Thedevelopmentofairaisesprofoundquestions.Perhapsforemostamongthem,though,isastraightforwardone.WhatdoesthismeanfortheeconomyManyhavegrandexpectations.NewresearchbyGoldmanSachs,abank,suggeststhat“widespreadaiadoptioncouldeventuallydrivea7%oralmost$7trnincreaseinannualglobalgdpoveraten-yearperiod.”Academicstudiespointtoathree-percentage-pointriseinannuallabour-productivitygrowthinfirmsthatadoptthetechnology,whichwouldrepresentahugeupliftinincomescompoundedovermanyyears.Astudypublishedin2021byTomDavidsonofOpenPhilanthropy,agrantmakingoutfit,putsamorethan10%chanceon“explosivegrowth”—definedasincreasesinglobaloutputofmorethan30%ayear—sometimethiscentury.Afeweconomists,onlyhalf-jokingly,holdoutthepossibilityofglobalincomesbecominginfinite.
Financialmarkets,however,pointtorathermoremodestoutcomes.Inthepastyearsharepricesofcompaniesinvolvedinaihavedoneworsethantheglobalaverage,althoughtheyhaveriseninrecentmonths(seechart).Interestratesareanotherclue.Ifpeoplethoughtthatthetechnologywasgoingtomakeeveryonerichertomorrow,rateswouldrisebecausetherewouldbelessneedtosave.Inflation-adjustedratesandsubsequentgdpgrowtharestronglycorrelated,pointsoutresearchbyBasilHalperinoftheMassachusettsInstituteofTechnology(mit)andcolleagues.YetsincethehypeaboutaibeganinNovember,long-termrateshavefallen—andtheyremainverylowbyhistoricalstandards.Financialmarkets,theresearchersconclude,“arenotexpectingahighprobabilityof…ai-inducedgrowthacceleration…onatleasta30-to-50-yeartimehorizon.”
Tojudgewhichgroupisright,itishelpfultoconsiderthehistoryofprevioustechnologicalbreakthroughs.Thisprovidessuccourtoinvestors.Foritisdifficulttomakethecasethatasinglenewtechnologybyitselfhasevernoticeablychangedtheeconomy,eitherforgoodorill.Eventheindustrialrevolutionofthelate1700s,whichmanypeoplebelievewastheresultoftheinventionofthespinningjenny,wasactuallycausedbyallsortsoffactorscomingtogether:increasinguseofcoal,firmerpropertyrights,theemergenceofascientificethosandmuchmorebesides.
Perhapsmostfamously,inthe1960sRobertFogelpublishedworkaboutAmerica’srailwaysthatwouldlaterwinhimaNobelPrizeineconomics.ManythoughtthatrailtransformedAmerica’sprospects,turninganagriculturalsocietyintoanindustrialpowerhouse.Infact,ithadaverymodestimpact,Fogelfound,becauseitreplacedtechnology—suchascanals—thatwouldhavedonejustaboutasgoodajob.Thelevelofper-personincomethatAmericaachievedbyJanuary1st1890wouldhavebeenreachedbyMarch31st1890ifrailwayshadneverbeeninvented.
Ofcourse,noonecanpredictwithanycertaintywhereatechnologyasfundamentallyunpredictableasaiwilltakehumans.Runawaygrowthisnotimpossible;noristechnologicalstagnation.Butyoucanstillthinkthroughthepossibilities.And,sofaratleast,itseemsasthoughFogel’srailwaysarelikelytobeausefulblueprint.Considerthreebroadareas:monopolies,labourmarketsandproductivity.
Anewtechnologysometimescreatesasmallgroupofpeoplewithvasteconomicpower.JohnD.RockefellerwonoutwithoilrefiningandHenryFordwithcars.TodayJeffBezosandMarkZuckerbergareprettydominantthankstotech.
Manypunditsexpectthatbeforelongtheaiindustrywillgeneratehugeprofits.InarecentpaperGoldman’sanalystsestimateinabest-casescenariogenerativeaicouldaddabout$430bntoannualglobalenterprise-softwarerevenues.Theircalculationassumesthateachoftheworld’s1.1bnofficeworkerswilladoptafewaigizmos,payingaround$400intotaleach.
