英國人工智能白皮書 A pro-innovation approach to AI regulation 20230330

Apro-innovationapproachtoAI

regulation

Ministerialforeword

Ibelievethatacommon-sense,outcomes-orientedapproachisthebestwaytogetrighttotheheartofdeliveringontheprioritiesofpeopleacrosstheUK.Betterpublicservices,highqualityjobsandopportunitiestolearntheskillsthatwillpowerourfuture–thesearetheprioritiesthatwilldriveourgoaltobecomeascienceandtechnologysuperpowerby2030.

ArtificialIntelligence(AI)willplayacentralpartindeliveringandenablingthesegoals,andthiswhitepaperwillensureweareputtingtheUKoncoursetobethebestplaceintheworldtobuild,testanduseAItechnology.Butwearenotstartingfromzero.Havinginvestedover￡2.5billioninAIsince2014,thispaper

buildsonourrecentannouncementsof￡110millionfor

our

AITechMissionsFund,

￡900milliontoestablisha

newAIResearchResourceandtodevelopanexascalesupercomputercapableofrunninglargeAImodels–backedupbyournew￡8millionAIGlobalTalentNetworkand￡117million

ofexistingfundingtocreatehundredsof

newPhDsforAIresearchers.

MostofusareonlynowbeginningtounderstandthetransformativepotentialofAIasthetechnologyrapidlyimproves.Butinmanyways,AIisalreadydeliveringfantasticsocialandeconomicbenefitsforrealpeople–fromimprovingNHSmedicalcaretomakingtransportsafer.RecentadvancesinthingslikegenerativeAIgiveusaglimpseintotheenormousopportunitiesthatawaitusinthenearfutureifwearepreparedtoleadtheworldintheAIsectorwithourvaluesoftransparency,accountabilityandinnovation.

MyvisionforanAI-enabledcountryisonewhereourNHSheroesareabletosavelivesusingAItechnologiesthatwereunimaginablejustafewdecadesago.Iwantourpolice,transportnetworksandclimatescientistsandmanymoretobeempoweredbyAItechnologiesthatwillmaketheUKthesmartest,healthiest,safestandhappiestplacetoliveandwork.ThatiswhyAIisoneofthisgovernment’s5technologiesoftomorrow–bringingstrongergrowth,betterjobs,andboldnewdiscoveries.ItisavisionthathasbeenshapedbystakeholdersandexpertsinAI,

whoseexpertiseandideasIamdeterminedtoseereflectedinourdepartment.

TheUKhasbeenattheforefrontofthisprogress,placingthirdintheworldforAIresearchanddevelopment.WearehometoathirdofEurope’stotalAIcompaniesandtwiceasmanyasanyotherEuropeancountry.Ourworld-leadingstatusisdowntoourthrivingresearchbaseandthepipelineofexpertisegraduatingthroughouruniversities,theingenuityofourinnovatorsandthegovernment’slong-termcommitmenttoinvestinAI.

ToensurewebecomeanAIsuperpower,though,itiscrucialthatwedoallwecantocreatetherightenvironmenttoharnessthebenefitsofAIandremainattheforefrontoftechnologicaldevelopments.ThatincludesgettingregulationrightsothatinnovatorscanthriveandtherisksposedbyAIcanbeaddressed.

Theseriskscouldincludeanythingfromphysicalharm,anunderminingofnationalsecurity,aswellasriskstomentalhealth.ThedevelopmentanddeploymentofAIcanalsopresentethicalchallengeswhichdonotalwayshaveclearanswers.Unlessweact,householdconsumers,publicservicesandbusinesseswillnottrustthetechnologyandwillbenervousaboutadoptingit.

Unlesswebuildpublictrust,wewillmissoutonmanyofthebenefitsonoffer.

