CEPR-政策洞察：我們能擁有支持工人的人工智能嗎？選擇一條為心靈服務(wù)的機器之路（英）

CEPRPOLICYINSIGHTNo.123

October2023

CanweHavePro-WorkerAI?

Choosingapathofmachinesinserviceofminds

DaronAcemoglu,DavidAutor,andSimonJohnson1

MassachusettsInstituteofTechnologyandCEPR

Summary

?Overthepast40years,thediffusionofdigitaltechnologiessignificantlyincreasedincomeinequality.

?GenerativeArtificialIntelligence(AI)willsurelyimpactinequality,butthenatureofthateffectdependsonexactlyhowthistechnologyisdevelopedandapplied.Nothingaboutthepathofthis(orany)technologyisinevitable.

?TheprivatesectorispursuingapathforgenerativeAIthatemphasisesautomationandthedisplacementoflabour,alongwithintrusiveworkplacesurveillance.

?Simplydisplacingworkersisnevergoodforthelabourmarket,evenwhenthedisplacedarehighlypaid.Displacedformerlyhigh-paidworkersareforcedtocompeteforjobswithlower-wageworkers,leadingtoadownwardcascadeinwagelevels.

?Abetterpathisavailable,alongwhichgenerativeAIwouldbecomplementarytomosthumans-augmentingtheircapabilities-includingpeoplewithoutafour-yearcollegedegree.

?Choosingthehuman-complementarypathisfeasiblebutwillrequirechangesinthedirectionoftechnologicalinnovation,aswellasincorporatenormsandbehaviour.

?ThegoalshouldbetodeploygenerativeAItocreateandsupportnewoccupationaltasksandnewcapabilitiesforworkers.IfAItoolscanenableteachers,nursepractitioners,nurses,medicaltechnicians,electricians,plumbers,andothermoderncraftworkerstodomoreexpertwork,thiscanreduceinequality,raiseproductivity,andboostpaybylevellingworkersup.

?Publicpolicyhasacentralroleinencouragingthispositivepathoftechnologytocomplementallworkers,elevatingtheachievablelevelofskillandexpertiseforeveryone.

1Theauthorsareco-directorsoftheMITShapingtheFutureofWorkInitiative,whichwasestablishedthrougha

generousgiftfromtheHewlettFoundation.Thisresearchprogramandrelatedresultswerealsomadepossiblewith

thesupportoftheNOMISFoundation.Acemoglu:InstituteProfessor,MIT;Autor:FordProfessorofEconomics,

MITDepartmentofEconomics;Johnson:KurtzProfessorofEntrepreneurship,MITSloanSchoolofManagement.

Relevantdisclosuresareavailableat

/power-and-progress

,under“PolicySummary.”.

CEPRPOLICYINSIGHTNo.123October2023

2

?Atthistime,thefivemostimportantfederalpoliciesshouldbe:

1.Equalisetaxratesonemployingworkersandonowningequipment/algorithmstoleveltheplayingfieldbetweenpeopleandmachines.

2.UpdateOccupationalSafetyandHealthAdministrationrulestocreatesafeguards(i.e.limitations)onthesurveillanceofworkers.Findingwaystoelevateworkervoiceonthedirectionofdevelopmentcouldalsobehelpful.

3.Increasefundingforhuman-complementarytechnologyresearch,recognisingthatthisisnotcurrentlyaprivatesectorpriority.

4.CreateanAIcentreofexpertisewithinthegovernment,tohelpshareknowledgeamongregulatorsandotherofficials.

5.Usethatfederalexpertisetoadviseonwhetherpurportedhuman-complementarytechnologyisappropriatetoadoptinpubliclyprovided

educationandhealthcareprogrammes,includingatthestateandlocallevel.

