WhatdoesthepublicthinkaboutAI?

US,2024

JonathanDupont,VinousAli,DeniseBaron,AmyPrice&SebWride

May2024

1

Contents

IntroductionandSummary3

Foreword6

Today8

HowdopeoplefeelaboutAI?9

HowmuchdopeopleunderstandhowAIworks?11

HowareAmericansusingAItoday?13

Opportunities17

ThePotentialBenefitsofAI18

AtWork21

Health23

Concerns25

WhichrisksdoAmericansworryabout?26

MisinformationandDeepfakes29

Automation33

Future37

ArtificialGeneralIntelligence38

Superintelligence40

Policy43

HowshouldtheGovernmentreact?44

China46

2

IntroductionandSummary

Thelast12monthshaveseenArtificialIntelligence(AI)andthediscoursearounditcontinuetorapidlyevolve.

Followingupon

lastyear’sresearch,

weran4extensivenewnationallyrepresentativepollsofadultsacrosstheUSandtheUK,askingthepublictheirviewsonarangeofAIissues:theirfeelingstowardsit,howtheyusedAItoday,howtheyexpectedittoevolve,andwhattheywantedtheGovernmenttodoinresponse.ThisreportexploresthefindingsfromourUSsurveywork.

WeaskedthemtheirviewsoneverythingfromAIagentstomisinformation,whetheranAIcouldpasstheTuringTestandhowimportantitwasfortheUStomaintainatechnologicalleadaheadofChina.

Herearesomeofthemoreinterestingthingsthatwefound:

Today

●ThemostcommonlychosenemotionbyAmericansfortheirfeelingsaroundAIwascuriosity-buttheywerealsoaroundafifthmorelikelytochoosenegativeemotionsthisyearthanlastyear.

●Onlyaroundathird(32%)ofAmericanssaidthattheywereconfidenttheycouldexplainhowmodernAImodelsworked.Inpractice,peoplestillseemedtobeunsureabouttheirrelativestrengthsandweaknesses:thinkingthattheywerebestatmath,andweakestatsoundingsympathetic.

●Onlyaround13%ofAmericansarecurrentlyusingoneoftheLLMbasedchatbotsregularly.Thatsaid,awarenessofChatGPTspecificallyhasnearlydoubledinthelastyear.

Uses

●WhileAmericansthoughtacceleratingscienceandimprovingeconomicproductivitycouldbeimportantbenefitsofAI,theyweremuchlesssureiftheywouldbenefitdirectly-just7%thoughtitwaslikelytoincreasewagesforworkers.

●AmajorityofAmericansunder35saidtheywouldbeinterestedintryinganAIpersonaltutor,personalassistant,workoutcoachorfinancialadvisor.43%ofAmericanmenunder35saythattheywouldbeinterestedinanAIgirlfriendorboyfriend.

3

●AroundhalfofAmericanworkerswhouseLLMchatbotssaythattheyworkedouthowtousethosetoolsthemselves.Onlyaroundathird(36%)wereinterestedinformaltraining-althoughthisproportionwassignificantlyhigherforolderworkers.

●Givingpeoplechoiceisessentialtomaintainsupport.61%ofAmericanssaytheywouldsupportAIdiagnosisaslongaspeoplehadachoice,whereastwo-thirdswereopposedifitwasforcedonpeople.

Concerns

●WhilethepotentialofAIcreatedunemploymentwastheriskAmericanssaidtheyweremostfamiliarwith,misinformationistherisktheyaremostcurrentlyworriedabout.Republicans,inparticular,areworriedaboutAIbeingbiasedagainstpeoplewithdifferentpoliticalviews.

●AroundahalfofAmericans(46%)werenotconfidentthattheycoulddetectfakeAIgeneratedcontentontheInternet-withconfidencesignificantlyfallingtheolderourrespondentgot.71%ofAmericansareworriedthatAIgeneratedcontentcouldbeusedtomanipulateanelection.

●ThosewithaBachelor’sDegreeorMaster’swereactuallyslightlymorelikelytobelievethatAIcoulddotheirjobwithintwentyyearsthanthosewithjustaHighSchooleducation.59%ofAmericansthinkitislikelythatAIwillincreaseunemployment.