Indeed,thepaceofchangeitselfcanbeunsettling.SomefearafutureinwhichAIreplacesordisplacesjobs,forexample.OurwhitepaperandourvisionforafutureAI-enabledcountryisoneinwhichourwaysofworkingarecomplementedbyAIratherthandisruptedbyit.Inthemodernworld,toomuchofourprofessionallivesaretakenupbymonotonoustasks–inputtingdata,fillingoutpaperwork,scanningthroughdocumentsforonepieceofinformationandsoon.AIintheworkplacehasthepotentialtofreeusupfromthesetasks,allowingustospendmoretimedoingthethingswetrainedfor–teacherswithmoretimetoteach,clinicianswithmoretimetospendwithpatients,policeofficerswithmoretimeonthebeatratherthanbehindadesk–thelistgoeson.

Indeed,sinceAIisalreadyinourday-to-daylives,therearenumerousexamplesthatcanhelptoillustratethereal,tangiblebenefitsthatAIcanbringonceanyrisksaremitigated.StreamingservicesalreadyuseadvancedAItorecommendTVshowsandfilmstous.OursatnavusesAItoplotthefastestroutesforourjourneys,orhelpsusavoidtrafficbyintelligentlypredictingwhere

congestionwillbeonourjourney.Andofcourse,almostallofuscarryasmartphoneinourpocketsthatusesadvancedAIinallsortsofways.Thesecommondevicesallcarriedrisksatonetimeoranother,buttodaytheybenefitusenormously.

Thatiswhyourwhitepaperdetailshowweintendtosupportinnovationwhileprovidingaframeworktoensurerisksareidentifiedandaddressed.However,aheavy-handedandrigidapproachcanstifleinnovationandslowAIadoption.Thatiswhywesetoutaproportionateandpro-innovationregulatoryframework.Ratherthantargetspecifictechnologies,itfocusesonthecontextinwhichAIisdeployed.Thisenablesustotakeabalancedapproachtoweighingupthebenefitsversusthepotentialrisks.

WerecognisethatparticularAItechnologies,foundationmodelsforexample,canbeappliedinmanydifferentwaysandthismeanstheriskscanvaryhugely.Forexample,usingachatbottoproduceasummaryofalongarticlepresentsverydifferentriskstousingthesametechnologytoprovidemedicaladvice.WeunderstandtheneedtomonitorthesedevelopmentsinpartnershipwithinnovatorswhilealsoavoidingplacingunnecessaryregulatoryburdensonthosedeployingAI.

Toensureourregulatoryframeworkiseffective,wewillleveragetheexpertiseofourworldclassregulators.TheyunderstandtherisksintheirsectorsandarebestplacedtotakeaproportionateapproachtoregulatingAI.Thiswillmeansupportinginnovationandworkingcloselywithbusiness,butalsosteppingintoaddressriskswhennecessary.Byunderpinningtheframeworkwithasetofprinciples,wewilldriveconsistencyacrossregulatorswhilealsoprovidingthemwiththeflexibilityneeded.

Forinnovatorsworkingatthecuttingedgeanddevelopingnoveltechnologies,navigatingregulatoryregimescanbechallenging.That’swhyweareconfirmingourcommitmenttotakingforwardakeyrecommendationmadebySirPatrickVallancetoestablisharegulatorysandboxforAI.Thiswillbringtogetherregulatorstosupportinnovatorsdirectlyandhelpthemgettheirproductstomarket.Thesandboxwillalsoenableustounderstandhowregulationinteractswithnewtechnologiesandrefinethisinteractionwherenecessary.

HavingexitedtheEuropeanUnionwearefreetoestablisharegulatoryapproachthatenablesustoestablishtheUKasanAIsuperpower.Itisanapproachthatwillactivelysupportinnovationwhileaddressingrisksandpublicconcerns.TheUKis

hometothrivingstart-ups,whichourframeworkwillsupporttoscale-upandcompeteinternationally.Ourpro-innovationapproachwillalsoactasastrongincentivewhenitcomestoAIbusinessesbasedoverseasestablishingapresenceintheUK.Thewhitepapersetsoutourcommitmenttoengaginginternationallytosupportinteroperabilityacrossdifferentregulatoryregimes.NotonlywillthiseasetheburdenonbusinessbutitwillalsoallowustoembedourvaluesasglobalapproachestogoverningAIdevelop.