Introduction

TheworldisabouttoexperiencetransformativeanddisruptiveadvancesingenerativeArtificialIntelligence.Amajorsetofconcernscentresaroundthelabourmarketandeconomicinequalityimplicationsoftheseadvances.WillAIeliminatejobsinnet?Willitfurtherinflamethedecades-longphenomenonofrisingeconomicinequality?Willitboostlabourearningsorinsteadmakemachinesmorevaluableandworkersmoreexpendable?

Theconsensusintheeconomicliteratureisthatpreviouswavesofdigitaltechnologies–includingpersonalcomputers,numericallycontrolledmachinery,robotics,andofficeautomation–haveincreasedinequality.Thisisbothbecausesomeofthesetechnologies,suchaspersonalcomputers,havebeenhighlycomplementarytomore-educatedworkers(Autoretal.1998,Autoretal.2003,GoldinandKatz2008),andalsobecausemanyofthesetoolshavebeenusedforautomatingwork,withanunequalimpactondifferenttypesofworkers(Autoretal.2003,AcemogluandRestrepo2022a,2022b).Whiledigitaltechnologieshaveundoubtedlycreatednewgoods/servicesandboostedproductivityinsomeactivities(e.g.BrynjolfssonandMcAfee2015),thereisalsoevidencethatproductivitygainsfromthesetechnologieshavesometimesfallenwellbelowexpectations(e.g.Acemogluetal.2016).

GenerativeAIwillhaveasubstantialimpactonthefutureofworkandthetrajectoryofinequality.ThenatureofthatimpactisnotaninevitableconsequenceofthetechnologyitselfbutinsteaddependsonhowsocietydevelopsandshapesAI.

?ThecurrentlypredominantdirectionforAIemphasisesautomation,displacementofskilledlabour,anddiminishedworkervoiceduetostepped-upmonitoringandsurveillance.

?Analternative,“human-complementary”pathcouldcontributemoretoproductivitygrowthandcouldhelpreduceeconomicinequality.

Inthenextsection,weoutlinewhattheautomationpathlookslikeandwhatitsimplicationswouldbeforwork,inequality,andproductivity.Wethendescribethealternativehuman-complementarypath,drawingonbothgeneralprinciplesandspecificexamples.Wealsoexplainwhy,despiteitsadvantages,thehuman-complementaryapproachisnotlikelytoprevailbasedoncurrentinvestmentsandcorporateattitudes.WesuggestpoliciesthatcouldhelpsteerAIdevelopmentandimplementationinthemoreconstructivedirection.

CEPRPOLICYINSIGHTNo.123October2023

3

THEAUTOMATIONPATH

Automation–thesubstitutionofmachinesand,morerecently,algorithmsfortaskspreviouslyperformedbyhumans–hasbeenaconstantsinceatleastthebeginningoftheIndustrialRevolution.Machinesweredevelopedtoperformtasksthathaveahighdegreeofpredictabilityandarecarriedoutinstableenvironments.Examplesincludespinningandweavingintheeighteenth-centurytextileindustry,harvestinginnineteenth-centuryagriculture,andmanyofficeandclericaltasksinthetwentiethcentury,suchastelephoneswitchboardoperationandroutinebookkeeping.Massproduction,whichvastlyreducedthecostofeverydayproducts,dependsfundamentally–thoughnotexclusively–onassemblylinesmadepossiblebyautomation.

Notallautomationishighlyproductive,however.Whenmachinesaredeployedtoperformtasksinwhichtheyarenotparticularlyeffective,lacklustreproductivitygainsfollow.Mostpeoplearefamiliarwiththefrustrationofseekingcustomerservicefromanairline,creditcardprovider,orcomputermanufacturer,onlytobedivertedthroughmazesofunhelpfulcomputermenus.Firmsmayfindsuchautomationtobecost-effective,butitisnotameaningfulproductivityadvance.

Whetherithaslargeorsmalleffectsonproductivity,automationtendstohavemajordistributionalconsequences.Thereasonisthatautomationdisplacesworkerswhowerespecialisedinthetasksthatautomationreallocatestomachinesandalgorithms.Theautomationofblue-collarandofficejobsusingdigitaltechnologieshasbeenanimportantdriveroftheriseininequalitysince1980(AcemogluandRestrepo2022a).

