規(guī)范機器學(xué)習(xí)：異構(gòu)性的挑戰(zhàn)

UniversityofPennsylvaniaCareyLawSchool

PennCareyLaw:LegalScholarshipRepository

AllFacultyScholarship

FacultyWorks

2-2023

RegulatingMachineLearning:TheChallengeofHeterogeneity

CaryCoglianese

UniversityofPennsylvaniaCareyLawSchool

AuthorORCIDIdentifier:

CaryCoglianese0000-0002-5496-2104

Followthisandadditionalworksat:

/faculty_scholarship

Partofthe

AdministrativeLawCommons

,

ArtificialIntelligenceandRoboticsCommons

,

Public

AdministrationCommons

,

ScienceandTechnologyStudiesCommons

,andthe

TheoryandAlgorithms

Commons

RepositoryCitation

Coglianese,Cary,"RegulatingMachineLearning:TheChallengeofHeterogeneity"(2023).AllFacultyScholarship.2921.

/faculty_scholarship/2921

ThisArticleisbroughttoyouforfreeandopenaccessbytheFacultyWorksatPennCareyLaw:LegalScholarshipRepository.IthasbeenacceptedforinclusioninAllFacultyScholarshipbyanauthorizedadministratorofPennCareyLaw:LegalScholarshipRepository.Formoreinformation,pleasecontact

biddlerepos@

.

REGULATINGMACHINELEARNING:

THECHALLENGEOFHETEROGENEITY

BY

CARYCOGLIANESE

EdwardB.ShilsProfessorofLawandProfessorofPoliticalScience,andDirector,PennProgramonRegulation,UniversityofPennsylvania.

1 ?2023CompetitionPolicyInternationalAllRightsReserved

Electroniccopyavailableat:/abstract=4368604

Electroniccopyavailableat:/abstract=4368604

Machine-learningalgorithmsincreasinglydrivetechnologi-caladvancesthatdelivervaluableimprovementsforsocietyandtheeconomy.Butthesealgorithmsalsoraiseimportantconcerns.Thewaymachine-learningalgorithmsworkau-tonomouslytofindpatternsinlargedatasetshasgivenrisetofearsofaworldthatwillultimatelycedecriticalaspectsofhumancontroltothedictatesofartificialintelligence.Thesefearsseemonlyexacerbatedbytheintrinsicopac-itysurroundinghowmachine-learningalgorithmsachievetheirresults.Toagreaterdegreethanwithotherstatisticaltools,theoutcomesgeneratedbymachinelearningcannotbeeasilyinterpretedandexplained,whichcanmakeithardforthepublictotrustthefairnessofproductsorprocessespoweredbythesealgorithms.

Forthesereasons,theautonomousandopaquequalitiesofmachine-learningalgorithmsmakethesedigitaltoolsbothdistinctiveandamatterofpublicconcern.Butwhenitcomestoregulatingmachinelearning,adifferentqualityofthesealgorithmsmattersmostofall:theirheterogeneity.TheMerriam-WebsterDictionarydefines“heterogeneity”as“thequalityorstateofconsistingofdissimilarordiverseelements.”Machinelearningalgorithms’heterogeneitywillmakeallthedifferenceindecidingwhentoregulatethem,whoshouldregulatethem,andhowtodesignregulationsimposedontheirdevelopmentanduse.

01

MACHINELEARNING’S

HETEROGENEITY

Oneofthemostimportantsourcesofmachinelearning’sheterogeneityderivesfromthehighlydiverseusestowhichitisput.Theseusescouldhardlyvarymorewidely.Con-siderjustasmallsampleofwaysthatdifferententitiesusemachine-learningalgorithms:

Socialmediaplatformsusethemtoselectandhighlightcontentforusers;

Hospitalradiologydepartmentsusethemtodetectcancerinpatients;

Creditcardcompaniesusethemtoidentifypoten-tialfraudulentcharges;

Commercialairlinesusethemtooperateaircraftwithauto-pilotingsystems;

Onlineretailersusethemtomakeproductrecom-mendationstovisitorstotheirwebsites;and

Politicalcampaignsusethemindecidingwhereandhowadvertise.