Ourapproachreliesoncollaborationbetweengovernment,regulatorsandbusiness.Initially,wedonotintendtointroducenewlegislation.Byrushingtolegislatetooearly,wewouldriskplacingundueburdensonbusinesses.Butalongsideempoweringregulatorstotakealead,wearealsosettingexpectations.Ournewmonitoringfunctionswillprovidearealtimeassessmentofhowtheregulatoryframeworkisperformingsothatwecanbeconfidentthatitisproportionate.Thepaceoftechnologicaldevelopmentalsomeansthatweneedtounderstandnewandemergingrisks,engagingwithexpertstoensurewetakeactionwherenecessary.Acriticalcomponentofthisactivitywillbeengagingwiththepublictounderstandtheir

expectations,raisingawarenessofthepotentialofAIanddemonstratingthatwearerespondingtoconcerns.

Theframeworksetoutinthiswhitepaperisdeliberatelydesignedtobeflexible.Asthetechnologyevolves,ourregulatoryapproachmayalsoneedtoadjust.Ourprinciples-basedapproach,withcentralfunctionstomonitoranddrivecollaboration,willenableustoadaptasneededwhileprovidingindustrywiththeclarityneededtoinnovate.Wewillcontinuetodevelopourapproach,buildingonourcommitmenttomakingtheUKthebestplaceintheworldtobeabusinessdevelopingandusingAI.Responsestotheconsultationwillinformhowwedeveloptheregulatoryframework–Iencourageallofthosewithaninteresttorespond.

Executivesummary

Artificialintelligence–theopportunityand

thechallenge

1.Artificialintelligence(AI)isalreadydeliveringwidesocietalbenefits,frommedicaladvances

[footnote1]

tomitigatingclimate

change

.[footnote2]

Forexample,anAItechnologydevelopedby

DeepMind,aUK-basedbusiness,cannowpredictthestructure

ofalmosteveryproteinknowntoscience

.[footnote3]

This

breakthroughwillacceleratescientificresearchandthedevelopmentoflife-savingmedicines-ithasalreadyhelpedscientiststomakehugeprogressincombatingmalaria,antibioticresistance,andplasticwaste.

2.TheUKScienceandTechnologyFramework

[footnote4]

setsoutgovernment’sstrategicvisionandidentifiesAIasoneof5criticaltechnologies.TheframeworknotestheroleofregulationincreatingtheenvironmentforAItoflourish.WeknowthatwehaveyettoseeAItechnologiesreachtheirfullpotential.Undertherightconditions,AIwilltransformallareasoflife

[footnote5]

andstimulatetheUKeconomybyunleashinginnovationanddrivingproductivity

,[footnote6]

creatingnewjobsandimprovingtheworkplace.

3.Acrosstheworld,countriesandregionsarebeginningtodrafttherulesforAI.TheUKneedstoactquicklytocontinuetoleadtheinternationalconversationonAIgovernanceanddemonstratethevalueofourpragmatic,proportionateregulatoryapproach.TheneedtoactwashighlightedbySirPatrickVallanceinhisrecentRegulationforInnovationreview.Thereportidentifiesthe

shorttimeframeforgovernmentinterventiontoprovideaclear,pro-innovationregulatoryenvironmentinordertomaketheUK

oneofthetopplacesintheworldtobuild

foundationalAIcompanies

.[footnote7]