ItisinevitablethatAIsystemswillbeusedforsomeautomation,bothfortechnicalandbusinessstrategyreasons.Onthetechnicalfront,amajorbarriertoautomationofmanyserviceandproductiontaskshasbeenthattheyrequireflexibility,judgment,andcommonsense–thingsthatarenotablyabsentfrompre-AIformsofautomation.Artificialintelligence,especiallygenerativeAI,canpotentiallymastersuchtasks(Susskind2021).AbroadswathofcomputersecuritytasksthatusedtobeperformedbyskilledhumanoperatorscannowbeperformedbyAIbots.Similarly,generativeAIsystemscanwriteadvertisingcopy,parselegaldocuments,transcribephysicians’medicalnotes,andperformlanguagetranslation.Itisunclearhowmuchthistypeofautomationwillcontributetoaggregateproductivitygrowthwhilethesetechnologiesareimmature,buttheycouldcontributetosizeableproductivitygainsascostsfallandreliabilityimproves.

Businessesmaychoosemachinesoverworkersforreasonsotherthanproductivity.Automationappealstomanagerswhoareseekinggreaterconsistencyandlessoppositionfromorganisedorunorganisedlabour(AcemogluandJohnson2023).

Alltoooften,businessesprefertofocusonautomationratherthancreatingnewjobtasksandenablingworkerstobuildnewskills.Automationisalwaysaneasycaseforamanagertomakebecauseitappearstosavecosts.Investingtomakeworkersmoreproductiveormoreusefulmaybeahardersell,sinceit’sseenasmessy,uncertain,andexpensive.Somemanagerssimplypreferto“hiremachines”,ratherthanhiringworkers,becausemachinesdon’tcomplainaboutpayorworkingconditions,andtheycertainlydon’tjoinunions.Butacountryisnotabusiness.Wehaveasharedinterestinensuringthatadultsareproductivelyemployed.Thispromoteseconomicresilience,socialcohesion,andastrongtaxbase.Policymakerscaremoreaboutthequalityandquantityofjobsthandoemployers,andpolicyshouldsupportinstitutions,incentives,andinvestmentswiththisinmind.

CEPRPOLICYINSIGHTNo.123October2023

4

Beyondeconomicincentives,thedominantintellectualparadigmintoday’sdigitaltechsector–amongbothbusinessleadersandacademicresearchers–favourstheautomationpath.AmajorfocusofAIresearchistoattainhumanparityinavastrangeofcognitivetasksand,moregenerally,toachieve“artificialgeneralintelligence”thatfullymimicsandthensurpassescapabilitiesofthehumanmind.Thisintellectualfocusencouragesautomationratherthanthedevelopmentofhuman-complementarytechnologies(AcemogluandJohnson2023).

Thereisawidelysharedoptimismthathuman-replacingautomationwillproducesomuchproductivitygainthatalltypesofworkerswillbenefit.Itistruethatifautomationissufficientlyproductive(andthusreducescostsbyasignificantamount),thiscangeneratedemandforothergoodsandservicesand,asaresult,workersmaybenefitaswell.

However,whileproductivitygainsareobviouslywelcome,therearetwoproblemswiththislineofthought.First,thebenefitsmaybehighlyunequallydistributedacrossdifferentskillgroups.Forexample,AI-basedproductivitygainsmightincreasethedemandforso-called“promptengineers”,butthiswouldnothelpworkersdisplacedfromaccountingorfinancialservicesjobs–assumingtheydonothaveacomparativeadvantageinpromptengineering(AcemogluandAutor2011,AcemogluandRestrepo2022a).Second,automationtendstoreducethelabourshareofnationalincome,soevenifworkersbenefit,mostofthegainsflowtoentrepreneursandcapitalowners(AcemogluandRestrepo2018).Therearelimitstohowmuchsharedprosperitycanbegeneratedexclusivelybyautomation.

