PolicyandSociety

ISSN:1449-4035(Print)1839-3373(Online)Journalhomepage:

/journals/rpas20

Governanceofarti?cialintelligence

ArazTaeihagh

Tocitethisarticle:ArazTaeihagh(2021)Governanceofarti?cialintelligence,Policyand

Society,40:2,137-157,DOI:

10.1080/14494035.2021.1928377

Tolinktothisarticle:

/10.1080/14494035.2021.1928377

?2021TheAuthor(s).PublishedbyInformaUKLimited,tradingasTaylor&Francis

Group.

Publishedonline:04Jun2021.

Submityourarticletothisjournal

Articleviews:19643

Viewrelatedarticles

ViewCrossmarkdata

Citingarticles:13Viewcitingarticles

FullTerms&Conditionsofaccessandusecanbefoundat

/action/journalInformation?journalCode=rpas20

Routledge

Taylor&FrancisGroup

POLICYANDSOCIETY

2021,VOL.40,NO.2,137–157

/10.1080/14494035.2021.1928377

Governanceofartificialintelligence

ArazTaeihagh

PolicySystemsGroup,LeeKuanYewSchoolofPublicPolicy,NationalUniversityofSingapore,Singapore

ABSTRACT

TherapiddevelopmentsinArtificialIntelligence(AI)andtheinten-sificationintheadoptionofAIindomainssuchasautonomousvehicles,lethalweaponsystems,roboticsandalikeposeseriouschallengestogovernmentsastheymustmanagethescaleandspeedofsocio-technicaltransitionsoccurring.Whilethereiscon-siderableliteratureemergingonvariousaspectsofAI,governanceofAIisasignificantlyunderdevelopedarea.ThenewapplicationsofAIofferopportunitiesforincreasingeconomicefficiencyandqualityoflife,buttheyalsogenerateunexpectedandunintendedconse-quencesandposenewformsofrisksthatneedtobeaddressed.ToenhancethebenefitsfromAIwhileminimisingtheadverserisks,governmentsworldwideneedtounderstandbetterthescopeanddepthoftherisksposedanddevelopregulatoryandgovernanceprocessesandstructurestoaddressthesechallenges.Thisintro-ductoryarticleunpacksAIanddescribeswhytheGovernanceofAIshouldbegainingfarmoreattentiongiventhemyriadofchal-lengesitpresents.Itthensummarisesthespecialissuearticlesandhighlightstheirkeycontributions.ThisspecialissueintroducesthemultifacetedchallengesofgovernanceofAI,includingemer-ginggovernanceapproachestoAI,policycapacitybuilding,explor-inglegalandregulatorychallengesofAIandRobotics,andoutstandingissuesandgapsthatneedattention.Thespecialissueshowcasesthestate-of-the-artinthegovernanceofAI,aimingtoenableresearchersandpractitionerstoappreciatethechallengesandcomplexitiesofAIgovernanceandhighlightfutureavenuesforexploration.

KEYWORDS

Governance;artificial

intelligence;AI;robotics;publicpolicy

1.Introduction

Artificialintelligence(AI)israpidlychanginghowtransactionsandsocialinteractionsareorganisedinsocietytoday.AIsystemsandthealgorithmssupportingtheiroperationsplayanincreasinglyimportantroleinmakingvalue-ladendecisionsforsociety,rangingfromclinicaldecisionsupportsystemsthatmakemedicaldiagnoses,policingsystemsthatpredictthelikelihoodofcriminalactivitiesandfilteringalgorithmsthatcategoriseandprovidepersonalisedcontentforusers(Helbing,

2019

;Mittelstadt,Allo,Taddeo,Wachter,&Floridi,

2016

).Theabilitytomimicorrivalhumanintelligenceincomplexproblem-solvingsetsAIapartfromothertechnologies,asmanycognitivetasks

CONTACTArazTaeihaghspparaz@.sg;araz.taeihagh@LeeKuanYewSchoolofPublicPolicy,NationalUniversityofSingapore,469BBukitTimahRoad,LiKaShingBuilding,Level2,#02-10259771Singapore?2021TheAuthor(s).PublishedbyInformaUKLimited,tradingasTaylor&FrancisGroup.

ThisisanOpenAccessarticledistributedunderthetermsoftheCreativeCommonsAttribution-NonCommercialLicense(

http://

/licenses/by-nc/4.0/

),whichpermitsunrestrictednon-commercialuse,distribution,andreproductioninanymedium,providedtheoriginalworkisproperlycited.