4.Whileweshouldcapitaliseonthebenefitsofthesetechnologies,weshouldalsonotoverlookthenewrisksthatmayarisefromtheiruse,northeuneasethatthecomplexity

ofAItechnologiescanproduceinthewiderpublic.Wealready

knowthatsomeusesofAIcoulddamageourphysica

l[footnote8]

andmentalhealth,

[footnote9]

infringeontheprivacyofindividuals

[footnote

10]

andunderminehumanrights

.[footnote11]

5.PublictrustinAIwillbeunderminedunlesstheserisks,andwiderconcernsaboutthepotentialforbiasanddiscrimination,areaddressed.Bybuildingtrust,wecanacceleratetheadoptionofAIacrosstheUKtomaximisetheeconomicandsocialbenefitsthatthetechnologycandeliver,whileattractinginvestmentandstimulatingthecreationofhigh-skilledAIjobs

.[footnote12]

InordertomaintaintheUK’spositionasaglobalAIleader,weneedtoensurethatthepubliccontinuestoseehowthebenefitsofAIcanoutweightherisks

.[footnote13]

6.Respondingtoriskandbuildingpublictrustareimportantdriversforregulation.Butclearandconsistentregulationcanalsosupportbusinessinvestmentandbuildconfidenceininnovation.Throughoutourextensiveengagement,industryrepeatedlyemphasisedthatconsumertrustiskeytothesuccessofinnovationeconomies.Wethereforeneedaclear,proportionateapproachtoregulationthatenablestheresponsibleapplicationofAItoflourish.InsteadofcreatingcumbersomerulesapplyingtoallAItechnologies,ourframeworkensuresthatregulatorymeasuresareproportionatetocontextandoutcomes,byfocusingontheuseofAIratherthanthetechnologyitself.

7.PeopleandorganisationsdevelopanduseAIintheUKwithintherulessetbyourexistinglaws,informedbystandards,guidanceandothertools.ButAIisageneralpurposetechnologyanditsusescancutacrossregulatoryremits.Asaresult,AItechnologiesarecurrentlyregulatedthroughacomplexpatchworkoflegalrequirements.Weareconcernedbyfeedbackfromacrossindustrythattheabsenceofcross-cuttingAIregulationcreatesuncertaintyandinconsistencywhichcanunderminebusinessandconsumerconfidenceinAI,andstifleinnovation.Byprovidingaclearandunifiedapproachtoregulation,ourframeworkwillbuildpublicconfidence,makingit

clearthatAItechnologiesaresubjecttocross-cutting,principles-basedregulation.

Ourpro-innovationframework

8.ThegovernmentwillputinplaceanewframeworktobringclarityandcoherencetotheAIregulatorylandscape.Thisregimeisdesignedtomakeresponsibleinnovationeasier.ItwillstrengthentheUK’spositionasagloballeaderinAI,harnessAI’sabilitytodrivegrowthandprosperity

,[footnote14]

andincreasepublictrustinitsuseandapplication.

9.Wearetakingadeliberatelyagileanditerativeapproach,recognisingthespeedatwhichthesetechnologiesareevolving.Ourframeworkisdesignedtobuildtheevidencebasesothatwecanlearnfromexperienceandcontinuouslyadapttodevelopthebestpossibleregulatoryregime.Industryhaspraisedourpragmaticandproportionateapproach.

10.Ourframeworkisunderpinnedby5principlestoguideandinformtheresponsibledevelopmentanduseofAIinallsectorsoftheeconomy:

?Safety,securityandrobustness

?Appropriatetransparencyandexplainability

?Fairness

?Accountabilityandgovernance

?Contestabilityandredress

11.Wewillnotputtheseprinciplesonastatutoryfootinginitially.NewrigidandonerouslegislativerequirementsonbusinessescouldholdbackAIinnovationandreduceourabilitytorespondquicklyandinaproportionatewaytofuturetechnologicaladvances.Instead,theprincipleswillbeissuedonanon-statutorybasisandimplementedbyexistingregulators.Thisapproachmakesuseofregulators’domain-specificexpertisetotailortheimplementationoftheprinciplestothespecificcontextinwhichAIisused.Duringtheinitialperiodofimplementation,wewillcontinuetocollaboratewithregulatorstoidentifyanybarrierstotheproportionateapplicationoftheprinciples,andevaluatewhetherthenon-statutoryframeworkishavingthedesiredeffect.