Anybusinesswouldbegladtocapturesomeofthiscash.Butinmacroeconomicterms$430bnsimplydoesnotmovethedial.Assumethatalloftherevenueturnsintoprofits,whichisunrealistic,andthatalloftheseprofitsareearnedinAmerica,whichisatadmorerealistic.Evenundertheseconditions,theratioofthecountry’spre-taxcorporateprofitstoitsgdpwouldrisefrom12%todayto14%.Thatisfarabovethelong-runaverage,butnohigherthanitwasinthesecondquarterof2021.
Theseprofitscouldgotooneorganisation—maybeOpenai.Monopoliesoftenarisewhenanindustryhashighfixedcostsorwhenitishardtoswitchtocompetitors.CustomershadnoalternativetoRockefeller’soil,forinstance,andcouldnotproducetheirown.Generativeaihassomemonopolisticcharacteristics.gpt-4,oneofOpenai’schatbots,reportedlycostmorethan$100mtotrain,asumfewfirmshavelyingaround.Thereisalsoalotofproprietaryknowledgeaboutdatafortrainingthemodels,nottomentionuserfeedback.
Thereis,however,littlechanceofasinglecompanybestridingtheentireindustry.Morelikelyisthatamodestnumberofbigfirmscompetewithoneanother,ashappensinaviation,groceriesandsearchengines.Noaiproductistrulyuniquesinceallusesimilarmodels.Thismakesiteasierforacustomertoswitchfromonetoanother.Thecomputingpowerbehindthemodelsisalsofairlygeneric.Muchofthecode,aswellastipsandtricks,isfreelyavailableonline,meaningthatamateurscanproducetheirownmodels—oftenwithstrikinglygoodresults.
Althoughgenerativeaimightnotcreateanewclassofrobberbarons,tomanypeoplethatwillbecoldcomfort.Theyaremoreconcernedwiththeirowneconomicprospects—inparticular,whethertheirjobwilldisappear.Terrifyingpredictionsabound.TynaEloundouofOpenai,andcolleagues,haveestimatedthat“around80%oftheusworkforcecouldhaveatleast10%oftheirworktasksaffectedbytheintroductionofllms”.EdwardFeltenofPrincetonUniversity,andcolleagues,conductasimilarexercise.Legalservices,accountancyandtravelagenciescomeoutatornearthetopofprofessionsmostlikelytoloseout.
Economistshaveissuedgloomypredictionsbefore.Inthe2000smanyfearedtheimpactofoutsourcingonrich-worldworkers.In2013twoatOxfordUniversityissuedawidelycitedpaperthatsuggestedautomationcouldwipeout47%ofAmericanjobsoverthesubsequentdecadeorso.Othersmadethecasethat,evenwithoutwidespreadunemployment,therewouldbe“hollowingout”,whererewarding,well-paidjobsdisappearedandmindless,poorlypaidrolestooktheirplace.
Whatactuallyhappenedtookpeoplebysurprise.Inthepastdecadetheaveragerich-worldunemploymentratehasroughlyhalved(seechart).Theshareofworking-agepeopleinemploymentisatanall-timehigh.Countrieswiththehighestratesofautomationandrobotics,suchasJapan,SingaporeandSouthKorea,havetheleastunemployment.ArecentstudybyAmerica’sBureauofLabourStatisticsfoundthatinrecentyearsjobsclassifiedas“atrisk”fromnewtechnologies“didnotexhibitanygeneraltendencytowardnotablyrapidjobloss”.Evidencefor“hollowingout”ismixed.Measuresofjobsatisfactionroseduringthe2010s.FormostofthepastdecadethepoorestAmericanshaveseenfasterwagegrowththantherichestones.
Thistimecouldbedifferent.ThesharepriceofChegg,afirmwhichprovideshomeworkhelp,recentlyfellbyhalfafteritadmittedChatgptwas“havinganimpactonournewcustomergrowthrate”.Thechiefexecutiveofibm,abigtechfirm,saidthatthecompanyexpectstopausehiringforrolesthatcouldbereplacedbyaiinthecomingyears.ButaretheseearlysignsatsunamiisabouttohitPerhapsnot.
Imagineajobdisappearswhenaiautomatesmorethan50%ofthetasksitencompasses.Orimaginethatworkersareeliminatedinproportiontothetotalshareofeconomywidetasksthatareautomated.Ineithercasethiswould,followingMsEloundou’sestimates,resultinanetlossofaround15%ofAmericanjobs.Somefolkcouldmovetoindustriesexperiencingworkershortages,suchashospitality.Butabigriseintheunemploymentratewouldsurelyfollow—inline,maybe,withthe15%brieflyreachedinAmericaduringtheworstofthecovid-19pandemicin2020.