138A.TAEIHAGH

traditionallyperformedbyhumanscanbereplacedandoutperformedbymachines(Bathaee,

2018

;Osoba&Welser,

2017

;S?tra,

2020

).

Whilethetechnologycanyieldpositiveimpactsforhumanity,AIapplicationscanalsogenerateunexpectedandunintendedconsequencesandposenewformsofrisksthatneedtobeeffectivelymanagedbygovernments.AsAIsystemslearnfromdatainadditiontoprogrammedrules,unanticipatedsituationsthatthesystemhasnotbeentrainedtohandleanduncertaintiesinhuman-machineinteractionscanleadAIsystemstodisplayunexpectedbehavioursthatposesafetyhazardsforitsusers(Heetal.,

2019

;Helbing,

2019

;Knudson&Tumer,

2011

;Lim&Taeihagh,

2019

).InmanyAIsystems,biasesinthedataandalgorithmhavebeenshowntoyielddiscriminatoryandunethicaloutcomesfordifferentindividualsinvariousdomains,suchascreditscoringandcriminalsentencing(Huq,

2019

;Kleinberg,Ludwig,Mullainathan,&Sunstein,

2018

).TheautonomousnatureofAIsystemspresentsissuesaroundthepotentiallossofhumanautonomyandcontroloverdecision-making,whichcanyieldethicallyquestionableoutcomesinmultipleapplicationssuchascaregivingandmili-tarycombat(Firlej&Taeihagh,

2021

;Leenesetal.,

2017

;Solovyeva&Hynek,

2018

).ResponsibilityandliabilityforharmsresultingfromtheuseofAIapplicationsremainambiguousundermanylegalframeworks(Leenesetal.,

2017

;Xu&Borson,

2018

)andtheautomationofroutineandmanualtasksindomainssuchasdataanalysis,service,manufacturinganddrivingenabledbymachine-learningalgorithms,chatbotsanddriverlessvehiclesareexpectedtodisplacemillionsofjobsthatwillnotbeevenlydistributedwithinandacrosscountries(Linkov,Trump,Poinsatte-Jones,&Florin,

2018

;Taeihagh&Lim,

2019

).ManagingthescaleandspeedofAIadoptionandtheirattendantrisksisbecominganincreasinglycentraltaskforgovernments.However,inmanyinstances,thebeneficiariesofthesetechnologiesdonotbearthecostsoftheirrisks,andtheserisksaretransferredtothesocietyorgovernments(Leenesetal.,

2017

;Soteropoulos,Berger,&Ciari,

2018

).

WhilethereisconsiderableliteratureemergingonvariousaspectsofAI,governanceofAIisanemergingbutsignificantlyunderdevelopedarea.ToenhancethebenefitsofAIwhileminimisingtheadverseriskstheypose,governmentsworldwideneedtounder-standbetterthescopeanddepthoftherisksposed.Thereisaneedtoreassesstheefficacyoftraditionalgovernanceapproachessuchastheuseofregulations,taxes,andsubsidies,whichmaybeinsufficientduetothelackofinformationandconstantchanges(Guihot,Matthew,&Suzor,

2017

),andthespeedandscaleofadoptionofAIthreatenstooutpacetheregulatoryresponsestoaddresstheconcernsraised(Taeihagh,Ramesh,&Howlett,

2021

).Assuch,governmentsfacemountingpressurestodesignandestablishnewregulatoryandgovernancestructurestodealwiththesechallengeseffectively.TheincreasingrecognitionofAIgovernanceacrossgovernment,thepublic(Chen,Kuo,&Lee,

2020

;Zhang&Dafoe,

2019

,

2020

)andindustryisevidentfromtheemergenceofnewgovernanceframeworksinthemeta-discourseonAIsuchasadaptiveandhybridgovernance(Leiser&Murray

2016

;Linkovetal.,

2018

;Tan&Taeihagh,

2021b

),andself-regulatoryinitiativessuchstandardsandvoluntarycodesofconducttoguideAIdesign(Guihotetal.,

2017

;IEEE

2019

).Thefirsthalfof2018sawthereleaseofnewAIstrategiesfromoveradozencountries,significantboostsinpledgedfinancialsupportbygovern-mentsforAI,andtheheightenedinvolvementofindustrybodiesinAIregulatorydevelopment(Cath,

2018

),raisingfurtherquestionsregardingwhatideasandinterests

POLICYANDSOCIETY139

shouldshapeAIgovernancetoensureinclusionanddiverserepresentationofall

membersof2016;Jobin,&2019.

society(Hemphill,

Ienca,Vayena,

)