12.Followingthisinitialperiodofimplementation,andwhenparliamentarytimeallows,weanticipateintroducingastatutorydutyonregulatorsrequiringthemtohavedueregardtotheprinciples.Somefeedbackfromregulators,industryand

academiasuggestedweshouldimplementfurthermeasurestosupporttheenforcementoftheframework.Adutyrequiringregulatorstohaveregardtotheprinciplesshouldallowregulatorstheflexibilitytoexercisejudgementwhenapplyingtheprinciplesinparticularcontexts,whilealsostrengtheningtheirmandatetoimplementthem.Inlinewithourproposaltoworkcollaborativelywithregulatorsandtakeanadaptableapproach,wewillnotmovetointroducesuchastatutorydutyifourmonitoringoftheframeworkshowsthatimplementationiseffectivewithouttheneedtolegislate.

13.Inthe2022

AIregulationpolicypaper,

[footnote15]

weproposedasmallcoordinationlayerwithintheregulatoryarchitecture.

IndustryandcivilsocietyweresupportiveofourintentiontoensurecoherenceacrosstheAIregulatoryframework.However,feedbackoftenarguedstronglyforgreatercentralcoordinationtosupportregulatorsonissuesrequiringcross-cuttingcollaborationandensurethattheoverallregulatoryframeworkfunctionsasintended.

14.Wehaveidentifiedanumberofcentralsupportfunctionsrequiredtomakesurethattheoverallframeworkoffersa

proportionatebuteffectiveresponsetoriskwhilepromotinginnovationacrosstheregulatorylandscape:

?Monitoringandevaluationoftheoverallregulatoryframework’seffectivenessandtheimplementationoftheprinciples,includingtheextenttowhichimplementationsupportsinnovation.Thiswillallowustoremainresponsiveandadapttheframeworkifnecessary,includingwhereitneedstobeadaptedtoremaineffectiveinthecontextofdevelopmentsinAI’scapabilitiesandthestateoftheart.

?AssessingandmonitoringrisksacrosstheeconomyarisingfromAI.

?Conductinghorizonscanningandgapanalysis,includingbyconveningindustry,toinformacoherentresponsetoemergingAItechnologytrends.

?SupportingtestbedsandsandboxinitiativestohelpAIinnovatorsgetnewtechnologiestomarket.

?Providingeducationandawarenesstogiveclaritytobusinessesandempowercitizenstomaketheirvoicesheardaspartoftheongoingiterationoftheframework.

?Promotinginteroperabilitywithinternationalregulatoryframeworks.

15.Thecentralsupportfunctionswillinitiallybeprovidedfromwithingovernmentbutwillleverageexistingactivitiesandexpertisefromacrossthebroadereconomy.TheactivitiesdescribedabovewillneitherreplacenorduplicatetheworkundertakenbyregulatorsandwillnotinvolvethecreationofanewAIregulator.

16.Ourproportionateapproachrecognisesthatregulationisnotalwaysthemosteffectivewaytosupportresponsibleinnovation.Theproposedframeworkisalignedwith,andsupplementedby,avarietyoftoolsfortrustworthyAI,suchasassurancetechniques,voluntaryguidanceandtechnicalstandards.Governmentwillpromotetheuseofsuchtools.Wearecollaboratingwithpartnerslikethe

UKAIStandardsHub

toensurethatouroverallgovernanceframeworkencouragesresponsibleAIinnovation(see

part4

fordetails).