Theproblemwiththisscenarioisthathistorysuggestsjobdestructionhappensfarmoreslowly.Theautomatedtelephoneswitchingsystem—areplacementforhumanoperators—wasinventedin1892.Ittookuntil1921fortheBellSystemtoinstalltheirfirstfullyautomatedoffice.Evenafterthismilestone,thenumberofAmericanmanualtelephoneoperatorscontinuedtogrow,peakinginthemid-20thcenturyataround350,000.Theoccupationdidnot(mostly)disappearuntilthe1980s,ninedecadesafterautomationwasinvented.aiwilltakelessthan90yearstosweepthelabourmarket:llmsareeasytouse,andmanyexpertsareastonishedbythespeedatwhichthegeneralpublichasincorporatedChatgptintotheirlives.Butreasonsfortheslowadoptionoftechnologyinworkplaceswillalsoapplythistimearound.
InarecentessayMarkAndreessenofAndreessenHorowitzoutlinedsomeofthem.MrAndreessen’sargumentfocusesonregulation.Inbitsoftheeconomywithheavystateinvolvement,suchaseducationandhealthcare,technologicalchangetendstobepitifullyslow.Theabsenceofcompetitivepressurebluntsincentivestoimprove.Governmentsmayalsohavepublic-policygoals,suchasmaximisingemploymentlevels,whichareinconsistentwithimprovedefficiency.Theseindustriesarealsomorelikelytobeunionised—andunionsaregoodatpreventingjoblosses.
Examplesabound.TraindriversonLondon’spubliclyrunUndergroundnetworkarepaidclosetotwicethenationalmedian,eventhoughthetechnologytopartiallyorwhollyreplacethemhasexistedfordecades.Governmentagenciesstillrequireyoutofillinpaperformsprovidingyourpersonalinformationagainandagain.InSanFrancisco,theglobalcentreoftheaisurge,real-lifecopsarestillemployedtodirecttrafficduringrushhour.
Manyofthejobsatriskfromaiareinheavilyregulatedsectors.ReturntothepaperbyMrFeltenofPrincetonUniversity.Fourteenofthetop20occupationsmostexposedtoaiareteachers(foreign-languageonesarenearthetop;geographersareinaslightlystrongerposition).Butonlythebravestgovernmentwouldreplaceteacherswithai.Imaginetheheadlines.Thesamegoesforcopsandcrime-fightingai.ThefactthatItalyhasalreadytemporarilyblockedChatgptoverprivacyconcerns,withFrance,GermanyandIrelandsaidtobeconsideringtheoption,showshowworriedgovernmentsareaboutthejob-destructiveeffectsofai.
Perhaps,intime,governmentswillallowsomejobstobereplaced.Butthedelaywillmakespacefortheeconomytodowhatitalwaysdoes:createnewtypesofjobsasothersareeliminated.Byloweringcostsofproduction,newtechcancreatemoredemandforgoodsandservices,boostingjobsthatarehardtoautomate.Apaperpublishedin2020byDavidAutorofmit,andcolleagues,offeredastrikingconclusion.About60%ofthejobsinAmericadidnotexistin1940.Thejobof“fingernailtechnician”wasaddedtothecensusin2000.“Solarphotovoltaicelectrician”wasaddedjustfiveyearsago.Theaieconomyislikelytocreatenewoccupationswhichtodaycannotevenbeimagined.
Modestlabour-marketeffectsarelikelytotranslateintoamodestimpactonproductivity—thethirdfactor.AdoptionofelectricityinfactoriesandhouseholdsbeganinAmericatowardstheendofthe19thcentury.Yettherewasnoproductivityboomuntiltheendofthefirstworldwar.Thepersonalcomputerwasinventedinthe1970s.Thistimetheproductivityboomfollowedmorequickly—butitstillfeltslowatthetime.In1987RobertSolow,aneconomist,famouslydeclaredthatthecomputeragewas“everywhereexceptfortheproductivitystatistics”.