Future

●55%ofAmericansbelievethatanAIwouldreachahumanlevelofintelligencebytheendofthe2030s,comparedto56%lastyear.Bytheendofthe2040s,theyexpectittobesignificantlymoreintelligent.

●Just23%ofAmericansthinkthattryingtocreateasuperintelligenceisagoodidea,

with44%admittingtheywereunawarethatthiswasagoalofleadingAIlabs.

Policy

●AmajorityofAmericansexpressedsupportforarangeofAIbasedpolicies,including

requiringAIgeneratedcontenttobeclearlymarked,makingAIcompanieslegallyliable,restrictingtheexportofadvancedAItechnology,andrequiringAIcompaniestodisclosedetailsabouttheirmodelsandtheirtraining.

●TheonlypolicieswetestedwhichmorepeoplesawasabadideathanagoodideawereintroducinganewtaxontheuseofAImodels,increasinggovernmentfundingofAIresearchandbanningnewresearchintoAI.

●WhenaskedtomakeaforcedchoicebetweentheUSprioritizingstayingatthetechnologicalfrontierorresponsibledevelopment,evenifthismeantothercountrieslikeChinatakingthelead,wesawmuchlessagreement.43%ofAmericansprioritizedstayingatthefrontier,comparedto35%prioritizingresponsibledevelopment-withthisdisagreementcrossingacrossdemographicsandpoliticalparties.

4

Foreword

SincethedebutofChatGPTin2023,therehasbeenasurgeofinterestinartificialintelligence(AI).

Thepossibilitiesofthistechnologyareimmense.Imagineeverystudenthavingaccesstoanempathetictutor,everypatientbeingabletoinstantlycontactamedicalexpert,andeverytravelerhavingareal-timetranslatorattheirfingertips.AsAIbecomesmorehelpful,intuitive,andreliable,thetechnologywillhaveaprofoundimpactonvirtuallyeveryareaoftheeconomyandsociety,fromhowpeopleworkandlearntohowtheyplayandsocialize.Inadditiontoboostingproductivityandcompetitiveness,therebyraisingstandardsofliving,AIhasthepotentialtoaddressmajorglobalchallenges,suchaspublichealth,cleanenergy,andenvironmentalconcerns.

Yetdespitesomanybrightspotsonthehorizon,itistemptingtofearchange.SomeworryabouthowAImayaffecttheirlivesorlivelihoodsoragonizeovertheriskthatAImayunleashbiasandmisinformation.Othersseeelementsoncerelegatedtosciencefiction,suchasself-drivingvehicles,virtualcompanions,andintelligentagents,becomingpartoftheirworldandfearthatthedarkestelementsofthesetales—machinesdestroyinghumancivilization—mayalsocometopass.

ItiswiththisbifurcatedvisionofthefuturethatpolicymakersmustconfrontthekeyquestionofhowtheUnitedStatesshouldrespondtotheriseofAI.Shoulditprioritizeacceleratinginnovationtomaximizebenefitsorslowingdowninnovationtominimizerisks?Indeed,thequestionisnotmerelyrhetoricalashundredsofexpertsissuedademandlastyearforasix-monthpauseonAIresearch.Inthisenvironment,publicopiniononAIiscrucial.Electedofficialsarechosentorepresentthewillofthevoters,andtheirpoliciesandprioritiestendtoreflectpublicsentiment.Atthesametime,electedofficialsarechosentoleadandmakeinformeddecisionsbasedontheirconstituents’long-terminterests.

ThispollshowsAmericanshaveamultitudeofviewsaboutAI.Theyarecurious,interested,worried,andamazed.Theythinkthetechnologywillhelpadvancescience,makejobslessmundane,andimprovehealthcare,buttheyalsoworrythatitcouldhelpcreatemisinformation,enablehacking,andcausejoblosses.Overall,theyaredividedbetweenAIoptimistswhobelievethetechnologywillmakethingsbetter,andAIpessimistswhothinkitwillmakethingsworse.Andmanyhavenotyetmadeuptheirmind.

ToremainagloballeaderinAI,theUnitedStatesfacestwobigchallenges.First,itmuststayonthefrontierofAIdevelopment,whichrequiresmaintainingarobustecosystemofAIskills,chips,datacenters,andmodels.Second,itmustleadinAIadoption,especiallyintradesectorsoftheeconomywhereitfacesglobalcompetition,andareaslikeeducationandgovernmentwheretherearesignificantrewards.Botheffortswillrequireextensivecoordinationandcooperationbetweenthepublicandprivatesectors,aswellasaregulatoryenvironmentthatfostersresponsibleinnovation.