Evenwithinthesameorganizations,differentmachine-learningalgorithmscanperformdifferentfunctions.Anau-tomobilemanufacturer,forexample,mightuseonetypeofmachine-learningalgorithmtoautomatecertainon-roadoperationsoftheirvehicles,whileusingothermachine-learningalgorithmsaspartofitsmanufacturingprocessesorformanagingitssupplychainandinventory.

Inadditiontotheirvarieduses,machine-learningalgo-rithmscanthemselvestakemanydifferentformsandpos-sessdiversequalities.Thesealgorithmsareoftengroupedintoseveralmaincategories:supervisedlearning,unsuper-visedlearning,semi-supervisedlearning,andreinforcementlearning.Withineachcategory,therangeofalgorithmsandtheirformscanbehighlydiverse.Na?veBayesianmodels,decisiontrees,randomforests,andneuralnetworksarejustafewtypesofsupervisedlearningmodels.1Evenwithinanysingletype,finerpointsabouthoweachmodelgeneratedbyanalgorithmisstructured,nottomentiondifferencesinthedatausedtotrainit,canleadeachapplicationofma-chinelearningalmosttofallwithinacategoryofitsown.

Despitethewidevariationinalgorithms,italsoremainsthatthesamemachine-learningmodelcanbeputtodifferentuseswithinasingleorganization.Forexample,Meta—thecorporationthatownsFacebookandInstagram—hasnot-edthat,eventhoughits“imageclassificationmodelsarealldesignedtopredictwhat’sinagivenimage,theymaybeuseddifferentlyinanintegritysystemthatflagsharm-fulcontentversusarecommendersystemusedtoshowpeoplepoststheymightbeinterestedin.”2

Addedtotheextremevariationinusesanddesignsofal-gorithmsisthefactthat,formanyuses,multipledifferentalgorithmsareusedincombinationwitheachothertosup-portautomatedsystems.Whatmayattimesbereferredtoas“an”algorithmisoftenactuallyasuiteorfamilyofalgo-rithms,integratedintoanautomatedsystemorprocessinamannerdesignedtoperformaspecifiedfunction.Further-more,thesealgorithmsandtheircombinationsareupdatedandchangedovertime,asneworrefinedalgorithmsareshowntodobetter.Today’sChatGPT,forexample,runsonmodelsthataremarkedlydifferentthanearlierlanguagemodels,anditwillonlybeupdated,enhanced,andmodi-fiedrepeatedlyintheyearstocome.

Finally,thesechangesinmachine-learningmodelscomeontopofthefactthatwhenthedataprocessedbyalearning

Differencesofexpertopinionevenexistoverwhatcountsasmachinelearning,withsomedatascientiststreatingformsofwhatothersseeasstandardregressionanalysisasatypeofmachinelearning.

MetaAI,SystemCards,ANewResourceforUnderstandingHowAISystemsWork(Feb.23,2022),

/blog/system-

cards-a-new-resource-for-understanding-how-ai-systems-work/

.

?2023CompetitionPolicyInternationalAllRightsReserved 3

algorithmchanges,thensotoocanitsperformance.Thismeansthat,forsomealgorithms,theirperformancecanbeconstantlyevolvingastheyencounterandprocessnewdata.3

Inshort,machine-learningalgorithmsplacethedefinitionofheterogeneityonsteroids.Thesealgorithmsvarywidelyacrossdifferenttypesanddifferentusesatanygiventime

—andtheyarehighlydynamic,withtheirperformanceevolvingovertime.Allthisheterogeneityholdscrucialimpli-cationsforwhetherandhowmachine-learningalgorithmsshouldberegulated.

02

DECIDINGTOREGULATE

MACHINELEARNING

Thefirstquestiontoask,ofcourse,iswhethermachinelearningneedstoberegulatedatall.4Regulationisatooldesignedtorespondtoandhelpsolvesocialandeconomicproblems.Butbythemselves,machine-learningalgorithmsarejustmathematicalconstructsandcreatenosocialoreconomicproblems.5Iftheywereusedonlyforintellectualpleasure—say,asahobbypursuedbyamathematicallyin-clinedsubsetofthepopulation—thentherewouldsurelybenoneedtoconsiderregulatingthem.Regulatingmachinelearningbecomesatopicofconversationonlywhenitisusedinwaysthathavetangibleeffectsonpeople.

