgrowthbyabout0.37percentage points and improved labor productivity by 0.36 percentage points.72 ThisaccountedforonetenthoftheaggregateGDPgrowthduringthisperiodforthecountriesexamined, even though industrial robotics accounted for only roughly 2.25% of capitalstocks inrobotic‐using industriesasof2007andmuch less intheearlieryearsanalyzed.GraetzandMichaelsalsoconductseveralsensitivitiesanddeterminethattheirresultsaregenerally robust to changes in capital and labor (including both employment andproductivity effects) which might be induced by an increased use of robots in theseindustries.Applyingtheresultsofthisstudy,weestimateapotentialeconomiceffectfromAI,shouldAI perform as industrial robotics did during the 1990s and 2000s. This estimate iscalculatedusingasimilarapproachtotheITinvestmentbenchmark.AswiththeprecedingITandbroadbandbenchmarks,thebenchmarkstudyonroboticsfocusedontheeffectsofroboticsinrelativelyhigh‐incomecountries.Thus,multiplyingtheprojectedGDPestimatesfor high‐income countries through 2025 by the 0.37% estimated economic impact ofindustrialrobotics,wecalculatethatifAIperformsequivalentlytohistoricalrobotics,thenAIcouldhaveacumulativeeconomiceffectof$2.23trillionthrough2025,or0.4%ofGDPduringthatperiod.73Asdiscussedabove,roboticautomationasabenchmarkispotentiallypreferabletoanyofIT investment, broadband, or mobile phones because robotics is likely to more closelyreflectAI’smoresector‐targetedeffects,andpotentiallydisruptivenature.Inaddition,theresults associatedwith applying industrial robots to AI yield estimates of AI’s economiceffectwhich fall between previous benchmarks thatwe assess as likely under‐estimates(i.e.venturecapitalandprivateinvestment)andbelowestimateswhichweassessaslikelyover‐estimates(i.e.ITinvestmentandtheupperboundsimpliedbybroadbandandmobilephones). We therefore find this a potentially reasonable benchmark for AI’s economicimpact.
IV. ConclusionOur study of the economic impacts of AI during the next decade relies on twomethodologies. First, a narrow and conservative analysis of private sector and venturecapital investment indicates that the economic impact of AI could be between $359.6billion and $773.2 billion over the next ten years. At a minimum, the current levels ofinvestmentinAIareasignaloftheeconomicpotentialofAI.However,givenlimitationsofthe available data on current investment levels in AI and other potential mechanismsbeyondreturnoninvestmentthroughwhichAImayalsoaffecteconomicgrowthwhichare
not captured by this approach, we conclude that this approach likely yields an underestimateofAI’seconomiceffectsoverthenexttenyears.Second, a set of benchmarks of recent and significant technologies that share similarcharacteristics with AI including general IT investment, broadband internet, mobilephones,andindustrialrobotics,providesausefulframeworkforestimatingAI’spotentialeconomiceffects.ThemostreasonablebenchmarkssuggestthattheeconomicimpactofAIcouldbebetween$1.49trillionand$2.95trillionoverthenexttenyears.Theseestimatesarederivedfromthemoreconservativestudiesofbroadbandinternet,mobilephones,andthestudyofindustrialrobotics.Atahighlevelthisrangeisconceptuallyreasonable.Itfallswell above our first, investment driven approach which we conclude is a likely under‐estimate,asnotedabove.ItalsofallswellbelowtheITbenchmark(implying$4.78trillionineconomiceffect)whichweconcludeisalikelyover‐estimategiventheimprobabilityofAI singularly attaining the same level of significanceas the entire IT sectorof the1990swithinthenexttenyears.At amore detailed level we also conclude that broadband,mobile phones, and roboticstechnologyrepresentreasonablebenchmarks forAI.All three technologiesembodycloseparallelswithAI’spotentialmechanismsforeconomicimpact,andpotentialdiffusionrates.For example, similar to AI’s potential both broadband internet and mobile phonesincreased the speed and accessibility of information allowing for increased productivityacrossmany sectors. Relative to broadband,mobile phones have the advantage of beingstudied globally, and not just in developed, high‐income countries. This benchmarkcorrespondingly yields the upper bound in our range of reasonable estimates at $2.95trillion. Relative to both broadband and mobile phones, robotics offers several furtheradvantages which make it a reasonable benchmark. Namely, robotics is likely to morecloselyreflectAI’smoresector‐targetedeffects,andAI’spotentiallydisruptivenature.GiventherangeofpotentialforAI’sdevelopmentoverthenextdecadewealsorelyonourbenchmarkstoprovideanestimateforAI’seconomicimpactshouldAI’sdevelopmentanddiffusion meet or exceed its strongest proponents current projections. This optimisticupperboundisestimatedtobeashighas$5.89trillionoverthenexttenyears.It is important to note that the benchmarks serve as informative approximations of AI’spotential, but arenot intended toperfectlypredict the future economic effectsofAI.Wecaveat our results with the points that AI’s future diffusion, impacts on developingcountries,andeffectsonemploymentarewide‐rangingandmuch‐debated.OurstudyisnotanchoredonanyspecificfutureofAIbutratherservestointroducearangeofpossibilitiesforAIgoingforward.Inthecourseofits60yearhistory,AIhasfrequentlybeenheraldedasonthecuspofbeingasignificantcontributortoglobaleconomicgrowth.GiventheAIthatexiststodayandtheavailabilityofdataandcomputingpower,AImaybeonthevergeofstartingtorealizeitsmuchanticipatedpotential.
![Whose Roads? Defining Bicyclists' And Pedestrians' Right To Use Public Roads [VTPI, 2013]](https://static.documents.pub/doc/80x56/577cd5b11a28ab9e789b6650/whose-roads-defining-bicyclists-and-pedestrians-right-to-use-public.jpg)