Policymakershavetheirworkcutoutforthem,notonlytodesigntherightpoliciesforAIthatkeepstheUnitedStatesoncourseintheglobalAIrace,butalsotobuildpublicsupportfortheseinitiatives.ThistaskisespeciallychallengingbecauseAI,andthecompaniesthatmake

6

it,areregularlyvilifiedinthemedia.AsmoreAmericansuseandexperiencethebenefitsofAIandrealizethattheirworstfearshavenotcometopass,hopefullypublicsupportwillfollow.

-DanielCastro,Director,CenterforDataInnovation

7

Today

。

HowdopeoplefeelaboutAI?

In1950,mathematicianandcomputerscientistAlanTuringproposedtheimitationtest:atestforintelligentbehaviorinamachineduringwhichahumanevaluatorwouldhavetodecidewhethertheentitytheyweretalkingtoinatextchatwasanotherhuman,oranAI.

WehavenotreachedthepointwhereAIscanreliablypassaTuringtest.Butwearegettingcloser:43%ofAmericanstoldusthattheywouldn’tbeconfidentthattheycouldtellwhetherachatwaswithanAIornotinlessthanaminute.Nearlyathirdsaidthattheywouldn’tbeconfidentthattheycouldtellinunder10minutes-andforAmericansover55,itwasamajority.

Overall,51%ofAmericanssaidthatAIwasdevelopingfasterthantheyexpected.That’suparoundafifthrelativelycomparedtowhenweaskedthesamequestionlastyear.

Overall,whenaskedaboutthenearestpotentialhistoricalcomparatortoAIasatechnology,AmericanspointtothecomputerortheInternet.Theydon’tseeitastransformativeastheprintingpressorelectricityassomepeoplebelieve-buttheyalsoseeitaslikelytobesignificantlymoreimportantthan,say,socialmediaonitsown.

Thisrapiddevelopmenthasledtomixedfeelings.Aswithlastyear,whenaskedtodescribehowtheyfeelaboutAI,themostcommonlychosenemotionbyAmericanswascuriosity-withamixofpositiveandnegativeemotionsafterthat.

9

Whatwasnoticeablewasthatnegativeemotionshaveslightlytickedupcomparedtolastyear,withAmericansagainaroundafifthmorelikelyrelativelycomparedtolastyeartochoosethem.

Equally,inanewquestion,whenweaskedaboutAmericans'expectationsofthefutureimpactofAIwesawamixedpicture.Onaverage,acrossarangeofpersonalandsocietalcategories,wesawthatAmericansweremoderatelymorelikelytohavepositiveexpectationsthannegativeones-butthisgapneverexceeded10percentagepoints,andasignificantproportionwereunsure,particularlyasitaffectedthempersonally.

10

HowmuchdopeopleunderstandhowAIworks?

Insomeways,AIismoreintuitivethanothertechnologies:oftenthebestwaytointeractwithitistotalktoithowyouwouldanotherhuman.Inotherways,itisverycomplex,andeventheworld’sleadingAIexpertstodaydonotfullyunderstandhowatransformermodelworksthewayitdoes.

Inourpolling,onlyaroundathird(32%)ofAmericanssaidthattheywereconfidenttheycouldexplainhowmodernAImodelsworked-andwhenwepushedonthisfurtherbyaskingaroundarangeoftermsrelatedtoAI,weunsurprisinglysawverylowlevelsofawareness.

11

Morerelevantlyperhaps,wealsosawonlymixedawarenessoftherelativestrengthsandweaknessesoftoday’smodels.ManypeopleinourpollseemedtothinkofAImodelsashavingthetraditionalsstrengthsandweaknessesofacomputer:goodatmathandwithaperfectmemory,butweakatcommonsensereasoningandsoundingempathetic.Inpractice,thisisalmosttheoppositeofthestrengthsandweaknessesoftoday’sLLMbasedmodels.

12

HowareAmericansusingAItoday?

Puttingasideabstractimpressions,howmuchareAmericansconcretelyawareandusingAItoolsthatareavailabletoday?