Ifmachinelearningistobeacandidateforregulation,then,itisbecauseoftheusesforwhichitgetsemployed.Thisisnotunlikeotherphysicalmachines.Whenothermachineshavehadconsequentialeffectsonthepublic,theyhave

cometoberegulated.TheNationalHighwayTrafficSafetyAdministration(“NHTSA”),forexample,longagostartingimposingregulatorystandardsondifferentpartsofanauto-mobilenotbecauseofsomethingintrinsicaboutthepartsthemselves,butratherbecauseofhowtheyareusedinve-hiclesandhowthoseusesaffectthesafetyofthevehicle.Machine-learningalgorithmsaremuchthesame.Theyareorwillbecomeobjectsofregulationbecauseofthesystemsinwhichtheyaresituatedandhowtheyultimatelyaffectsystemoutcomesinwaysthattouchpeople’slivesandlive-lihoods.

Becausemachine-learningalgorithmscanbeusedinsomanydifferentways,thismeansthattheregulatoryprob-lemstheycancreatewillvaryquitewidelyaswell.Look-ingacrossahostofdifferentusesofmachinelearning,itispossibletosaythatthepotentialproblemscoverthegamutofclassicmarketfailuresthatjustifyregulation.Machine-learningalgorithmsusedaspartofautomatedpricingsys-temsbyonlineretailers,forexample,maycontributetoanti-competitivebehaviorinthemarketplace.6Machine-learningalgorithmsusedinmedicaltreatmentsandconsumerprod-uctscancontributetothekindofinformationasymmetriesthattypicallyjustifyconsumerprotectionregulation.7Andanypedestrianputatanincreasedriskfromaself-drivingcarshouldeasilybeabletoseeanotherobviousmarketfailure—anexternality—createdbyvehiclesthatoperateautonomouslyusingsensorsandmachine-learningalgo-rithms.

Regulationisoftenjustifiedbymorethanjusttheseclassicmarketfailures.Itcanalsobeused,forexample,asatoolforpreventinginjusticesandprotectingcivilrights,suchaswhenregulationsaimtocombatemploymentdiscrimina-tion.8Groundsexistforregulatingmachinelearningonthisbasisaswell.Whensociety’sprevailingbiaseshavebeenreflectedinthedesignofmachine-learningalgorithmsorinthedataonwhichtheyaretrained,thesealgorithmscanendupreinforcing,ifnotevenexacerbating,existingin-

See,e.g.,JessaBoubker,WhenMedicalDevicesHaveaMindofTheirOwn:TheChallengesofRegulatingArtificialIntelligence,47Am.

J.L.&med.427,434(2021)(indicatingthat,ifanalgorithmiscontinuouslylearning,it“willnotalwaysbeabletopredicthowasoftwareisgoingtoreactinreal-timebasedonnewdata”).

Inposingthequestionintermsofwhetherto“regulatemachinelearning,”Imeantodistinguishitfromthequestionofwhethertoimposeantitrustregulationonthestructuralorotherbusinessdecisionsoffirmsthatrelyheavilyonmachinelearning—namely,theso-calledbigtechfirms.DecidingtoimposeregulatoryscrutinyonmergersandacquisitionsinthebigtechspaceisnotwhatImeanherebyregulatingmachinelearning.Onlyifmachine-learningtoolsarethemselvesdirectlyusedtoimpedecompetitionorconcentratemarketpowerwouldantitrustlawbecomerelevantforregulatingmachinelearninginthesenseImeanhere.

Thisisputtingtotheside,ofcourse,thefactthatprocessingdatausingmachine-learningalgorithmscanresultinexternalitiesfromtheproductionofenergyneededtopowerthenecessarycomputerhardware.

CaryCoglianese&AliciaLai,AntitrustbyAlgorithm,StAn.ComputAtionALAntitruSt,Vol.2,no.1,2022,at4.

Cf.id.at18(describingthedifficultyinsupportingalgorithmicforecastswithintuitiveexplanations,whichmayruninsometensionwithconsumerprotectionprinciplesfavoringdisclosureandtransparency).

See,e.g.,OlatundeC.A.Johnson,BeyondthePrivateAttorneyGeneral:EqualityDirectivesinAmericanLaw,87n.Y.u.L.rev.1339(2012)(providinganoverviewofcivilrightsregulationintheUnitedStates).