Theworldisstillwaitingforaproductivitysurgelinkedtorecentinnovations.Smartphoneshavebeeninwidespreaduseforadecade,billionsofpeoplehaveaccesstosuperfastinternetandmanyworkersnowshiftbetweentheofficeandhomeasitsuitsthem.OfficialsurveysshowthatwelloveratenthofAmericanemployeesalreadyworkatfirmsusingaiofsomekind,whileunofficialsurveyspointtoevenhighernumbers.Still,though,globalproductivitygrowthremainsweak.
aicouldeventuallymakesomeindustriesvastlymoreproductive.ApaperbyErikBrynjolfssonofStanfordUniversity,andcolleagues,examinescustomer-supportagents.Accesstoanaitoolraisesthenumberofissuesresolvedeachhourby14%onaverage.Researchersthemselvescouldalsobecomemoreefficient:gpt-xmaygivethemanunlimitednumberofalmost-freeresearchassistants.Othershopeaiwilleliminateadministrativeinefficienciesinhealthcare,reducingcosts.
Buttherearemanythingsbeyondthereachofai.Blue-collarwork,suchasconstructionandfarming,whichaccountforabout20%ofrich-worldgdp,isoneexample.Anllmisoflittleusetosomeonepickingasparagus.Itcouldbeofsomeusetoaplumberfixingaleakytap:awidgetcouldrecognisethetap,diagnosethefaultandadviseonfixes.Ultimately,though,theplumberstillhastodothephysicalwork.Soitishardtoimaginethat,inafewyears’time,blue-collarworkisgoingtobemuchmoreproductivethanitisnow.Thesamegoesforindustrieswherehuman-to-humancontactisaninherentpartoftheservice,suchashospitalityandmedicalcare.
Itisevenpossiblethattheaieconomycouldbecomelessproductive.Lookatsomerecenttechnologies.Smartphonesallowinstantcommunication,buttheycanalsobeadistraction.Withemailyouareconnected24/7,whichcanmakeithardtofocus.Apaperin2016byresearchersattheUniversityofCaliforniaatIrvine,MicrosoftResearchandmitfinds“thelongerdailytimespentonemail,thelowerwasperceivedproductivity”.Somebossesnowbelievethatworkingfromhome,onceseenasaproductivity-booster,givestoomanypeopletheexcusetoslackoff.
Generativeaiitselfcouldactasadrainonproductivity.Whathappens,forinstance,ifaicancreateentertainmentperfectlytailoredtoyoureverydesireMoreover,fewpeoplehavethoughtthroughtheimplicationsofasystemthatcangeneratevastamountsoftextinstantly.gpt-4isagodsendforanimbyfacingaplanningapplication.Infiveminuteshecanproduceawellwritten1,000-pageobjection.Someonethenhastorespondtoit.Spamemailsaregoingtobehardertodetect.Fraudcasescouldsoar.Bankswillneedtospendmoreonpreventingattacksandcompensatingpeoplewholoseout.
Inanai-heavyworldlawyerswillmultiply.“Inthe1970syoucoulddoamulti-million-dollardealon15pagesbecauseretypingwasapainintheass,”saysPrestonByrneofBrownRudnick,alawfirm.“aiwillallowustocoverthe1,000mostlikelyedgecasesinthefirstdraftandthenthepartieswillargueoveritforweeks.”AruleofthumbinAmericaisthatthereisnopointsuingfordamagesunlessyouhopefor$250,000ormoreincompensation,sinceyouneedtospendthatmuchgettingtocourt.Nowthecostsoflitigationcouldfalltoclosetozero.Meanwhile,teachersandeditorswillneedtocheckthateverythingtheyreadhasnotbeencomposedbyanai.Openaihasreleasedaprogramthatallowsyoutodothis.Itisthusprovidingtheworldasolutiontoaproblemthatitstechnologyhascreated.
aimaychangetheworldinwaysthattodayareimpossibletoimagine.Butthatisnotthesamethingasturningtheeconomyupsidedown.AsFogelnotedinhisstudy:“TheprecedingargumentisaimednotatrefutingtheviewthattherailroadplayedadecisiveroleinAmericandevelopmentduringthe19thcentury,butratheratdemonstratingthattheempiricalbaseonwhichthisviewrestsisnotnearlysosubstantialasisusuallypresumed”.Sometimeinthemid-21stcenturyafutureNobelprizewinner,examininggenerativeai,maywellreachthesameconclusion.（完）