Inourpolling,thehighestawarenesswasforexistingAItoolsthathavebeenaroundforalongtime:AmazonAlexa,GoogleAssistantandApple’sSiri.

Thatsaid,ChatGPTwasnotfarbehindthebigthree-andcomparedtolastyear’spoll,awarenessofOpenAI’stoolhadnearlydoubled.

13

ForChatGPT,wecanalsocompareusageyearonyear-withtheproportionwhosaytheyhaveuseditmultipletimesincreasingfrom20%to35%.

Overallthough,whileawarenessmaybehigh,thishasnotyetnecessarilyturnedintoregularusageforeveryone.Inourpolling,just13%ofAmericansselfdescribeasusingoneoftheLLMbasedchatbotsregularly-withasignificantgradientacrossbothageandgender.

14

Inourpolling,wesawsomeevidencethatpeoplearecontinuingtodiscoverthesetoolsthough,suggestingthatusagemaycontinuetogrowreasonablyfast.Onaverage,over40%ofusersofthetoolssaidtheyhadonlystartedusingtheminthelast3months.

Similarly,forthosewhoareusingthesetools,thepictureseemstobethatwhiletheirusersfindthemoverwhelminglyhelpful,onlyaroundathirdofregularusersseethemascurrentlyessential:

83%

33%

45%

ofAmericansusingLLMbasedchatbots

ofAmericansusingLLMbasedchatbots

ofAmericansusingLLMbasedchatbots

saytheyfindthemhelpful

saytheyhavebecomean

essentialtooltheyuse

regularly

saytheyusethemfrom

timetotime,butwouldnot

missitifitdidn’texist

Whenweaskedwhatusecasespeoplehadtried,themostcommonwastohelpexplainsomething-witharoundtwothirdsofuserssayingtheyhaddonethis.Afterthat,aroundahalfofuserssaidthattheyusedthemtohelpbrainstormideasorwritetext.

15

16

Opportunities

17

ThePotentialBenefitsofAI

AIislikelytobeoneofthemostsignificanteconomicdriversinthenexttwentyyears.TheIMFthisyearestimatedthatAIcouldboostproductivityinanadvancedeconomyliketheUKby1.5%1,similartopredictionslastyearbyGoldmanSachsfortheUS.2

Inourpolling,whenweaskedaboutthepotentialbenefitsfromAIwesawaninterestingdichotomy:whilethemostwidelyrecognisedbenefitswereacceleratingscientificadvancementandincreasingproductivityacrosstheeconomy,respondentsweremuchlesslikelytobelievethatthiswouldtranslateintoincreasedwagesforworkers,withthisbeingtheleastpopularchoice.,

Whenitcametopersonalusecases,however,wesawawidespreadinterestinatleastgivingAIatryinavarietyofroles:frombasicresearchtogivingearlywarningofanewmedicalcondition.

1

/en/Publications/WEO/Issues/2024/04/16/world-economic-outlook-april-2

024

2

/intelligence/pages/ai-may-start-to-boost-us-gdp-in-2027.

html

18

2024islikelytobeayearwherethereisanincreasedfocusonthecreationofagents.Agentsaredesignednottojusttobeabletoanswerquestions,buttoactuallycarryoutbasictasksforyou.BothOpenAI3andGoogle4havebeenexplicitthatthisisthenextleapforward.

Inourpolling,weaskedaboutarangeofpotentialAIusecases,fromactingasapersonalassistanttoactingasavirtualworkoutcoach.Overall,wesawmorecautionherethanthemoregenericAIusecasesabove-althoughyoungeradultsweremorepreparedtogiveAIagentsago.AmajorityofAmericansunder35saidtheywouldatleasttryanAIpersonaltutor,personalassistant,workoutcoachorfinancialadvisor.

3

/2024/05/01/1091979/sam-altman-says-helpful-agents-ar

e-poised-to-become-ais-killer-function/

4

https://blog.google/intl/en-africa/products/explore-get-answers/google-io-2024-an-io-for-a

-new-generation/

19

OneofthemorecontroversialAIagentusecaseshasbeenusingthemforcompanionship,asavirtualfriendorevenboyfriendandgirlfriend.In2023,AIcompanionappReplikawasalreadyreportingover10millionusersworldwide.5

Inourpolling,only20%ofAmericanssaidthattheywereinterested-butwealsosawabigdifferenceacrossdemographics.43%ofAmericanmenunder35saythattheywouldbeinterestedinanAIgirlfriendorboyfriend.