4 ?2023CompetitionPolicyInternationalAllRightsReserved

justices.9Machinelearningusedaspartofanemployer’shiringprocess,forexample,canthuscreatetheproblemsthatantidiscriminationregulationhasbeenestablishedtosolve.10

Privacyisanothercivilrightsconcernthatisoftenraisedinthecontextofcallsforregulationofmachinelearning.Oneworrycentersonprotectingtheprivateinformationcontainedintheextensivedataonwhichthesealgorithmsdraw—aswellasensuringindividualnoticeoforconsenttotheuseofsuchinformation.Stillanotherconcernarisesfromtheabilityofmachine-learningalgorithmstomakeaccurateinferencesaboutcertainprivatecharacteristicsthatarenotcontainedinthedatathemselves.Yetanotherconcerncentersonhowmachine-learningalgorithmscanmakepossibletheuseoffacialrecognitionandothertoolsthatcantrackindividu-als’whereaboutsandcontributetofearsofa“surveillancestate.”11

Andthenthereareahostofotherpublicpolicyconcernssur-roundingmachine-learningalgorithmsthatlieattheheartofmanyconversationsaboutregulatingartificialintelligence.12TheavailabilityofChatGPT,forexample,hasraisednewquestionsaboutwhatartificialintelligencemeansforedu-cation.13Socialmediaplatformsusemachine-learningal-gorithmstopushcontenttousersinwaysthataccentuateconflict,keepusersdistracted,ormakethemcravemoretimeontheirsmartphones.14Digitaltoolsdrivenbyma-chine-learningalgorithmscanalsogeneratenewartworkfromexistingworks,raisingquestionsaboutownershiprightsandrulesaboutappropriation.15Thesetoolscanbeusedperniciouslytoo,suchasbyfacilitatingnewoppor-tunitiesforfraudthroughdeepfakes.16Perniciousactors

canalsouseartificialintelligencetopropagatecyberattacksthatthreatenbothdigitalandphysicalassets.17

Asshouldbeevident,theheterogeneoususesformachine-learningalgorithmsleadtoavarietyofregulatoryconcerns.Itissurelyaxiomatictoobservethatwhenthetypesofregulatoryproblemsvary,regulationitselfmustvaryaswelltofitthenatureoftheproblem.Attheveryleast,regulationmustbedesignedinawaythataccommodatesvariationinusesandeithertargetsdiverseproblemsorprovidesappropriateincentivesforregulatedentitiestofindandaddressthoseproblems.18

03

WHOSHOULDREGULATE

MACHINELEARNING?

Beforeturningtohowregulationmightbedesignedtoac-commodatemachinelearning’sheterogeneity,apriorques-tionarisesaboutwhattypeofinstitutionshouldregulatemachinelearning,wheneverthatregulationisjustified.

Withrespecttoothertechnologiesandtheirregulatoryproblems,theneedforregulationtobeadaptedtofitdiffer-entcircumstanceshasledgovernmentstoestablishdiffer-entregulatorybodies,eachtargetingacircumscribedrange

See,e.g.,DorothyRoberts,DigitizingtheCarceralState,132HArv.L.rev.1695,1698(2019)(reviewingvirginiAeubAnkS,AutomAtinginequALitY:HowHigH-teCHtooLSprofiLe,poLiCe,AndpuniSHtHepoor(2018));SandraG.Mayson,Biasin,BiasOut,128YALeL.J.2218(2019).

JeffreyDastin,AmazonScrapsSecretAiRecruitingToolThatShowedBiasAgainstWomen,reuterS(Oct.10,2018,7:04pm),

HttpS://

www.reuterS.Com/ArtiCLe/uS-AmAzon-Com-JobS-AutomAtion-inSigHt/AmAzon-SCrApS-SeCret-Ai-reCruiting-tooL-tHAt-SHowed-biAS-AgAinSt-women-

iduSkCn1mk08g

.

Anumberofjurisdictionshaveprohibitedlawenforcementagenciesfromusingfacialrecognitiontools.SeeCaryCoglianese&KatHefter,FromNegativetoPositiveAlgorithmRights,30wm.&mArYbiLLrtSJ.883,886n.15(2022).

Id.at886-893.