17.Inkeepingwiththeglobalnatureofthesetechnologies,wewillalsocontinuetoworkwithinternationalpartnerstodeliverinteroperablemeasuresthatincentivisetheresponsibledesign,developmentandapplicationofAI.Duringourcallforviews,

industry,academiaandcivilsocietystressedthatinternationalalignmentshouldsupportUKbusinessestocapitaliseonglobalmarketsandprotectUKcitizensfromcross-borderharms.

18.TheUKisfrequentlyrankedthirdintheworldacrossarangeofmeasures,includinglevelofinvestment,innovationandimplementationofAI

.[footnote16]

TomaketheUKthemostattractiveplaceintheworldforAIinnovationandsupportUKcompanieswishingtoexportandattractinternationalinvestment,wemustensureinternationalcompatibilitybetweenapproaches.Countriesaroundtheworld,aswellasmultilateralforums,areexploringapproachestoregulatingAI.Thankstoourreputationfor

pragmaticregulation,theUKisrightlyseenbyinternationalpartnersasaleaderinthisglobalconversation.

Part1:Introduction

1.1Thepowerandpotentialofartificial

intelligence

19.AIisalreadydeliveringmajoradvancesandefficienciesinmanyareas.AIquietlyautomatesaspectsofoureveryday

activities,fromsystemsthatmonitortraffictomakeourcommutes

smoother

,[footnote17]

tothosethatdetectfraudinourbankaccounts

.[footnote18]

AIhasrevolutionisedlarge-scalesafety-critical

practicesinindustry,likecontrollingtheprocessofnuclear

fusion

.[footnote19]

Andithasalsobeenusedtoacceleratescientificadvancements,suchasthediscoveryofnewmedicine

[footnote20]

orthetechnologiesweneedtotackleclimatechange

.[footnote21]

20.Butthisisjustthebeginning.AIcanbeusedinahugevarietyofsettingsandhastheextraordinarypotentialtotransformoursocietyandeconomy

.[footnote22]

Itcouldhaveasmuchimpactaselectricityortheinternet,andhasbeenidentifiedasoneofthe5criticaltechnologiesintheUKScienceandTechnologyFramework

.[footnote23]

AsAIbecomesmorepowerful,andasinnovatorsexplorenewwaystouseit,wewillseemore

applicationsofAIemerge.Asaresult,AIhasahugepotentialto

drivegrowth

[footnote24]

andcreatejobs

.[footnote25]

Itwillsupportpeople

tocarryouttheirexistingjobs,byhelpingtoimproveworkforce

efficiencyandworkplacesafety

.[footnote26]

Toremainworldleaders

inAI,attractglobaltalentandcreatehigh-skilledjobsintheUK,wemustcreatearegulatoryenvironmentwheresuchinnovationcanthrive.

21.Technologicaladvanceslikelargelanguagemodels(LLMs)areanindicationofthetransformativedevelopmentsyettocome

.[footnote27]

LLMsprovidesubstantialopportunitiestotransform

theeconomyandsociety.Forexample,LLMscanautomatetheprocessofwritingcodeandfixingprogrammingbugs.Thetechnologycansupportgeneticmedicinebyidentifyinglinksbetweengeneticsequencesandmedicalconditions.Itcansupportpeopletoreviewandsummarisekeypointsfromlengthydocuments.Inthelast4years,LLMshavebeendeveloped

beyondexpectationsandtheyarebecomingapplicabletoan

increasinglywiderange

developmentofLLMand

3.3.3

below.

oftasks

.[footnote28]

Weexpandontheotherfoundationmodelsin

section

Box1.1:ExamplesofAIopportunities

AIhelpspiecetogetherthefirstcompleteimageofablackhole

AIcanenablescientificdiscovery.Acomputervisionmodelwasusedtopiecetogetherthefirsteverimageofablackhole55

millionlightyearsaway,combiningimagesfrom8telescopes

aroundtheworld

.[footnote29]

AIsolvesdecadesoldprotein-foldingpuzzle

AnAIcompanybasedintheUKtrainedneuralnetworkstopredictthestructuresofproteins,