5

/faculty/Pages/item.aspx?num=63508

20

AtWork

ThecurrentwaveofAIhypewaslargelydrivenbythearrivalofChatGPT-buttowhatextentarepeopleactuallyusingLLMbasedchatbotslikeitatwork?

Inourpoll,justoveraquarterofAmericanworkerstoldusthattheyhadusedachatbotatwork-butabouttwo-thirdsofthosewhohadusedthemsaidtheyfoundthemhelpfulorveryhelpful.

28%

68%

38%

ofAmericanworkerssaythat

ofAmericanworkersusing

LLMbasedchatbotssaythat

theyfindthem

ofAmericanworkersusingLLMbasedchatbotssaythat

HaveusedanLLMchatbottoolatwork

helpfulorveryhelpful

theyhavebecomean

essentialtooltheyuse

regularly

Aretheyjustexperimentingwiththesetoolshowever,orhavetheyalreadybecomeanormalpartofpeople’sworkflowinthesamewayasemailorsearch?

21

Overall,around19%ofworkerssaidtheywereusingthesetoolsregularly,withthemajorityofthoseusingthematleastmultipletimesaweek.Wesawabigageandgendergradienthere,however,withyoungmensignificantlymorelikelytobesayingtheywereusingthesetoolsregularly.

ThoseworkerswhoarealreadyusingAItoolsseemtobeclassicearlyadopters:aroundhalfofthemsaidtheyhadrespectivelydecidedtousethesetoolsontheirown,workedouthowtousethemandsaytheylearnbestfromexploringandexperimentingthemselves.

48%

53%

46%

ofAmericanworkersusing

LLMbasedchatbotssaythat

they

ofAmericanworkersusing

LLMbasedchatbotssaythat

they

ofAmericanworkersusingLLMbasedchatbotssaythat

workedouthowtousethoseAItoolsthemselves

decidedtousethoseAItoolsthemselves

theylearnbestfrom

exploringand

experimentingwithAItools

themselves

Whileoverallmoreworkerswantedtoteachthemselvesratherthanhaveformaltraining,thiswasnottrueforeveryone:workersover55yearsoldweresignificantlymorelikelytoexpressadesireforAIskillstraining.

22

Health

Alongsidetheeconomy,oneofthemostsignificantotheropportunitiesfromAIistospeedupthediagnosisandtreatmentofhealthconditions.

Giventhemanysensitivitiesinthisspace,whenfirstaskedAmericansareunderstandablyunsureaboutusingAItodiagnose,withjustoverathird(37%)sayingtheywouldsupportthis.

37%

32%

32%

ofAmericans

ofAmericanssaythat

ofAmericanssaythat

SupportusingAItodiagnosepatients

OpposeusingAItodiagnosepatients

unsure

Whenwedugmoreintothereasonsforthiscaution,themostsignificantreasonswereconcernsovertheAIsystem’sreliabilitywith73%sayingtheyworriedtheAIsystemwouldgiveincorrectdiagnoses.

Withbasicprotectionsinplacehowever,wesawthatitwaspossibletoovercomeconcernsaboutAIdiagnosis.

23

Forexample,70%ofAmericanssaytheywouldsupportAIdiagnosisifitwasdoublecheckedbyahumandoctorandjustgivingpeoplethechoicewhethertheyuseditornotincreasedsupportto61%.Theonlysituationherewherewesawstrongoppositionwasinascenariowherepatientswouldbeoutrightforcedtousethesystem,two-thirdsofAmericansopposedtothis.

24

Concerns

25

WhichrisksdoAmericansworryabout?

Eversincesciencefictionwritersfirstconceivedtheideaofanartificialintelligence,wehavebeeninundatedwithstoriesaboutthemanywaystheycangowrong.

Giventhis,perhapsunsurprisingly,wesawareasonablyhighlevelofself-reportedfamiliarityacrossarangeofrisks,withthemostcommonbeingthepotentialforunemployment.Interestingly,onepotentialriskthatdidnotseemtohavecutthroughyetwasthepotentialfor

AItosignificantlyincreaseelectricityconsumption,withonlyaroundathirdofAmericanssayingtheywerefamiliarwiththis.