KalleyHuang,AlarmedbyA.I.Chatbots,UniversitiesStartRevampingHowTheyTeach,n.Y.timeS(Jan.16,2023),

https://www.nytimes.

com/2023/01/16/technology/chatgpt-artificial-intelligence-universities.html

.

BarbaraOrtutay&DavidKlepper,FacebookWhistleblowerTestifies:FiveHighlights,ASSoC.preSS(Oct.5,2021),

HttpS://ApnewS.Com/

ArtiCLe/fACebook-frAnCeS-HAugen-CongreSS-teStimonY-Af86188337d25b179153b973754b71A4

.Seegenerallytimwu,tHeAttentionmerCHAntS:tHeepiCSCrAmbLetogetinSideourHeAdS(2016).

ElizabethPenava,AIArtIsinLegalGreyscale,reguL.rev.(Jan.24,2023),

/2023/01/24/penava-ai-art-is-

in-legal-greyscale/

.

toddC.HeLmuS,rAndCorp.,ArtifiCiALinteLLigenCe,deepfAkeS,AnddiSinformAtion:Aprimer(2022).

BlessingGuembe,AmbroseAzeta,SanjayMisra,VictorChukwudiOsamor,LuisFernandez-Sanz&VeraPospelova,TheEmergingThreatofAI-DrivenCyberAttacks:AReview,36AppLiedA.i.1(2022).

Forarelateddiscussion,seeCaryCoglianese,RegulatingNewTech:Problems,Pathways,andPeople,teCHregCHron.,Dec.2021,at65-73.

5

ofproblems.Theproblemscreatedbyanticompetitivebe-havior,afterall,aredifferentthanthosecreatedbyindustrialpollution,whichareinturndifferentthantheproblemsofunsafeandineffectiveconsumerproducts.Asaresult,an-titrustregulatoryinstitutionsexisttotargetanticompetitivebehavior;environmentalregulatorybodiesspecializeinre-ducingpollution;anddrugandconsumersafetyregulatorsaimtoprotectconsumersfromunsafeproducts.Asinglefirmwillneedtocomplywiththeregulationsofseveraldis-tinctregulatorswithrespecttodifferentfacetsofitsopera-tionsandmarketbehavior.

Thesedifferent,specializedregulatorybodieshavethead-vantageoveragenerallegislatureinthattheycandrawuponthespecializedknowledgeneededtoaddressthedif-ferenttypesofproblems,theiroriginsindifferentindustries,andtheireffectsondifferentsubsetsofthepopulation.Thisisnottosaythat,evenwithintheirspecializations,regula-torsdonotconfrontheterogeneity.Onthecontrary,antitrustregulatorsareusuallytaskedwithlookingacrossallsectorsoftheeconomyfordifferentwaysbusinessesmightengageinanticompetitivebehavior.Environmentalregulatorsarecommonlytaskedwithregulatingavarietyoftypesofpollu-tion,suchastotheair,water,andland,andfromamyriadofdifferentbusinesses,largeandsmall.Evenregulatorybod-ieswithrelativelynarrowtargets—suchastheU.S.Nucle-arRegulatoryCommission,whichtargetsasingleindustryfortheimportantbutstillcircumscribedproblemofnuclearsafety19—willfacesomedegreeofheterogeneityinthedif-ferentsourcesofrisksanddifferentscenariosthatmustbeaccountedforifregulationistobeeffective.Nevertheless,becauseofthevalueofspecializedexpertise,nuclearregu-latorsexisttolookatnuclearsafetyandarenotresponsiblefor,say,ensuringthesafetyandsoundnessofbanks.Thisiswhy,asaprescriptivematter,environmentalregulatorsdonotalsoseektocombatanticompetitivemarketconduct,andantitrustregulatorsarenotresponsibleforaddressingpollutionproblems.