Anothercommonpredictionisthat,becausegenerativeAImayautomatemanagerialorknowledgetaskstypicallyperformedbyprofessionalworkers,itcouldhaveequalisingeffects.Forinstance,ifaccountantsandfinancialanalystslosetheirjobs,thismightreduceinequalitybetweenretailworkersandhighlypaidfinancialsectorworkers.Thislogicisfaulty.Studiesofpreviouswavesofdigitalautomationshowthatworkersdirectlydisplacedbynewtechnologiesnotonlyexperiencelowerpaygrowthbutalsostartcompetingwithothergroupsoflower-paidworkers,whosepaythendeclines(e.g.AcemogluandRestrepo2022a).Simplyput,displacingworkerswillneverbegoodforworkersorforthelabourmarket.Instead,AIcanreduceinequalityifitenableslower-rankedworkerstoperformmorevaluablework–butnotifitmerelyknocksrungsoutoftheexistingjobladder.

TheHumanComplementaryPath

Newtechnologiesneednotmerelyreplaceworkersinexistingtasks.Theymayalsocomplementworkersbyenablingthemtoworkmoreefficiently,performhigher-qualitywork,oraccomplishnewtasksthatwerepreviouslyinfeasible(AcemogluandRestrepo2018,Autoretal.2022,AcemogluandJohnson2023).Forexample,evenasmechanisationgraduallypushedmorethanhalfoftheUSlabourforceoutofagriculture,arangeofnewblue-collarandclericaltasksinfactoriesandnewlyemergingserviceindustriesgeneratedsignificantdemandforskilledlabour.Theexpansionofemploymentinindustryandservicesbetween(roughly)theyears1870and1970ledtoworkthatwasnotonlybetterpaidbutalsolessdangerousandlessphysicallyexhausting,andincreasinglyrewardedtheformalliteracyandnumeracyskillscreatedbytheexpansionofuniversalpublichighschooleducation.

Thisvirtuouscombination-automationoftraditionalworkalongsidecreationofnewtasks–proceededinrelativebalanceformuchofthetwentiethcentury.Butsometimeafterapproximately1970,thisbalancewaslost.Whileautomationhasmaintaineditspaceorevenacceleratedovertheensuingfivedecades,theoffsettingforceofnew

CEPRPOLICYINSIGHTNo.123October2023

5

taskcreationhasslowed,particularlyforworkerswithoutfour-yearcollegedegrees(AcemogluandRestrepo2019,Autoretal.2022).Non-collegeworkershavebeendisplacedfromfactoriesandofficesbycomputerisationand,forblue-collarworkers,alsobyimportcompetition(Autoretal.2013),butnonewequivalentlywell-paidopportunitieshaveemergedtoattracttheseworkers.Asaresult,non-collegeeducatedworkersareincreasinglyfoundinlow-paidservicessuchascleaning,security,foodservice,recreation,andentertainment.Thesejobsaresociallyvaluable,buttheyrequirelittlespecialisededucation,training,orexpertise,andhencepaypoorly.

ThecriticalquestionwefaceintheneweraofgenerativeAIiswhetherthistechnologywillprimarilyacceleratetheexistingtrendofautomationwithouttheoffsettingforceofgoodjobcreation–particularlyfornon-collegeworkers–orwhetheritwillinsteadenabletheintroductionofnewlabour-complementarytasksforworkerswithdiverseskillsetsandawiderangeofeducationalbackgrounds.

Thereisacaseforqualifiedoptimism:generativeAIoffersanopportunitytocomplementworkerskillandexpertise.