Whethertheywerefamiliarwithrisksornot,howworriedweretheyaboutthem?

Whenwechangedthequestiontothis,wesawconcernsfromunemploymentfallslightlydownthelist,whilemisinformationwenttothetop-perhapsreflectingitsperceivedgreaterurgency.

26

Manyoftheseworrieswererelativelynonpartisan.BothRepublicansandDemocratswerealmostequallyworriedaboutthepotentialofAIdrivenunemploymentormisinformation.Republicans,however,weremorelikelytobeworriedaboutthepotentialforAItobebiasedagainstpeoplewithdifferentpoliticalviews,orbeingusedbycriminalsorforGovernmentsurveillance.Democrats,bycontrast,hadgreaterconcernsoverpersonaldeepfakes,increasedelectricityconsumptionandbiasagainstmarginalizedgroups.

27

WhileAImaycreatesomerisk,howmuchistheriskitcreatesadditionaltotherisksthatalreadyexist?Afterall,somehaveargued,ifyouwantto,youcanalreadycreatefakeimagesofsomeone,whilenewAIdriventechnologiesmayactuallyhelpreduceriskinmanyoftheseareas.

Inourpolling,Americanswererelativelyunconvincedbythisargument:acrosstherangeofriskswepresentedthem,fromembarrassingvideostohumanextinction,theyseemedtobelievethatAIrepresentedasignificantincreaseinrisk.

28

MisinformationandDeepfakes

2024isanelectionyear:bothfortheUSandacrosstheworld,withmorethan2billionpeopleexpectedtovotethisyear.6Giventhat,unsurprisingly,62%ofAmericansareworriedaboutthepotentialimpactofmisinformationontheUSelection-withthelevelofworryslightlyhigheramongRepublicans.

6

/agenda/2023/12/2024-elections-around-world/

29

Perhapsunsurprisingly,votersweremorelikelytothinkthattheothersidewouldbenefitmostfrommisinformation:BidensupportersthoughtTrumpwouldbehelpedmost,andTrumpsupportersthoughtBidenwould.

30

While71%wereworriedAIgeneratedcontentwouldbeusedtomanipulateanelection,thiswasactuallyjustoneoftheirworries:withequalconcernsaroundthepotentialforAItoconpeopleoutoftheirmoneyorcreatesexuallyexplicitdeepfakeswithoutconsent.

Similarly,thegroupsthatourrespondentsweremoreworriedaboutusingAItomanipulatepeoplewerenotdomesticpoliticiansbutcriminals,terroristsandforeigngovernments.

31

WhenaskedtochoosewhethertheythoughtAIgeneratedcontentwouldmaketheproblemofmisinformationworse-orthatpeoplecouldalreadymakefakecontentiftheywantedto,andAIwon’tchangethis-amajoritysaidthattheythoughtAIgeneratedcontentwouldmaketheproblemsofmisinformationsignificantlyworse.

OnereasonmanypeoplearelikelyconcernedisthataroundahalfofAmericans(46%)werenotconfidentthattheycoulddetectfakeAIgeneratedcontentontheInternet-withconfidencesignificantlyfallingtheolderourrespondentgot.

Overalltherewasstrongsupportforbetterlabeling,with62%ofAmericanssaythatGovernmentsandcompaniesneedtodomuchmoretobetterlabelandrestrictmisleadingAIgeneratedcontent.

Whenweaskedindirectly,itwastheelderlyandchildrenthatweremostseentobeatriskofbeingmisled,with57%and51%pointingtothemrespectively.Only2%ofrespondentssaidthattherewasnobodythattheywereworriedmightbemisled.

CouldAItoolsbepartofthesolutiontomisinformation,helpingtospotandcounterit?Atpresent,itseemsAmericansneedmoreconvincingonthispoint:just38%ofAmericanssaidthattheythinkitislikelythatnewAItoolscouldhelpreducemisinformation.

32

Automation

Newtechnologieshavealwayschangedthestructureoftheeconomy-butoneofthemoreunusualthingsaboutAIisthatthereissignificantlymoreuncertaintyaboutwhoitislikelytoaffectandhow.

Whenweaskedpeopletogiveascoreoutof