Itmaybetemptingtoconcludethatmachine-learningal-gorithmsarelikenuclearpowerplantsandthattheyneedtheirownregulator.Recently,U.S.RepresentativeTedLieu,forexample,hasarguedthat“[w]hatweneedisadedicat-edagencytoregulateA.I.”20Certainly,machine-learningalgorithmsdorequirespecializedskillstounderstandhow

theyworkandhowtheycangoawry.Regulatingmachine-learningalgorithms’impactonanysegmentofsocietyortheeconomywillrequiresophisticatedknowledgeaboutartificialintelligence.Butbecausetheregulatoryproblemsthatmachine-learningalgorithmsareassociatedwithcanbesovaried—andoftensocloselyconnectedtolong-standingregulatoryproblemsthatalreadyhavededicat-edregulatoryinstitutions—itisunrealistictoexpectthatanysingleregulatorcouldeversufficientlyregulatealltheproblematicaspectsofmachinelearning.Regulatingalgo-rithmicstockmarkettradingwillnecessarilyrequiregreatexpertiseaboutfinancialmarkets.Asimilarneedforsub-stantiveexpertisewillapplywhenregulatingtheeffectsofmachine-learningalgorithmsonthesafetyofmedicaldevices,theoperationofautomobiles,andthepricingbe-havioroffirms.NodedicatedAIregulatoryagencycouldpossiblypossessalloftheadditionalrelatedtechnicalknowledgeandcapacityneededtoregulatealgorithms’manyuses.

Itmaybetemptingtoconcludethatmachine-learningalgorithmsarelikenuclearpowerplantsandthattheyneedtheirownregulator

Giventhemanywaysthatmachine-learningalgorithmsareintertwinedwithdifferentproblems,manyofwhichareal-readyaddressedbyexistingregulatorybodies,itisnotsur-prisingthattheseexistingregulatorshavesofartakentheleadinrespondingtopotentialproblemsrelatedtomachinelearning.WithintheDepartmentofTransportation,forex-ample,NHTSAhasissuedregulatoryguidanceforautomo-bilemanufacturersonsafetyassessmentsforautonomousvehicletechnology.21Itorderedthesemanufacturerstofilereportsoncrashesinvolvingtheirautonomousvehicles.22NHTSAalsorecentlyproddedTeslatorecallmorethan350,000ofitsvehiclesoversafetyconcernsrelatedtoitsdriverassistancesoftware.23

AboutNRC,u.S.nuCLeArreguL.Comm’n,

/about-nrc.html

(lastvisitedFeb.4,2023).

TedLieu,I’maCongressmanWhoCodes.A.I.FreaksMeOut.,n.Y.timeS(Jan.23,2023),

/2023/01/23/opinion/

ted-lieu-ai-chatgpt-congress.html

.

U.S.Dep’tTransp.Nat’lHighwayTrafficSafetyAdmin.,FederalAutomatedVehiclesPolicy(Sept.2016),

/sites/

/files/documents/av_policy_guidance_pdf.pdf

.

FirstAmendedStandingGeneralOrder,U.S.Dep’tTransp.Nat’lHighwayTrafficSafetyAdmin.,IncidentReportingforAutomatedDriv-ingSystems(ADS)andLevel2AdvancedDriverAssistanceSystems(ADAS),OrderNo.2021-01(August2021),

/

sites//files/2021-08/First_Amended_SGO_2021_01_Final.pdf

.

NealE.Boudette,TeslatoRecall362,000CarsWithIts“FullSelfDriving”System,n.Y.timeS(Feb.16,2023),

https://www.nytimes.

com/2023/02/16/business/tesla-recall-full-self-driving.html

.

6 ?2023CompetitionPolicyInternationalAllRightsReserved

Separately,theU.S.FoodandDrugAdministration(FDA)hasdevelopedanactionplanforaddressingtheuseofma-chinelearninginmedicaldevices,announcingitwilltreatthemunderaseparatecategoryforinnovativedevices.24In2020,FDAapprovedthefirstAI-basedcardiacultrasoundsoftwareunderthisalternativetrack.25

AsexistingregulatorybodiesgoforwardtoaddressAI-relat-edproblemswithintheirdomains,theywillcertainlyneedtodevelopfurthertheirdatascienceexpertise.Itisnotincon-ceivablethattheycouldbenefitfromacentralizedexpertbodythatcanprovideguidanceandsupport.Already,theNationalInstituteofStandardsandTechnology(NIST)with-intheU.S.DepartmentofCommercehasissuedageneral-izedriskmanagementframeworkforartificialintelligencethatcouldbeofvalueifcustomizedtofittheneedsofothermorespecializedregulatorysettings.26NIST’sframeworkjoinsothersimilardocumentsissuedbyotherfederalenti-ties—suchastheU.S.GovernmentAccountabilityOffice,27theWhiteHouseOfficeofScienceandTechnology,28andtheAdministrativeConferenceoftheUnitedStates29—thatarticulategeneralprinciplestofollowwhenusingmachine-learningtools.ThefederalgovernmenthasalsoestablishedanAICenterofExcellencewithintheGeneralServicesAd-ministration.30