Becausesomanyoftheroutinetasksthatworkerspreviouslyperformedhavealreadybeenautomated,alargefractionofcurrentjobsrequirenon-routineproblem-solvinganddecision-makingtasks.Empoweringworkerstoperformthesetasksmoreeffectively,andtoaccomplishevenmoresophisticateddecision-makingtasks,willrequireprovidingworkerswithbetterinformationanddecision-supporttools.GenerativeAIisparticularlywell-suitedtothistypeofinformationprovision.Anironyoftoday’sinformationeraisthatpeopleareoverwhelmedbyinformationbutoftenlackthetimeandexpertisetoparsethisinformationeffectively.GenerativeAIisparticularlywell-suitedtoaddressthisproblem.Withproperdevelopment,AItoolscanhelpsurfacerelevantinformationattherighttimetoenablebetterdecision-making.

Additionally,andcloselyrelated,humanproductivityisoftenhamperedbylackofspecificknowledgeorexpertise,whichcouldbereadilysupplementedbyAI.Forexample,anelectricianmightbeunabletodiagnosearareproblemthatshecouldreadilyaddressifgivenrelevanttoolsorappropriatetraining.OrahighlytrainedimmigranttotheUSmaybeinhibitedfromfullyusingherabilitiesbecauseoflimitedEnglishlanguageskills.GenerativeAItoolscanassistinsuchcasesbyboostinghumanexpertise,supportingworkersinunfamiliarsituations,providingon-the-spottraining,andimprovingallformsofinformationtranslation.Overall,AIholdsgreatpotentialfortrainingandretrainingexpertworkers,suchaseducators,medicalpersonnel,softwaredevelopers,andotherworkerswithmodern“crafts”(suchaselectriciansandplumbers).

Finally,whilegenerativeAImaytakeovermoreoftheoperationaltasksincertainoccupations,suchasaccounting,financialanalysis,orcomputerprogramming,ifdevelopedintherightmanner,itcouldcreatenewdemandsforhumanexpertiseandjudgmentinoverseeingtheseprocesses,communicatingwithcustomers,andenablingmoresophisticatedservicesthatleveragethesetools.

Severalrecentstudiesprovide“proof-of-concept”examplesthatdemonstratehowgenerativeAIcansupplementexpertiseratherthandisplaceexperts.Pengetal.(2023)demonstratethatGitHubCopilot,agenerativeAI-basedprogrammingaid,cansignificantlyincreaseprogrammerproductivity.2Inacontrolledexperiment,

2

/features/copilot

October2023

CEPRPOLICYINSIGHTNo.123

6

thetreatmentgroupthatwasgivenaccesstothisgenerativeAItoolcompletedtherequiredprogrammingtaskabout56%fasterthanthecontrolgroupwithoutaccesstoCopilot.

NoyandZhang(2023)performedarelatedonlinerandomisedcontrolledtrial,focusedonwritingtasks.Amongthesetofwhite-collarworkersrecruitedforthestudy,halfwererandomlygivenaccessto(andencouragedtouse)ChatGPTforwritingtasks.NoyandZhang(2023)foundsignificantimprovementsinthespeedandqualityofwritingoutput.Mostimportantly,thebiggestimprovementswereconcentratedamongtheleast-capablewriters.AlthoughgenerativeAIdidnotmaketheleast-skilledwritersaseffectiveasthemost-skilledwriters,itmadeallwritersfasterandsubstantiallyreducedthequalitygapbetweenthetwogroups.

Finally,Brynjolfssonetal.(2023)evaluatedtheuseofgenerativeAItoolsusedforprovidingbackgroundinformationtocustomerserviceagents.Theyalsoestimatedasignificantimprovement(about14%)inproductivity,andlikeNoyandZhang’sstudy,thesegainswerethemostpronouncedfornoviceworkers.UsingtheseAItools,noviceworkerswereabletoreachalevelofproficiencywithinthreemonthsthatpreviouslytookworkerstenmonthstoattain.

Inallthreecases,generativeAItoolsautomateandaugmenthumanworksimultaneously.Theautomationstemsfromtime-savings:AIwritesthefirstdraftofcomputercode,advertisingcopy,andcustomersupportresponses.AugmentationhappensbecauseworkersarecalledupontoapplyexpertiseandjudgmenttointermediatebetweentheAI’ssuggestionsandthefinalproduct–whetheritissoftware,text,orcustomersupport.