Nevertheless,ashelpfulasthesegeneral,cross-cuttinginitiativesmaybe,existingregulatorsstillneedtobuilduptheirowncapacitytounderstandandregulateAItools,giv-enhowintertwinedtheycanbewithsomanylongstandingregulatoryproblems.Admittedly,evenwithsufficientca-pacitywithinexistingagencies,somekindsofnewprob-lemswillfallthroughthecracks.Illeffectsfromsocialme-diaplatforms’useofalgorithms,forexample,havesofarhaveelidedseriousgovernmentaloversight.Nevertheless,ratherthanhopingthatanewomnibusAIregulatorybodycanswoopintosavethedaybyregulatingallusesofma-chinelearning,policymakerswoulddowelltolookinstead

toempowerexistingcentersofregulatoryexpertise.Wheregapsoroverlapsexistincurrentregulatoryauthority,poli-cymakerscanthenworktofillthosegapsorworkoutanyconflictingjurisdictions.Gapscouldbefilledeitherbycreat-ingnewregulatorybodiesfocusedonunattendedproblemsorbyassigningthosenewproblemstoexistingregulatorswithrelevantexpertise.

04

HOWTOREGULATE

MACHINELEARNING

Nomatterwhichinstitutionstakeresponsibilityforregulat-ingmachinelearning,theywillstillconfrontheterogene-ity.Evenwithinaspecifiedindustryandevenwithrespecttosomeidenticalusesofmachinelearning,heterogeneitywillremainbecauseboththealgorithmsthemselvesandthedatatheyusevarysowidely.Moreover,thealgorithmsandtheautomatedsystemsofwhichtheyareapartarechangingovertime.Asaresult,evenwithinspecializeddomains,regulatorswillneedtopursuemeasuresthattakeintoaccountthevariedanddynamicnatureofthesealgorithms.

Forthisreason,itisimpossibletospecifyatidy,one-size-fits-allformulaforhowregulatorsshouldapproachtheirtaskofregulatingmachinelearning.Butatabroadbrush,itispossibletosaythatregulatorswillneedtoapproachtheirworkwithagility,flexibility,andvigilance.

U.S.Food&DrugAdmin.,ArtificialIntelligenceandMachineLearning(AI/ML)SoftwareasaMedicalDeviceActionPlan(Sept.22,2021),

/medical-devices/software-medical-device-samd/artificial-intelligence-and-machine-learning-aiml

-enabl

ed-medi

-

cal-

devices;U.S.Food&DrugAdmin.,ClinicalDecisionSupportSoftwareGuidanceforIndustryandFoodandDrugAdministrationStaff(Sept.28,2022),

/media/109618/download

.

PressRelease,U.S.Food&DrugAdmin.,FDAAuthorizesMarketingofFirstCardiacUltrasoundSoftwareThatUsesArtificialIntelli-gencetoGuideUser(Feb.7,2020),

/news-events/press-announcements/fda-authorizes-marketing-first-cardiac-ultra-

sound-software-uses-artificial-intelligence-guide-user

.

nAt’LinSt.ofStAndArdS&teCH.(niSt),ArtifiCiALinteLLigenCeriSkmAnAgementfrAmework(Jan.2023),/nistpubs/ai/NIST.AI.100-1.pdf.

u.S.gov’tACCountAbiLitYoff.,GAO-21-519SP,ArtifiCiALinteLLigenCe:AnACCountAbiLitYfrAmeworkforfederALAgenCieSAndotHerentitieS

(June2021),

/assets/gao-21-519sp.pdf

.

wHiteHouSeoff.ofSCi.&teCH.poL’Y,bLueprintforAnAibiLLofrigHtS:mAkingAutomAtedSYStemSworkfortHeAmeriCAnpeopLe,

/ostp/ai-bill-of-rights

.

Admin.Conf.oftheU.S.,AdministrativeConferenceStatement#20:AgencyUseofArtificialIntelligence,86Fed.Reg.6616,6616n.1(Jan.22,2021).

gen.ServS.Admin.,ACCeLe