PromisingApplications

Lookingforward,weseeatleastthreemajorsectorswherehuman-complementaryAIcouldbetransformative,bothforproductivityandforsharedprosperity.

EDUCATION

GenerativeAItoolscanenablemajoradvancementsineducation,togetherwithnewproductivity-enhancingrolesforeducators.Classroominstructionishinderedbythefactthattheteachermustchooseonepaceatwhichanentireclassproceeds,evenifitistoofastforsomestudentsandtooslowforothers.Individualisededucationprogrammesandpersonalisedteachingtoolscanbeeffectiveinenablinglesspreparedstudentstoexcel,butthesetoolsarelabour-intensiveandhenceexpensive.AI-enabledtoolshavethepotentialtovastlyimprovetutoringandself-instruction.

Khanmigo,anappbuiltonChatGPT-4,isaninfinitelypatient,highlyadaptabletutorthatcanbreakcomplexproblemsintotheirconstituentparts,walkstudentsstep-by-stepthroughsolvingthem,andprovidehintsandexamplesalongthewaywithoutdirectlyansweringquestions.3ResearchshowsthatLargeLanguageModels(LLMs)cananticipatewhichpartsofaproblemhumanswillfinddifficultandsuggestsimplificationstoimproveunderstanding(seeBubecketal.2023).

Thesetechnologiescansupporteducatorsaswellasstudents.Teacherscouldfocusmoreoftheirtimeoninstructionandlessonremediation.Theycoulddevelopricherlessonplansthatharnessnewtools,suchasvisualisation,simulation,andeven

3

/khan-labs#khanmigo

October2023

CEPRPOLICYINSIGHTNo.123

7

real-time“interaction”withfictitiousorhistoricalfigures.TherightAIinvestmentscouldenablesignificantgainsinstudentlearning,especiallyamongcurrentlyunderperformingpupils.Suchreorganisationwouldalsocreatearangeofnewtasksforeducators-andwouldalmostcertainlynecessitatefurtherproductivetrainingforexistingteachersandhiringofadditionalAI-savvyteachers.

Itisnotclear,however,thattheprivatesector(oranypublicschooldistrict)hastherightincentivestodevelopsuchAItools.WillAIbedevelopedtoreducegapsbetweenmore-capableandless-capablelearners–asChatGPTdoesforprofessionalwriters–orwillitinsteadbedeployedtoreduceteacherheadcountsandsubstituteforpersonalattention?Thisisnotaquestionofwhatthetechnologycandobuthowwecollectivelydecidetodevelopanddeployit.ApplicationsofAIthatappeartoreducelabourcostsmaybemoreappealingtomanycash-strappedschooldistrictsintheUnitedStates.Thus,ourconcernisthatAIwillbeusedtoautomateteaching,testing,andgrading,ratherthanfornewpersonalisededucation-targetingtasks.

HEALTHCARE

GiventhatoneinfiveUSdollarsisspentonhealthcare,anytechnologythatimprovesefficiency,lowerscosts,orbroadensaccesstothehealthcaresystemhaspotentiallyenormousbenefits.GenerativeAItoolscanimprovehealthcaredeliveryandaccessibility,enablingproductivitygainsandgeneratingvaluablenewworkertasks.Forexample,generativeAItoolscouldsupportexpandingscopeofpracticeboundaries,enablingmedicalprofessionalsatalllevelstoaccomplishabroaderrangeoftasks.Justasnursepractitionershaveprovedeffectiveatdiagnosing,prescribing,andtreating(Asubontengetal.1995,Lietal.2012)–tasksformerlydoneexclusivelybydoctors–thedecision-supportcapacitiesofAIcanenablealargersetoftrainedmedicalprofessionalstoaccomplishexperttaskswithoutexclusivelyrelyingonthemostelitemedicalprofessionals.UsingAI,qualifiednursepractitioners,nurses,andhealthtechnicianscoulddiagnoseroutinehealthproblems,recommendcoursesoftreatment,andmoreefficientlyroutepatientstofurthercareoptions.

MODERNCRAFTWORKERS(MCWS)

Inlinewiththeelectricianexampleabove,generativeAItoolscanbetransformativeformoderncraftworkers(MCWs)morebroadly.4TheUSiscurrentlyundertakingamajorinfrastructureinvestmentagenda,withgrowingemploymentinmanufacturing,greenenergyproduction,andchipproduction,amongothersectors.SkilledMCWsareinshortsupplyduetoanagingpopulationanddecadesofover-investmentinhighereducationattheexpenseofvaluablevocationaltraining.AIcanbeusedtosupporttrainingandenableMCWstocarryoutawiderrangeoftasksthatrequirespecialisedexpertise.AIcanproviderelevantinformation,real-timeinstruction,anddecision-makingsupportinelectricalwork,plumbing,expertrepair,design,andconstruction,amongotheractivities.ThecurrentgenerationofAItechnologiescannotreplacetheworkofMCWs–whosetasksrequiredexterity,flexibility,andjudgmentthatarefarbeyondthegraspofcurrentrobotics(evenAI-enhancedrobotics).ButAIcanenabletheseworkerstodomorewiththeirskillsbysolvingabroaderanddeeperrangeofappliedproblemsinthefield.

4Alsoknownas“moderntrades”or“tradespeople”.

CEPRPOLICYINSIGHTNo.123October2023

8

WhatCanBeDone

Taxsystem:ThecurrentUStaxcodeplacesaheavierburdenonfirmsthathirelabourthanthosethatinvestinalgorithmstoautomatework(Acemogluetal.2020).Weshouldaimtocreateamoresymmetrictaxstructure,wheremarginaltaxesforhiring(andtraining)labourandforinvestinginequipment/softwareareequated.Thiswillshiftincentivestowardhuman-complementarytechnologicalchoicesbyreducingthebiasofthetaxcodetowardphysicalcapitaloverhumancapital.

Labourvoice:ThedirectionofAIwillhaveprofoundconsequencesforallworkers.Creatinganinstitutionalframeworkinwhichworkersalsohaveavoicewouldbeaconstructivestep–andthereisanimportantroleforcivilsocietyinpressingforthistohappen,includingthrougharticulatingneedsatthelocalandstatelevel.Ataminimum,federalgovernmentpolicyshouldrestrictdeploymentofuntested(orinsufficientlytested)AIforapplicationsthatcouldputworkersatrisk–forexample,inhigh-stakespersonneldecision-makingtasks(includinghiringandtermination)orinworkplacemonitoringandsurveillance.Healthandsafetyrulesneedtobeupdatedaccordingly.

Fundingformorehuman-complementaryresearch:Giventhatthecurrentpathofresearchhasabiastowardautomation,additionalsupportfortheresearchanddevelopmentofhuman-complementaryAItechnologiescouldhavesignificantimpact.Itishardtotargethuman-complementaryworkintheabstract.Itisfeasible,however,tofocusonspecificsectorsandactivitieswhereopportunitiesarealreadyabundant.Theseincludeeducation,healthcare,andMCWtraining.JustasDARPAorchestratedinvestmentsandcompetitionstofosterthedevelopmentofself-drivingcarsanddexterousrobotics,thefederalgovernmentshouldfostercompetitionandinvestmentthatpairsAItoolswithhumanexpertise,aimingtoimproveworkinvitalsocialsectors.

AIexpertisewithinthefederalgovernment:AIwilltoucheveryareaofgovernmentinvestment,regulation,andoversight,including(butnotlimitedto):transportation,energyproduction,labourconditions,healthcare,education,environmentalprotection,publicsafety,andmilitarycapabilities.Developing