ThisspecialissueintroducesthemultifacetedchallengesofgovernanceofArtificialIntelligence,includingemerginggovernanceapproachestoAI,policycapacitybuilding,andexploringlegalandregulatorychallengesofAIandRobotics.ThisintroductionunpacksAIanddescribeswhytheGovernanceofAIshouldbegainingfarmoreattentiongiventhemyriadofchallengesitpresents.Theintroductionthensummarisesofthespecialissuearticlesarepresented,andtheirkeycontributionsarehighlighted.Thankstothediversesetofarticlescomprisingthisspecialissue;ithighlightsthestate-of-the-artinthegovernanceofAIanddiscussestheoutstandingissuesandgapsthatneedattention,aimingtoenableresearchersandpractitionerstoappreciatethechallengesthatAIbringsbetterandunderstandthecomplexitiesofgovernanceofAIandfutureavenuesforexploration.

2.AI–backgroundandrecenttrends

ConceptionsofAIdatebacktoearliereffortsindevelopingartificialneuralnetworkstoreplicatehumanintelligence,whichcanbereferredtoastheabilitytointerpretandlearnfromtheinformation.Originallydesignedtounderstandneuronactivityinthehumanbrain,moresophisticatedneuralnetworksweredevelopedinthelate20thcenturywiththeaidofadvancementsinprocessingpowertosolveproblemssuchasimageandspeechrecognition(Izenman

2008

).TheseeffortsledtotheintroductionoftheconceptofAIascomputerprograms(ormachines)thatcanperformpredefinedtasksatmuchhigherspeedsandaccuracy.InthemostrecentwaveofAIdevelopmentsfacilitatedbyadvance-mentsinbigdataanalytics,AIcapabilitieshaveexpandedtoincludecomputerprogramsthatcanlearnfromvastamountsofdataandmakedecisionswithouthumanguidance,commonlyreferredtoasMachine-learning(ML)algorithms(Izenman

2008

).Unlikeearlieralgorithmsthatrelyonpre-programmedrulestoexecuterepetitivetasks,MLalgorithmsaredesignedwithrulesabouthowtolearnfromdatathatinvolves‘inferentialreasoning’,‘perception’,‘classification’,and‘optimisation’toreplicatehumandecision-making(Bathaee,

2018

;Linkovetal.,

2018

).Thelearningprocessinvolvesfeedingthesealgorithmswithlargedatasets,fromwhichtheyseekandtestcomplexmathematicalcorrelationsbetweencandidatevariablestomaximisepredictionsofaspecifiedoutcome(Kleinbergetal.

2018

;Brauneis&Goodman,

2018

).Asthesealgorithmsadapttheirdecision-makingruleswithmoreexperience,ML-drivendecisionsareprimarilydepen-dentonthedataratherthanonpre-programmedrulesand,thus,typicallycannotbepredictedwellinadvance(Mittelstadtetal.,

2016

).

AmongAIexpertsandresearchers,thereisabroadconsensusthatAIstill‘fallsshort’ofhumancognitiveabilities,andmostAIapplicationsthathavebeensuccessfultodatestemfrom‘narrowAI’or‘weakAI’,whichrefertoAIapplicationsthatcanperformtasksinspecificandrestricteddomains,suchaschess,image,andspeechrecognition(Bostrom&Ludkowsky

2014

;Lele,

2019b

).NarrowAIisexpectedtoautomateandreplacemanymid-skillprofessionsduetotheirabilitytoexecuteroutine,cognitivetasksatmuchhigherspeedsandaccuracythantheirhumancounterparts(Lele,

2019b

b;Linkovetal.,

2018

).Infuture,itisexpectedthatthisformofAIwilleventuallyachieve‘GeneralAI’or‘a(chǎn)rtificialgeneralintelligence’,alevelofintelligence

140A.TAEIHAGH

comparabletoorsurpassinghumansduetotheabilitytogeneraliseacrossdifferentcontextsthatcannotbeprogrammedinadvance(Bostrom&Ludkowsky

2014

;Wang&Siau,

2019

).ThisintroductionandthearticlescomprisingthisspecialissuefocusonapplicationsofnarrowAI.

BothindustryandgovernmentsworldwidehaveenthusedoverthepotentialsocietalbenefitsarisingfromAIandthus,haveacceleratedthetechnology’sdevelopmentanddeploymentacrossvariousdomains.SomeoftheimpetusesfordeployingAIincludeincreasingeconomicefficiencyandqualityoflife,meetinglabourshortages,tacklingageingpopulationsandstrengtheningnationaldefence,andtheyvarybetweengovern-mentsaccordingtoeachnation’suniquestrategicconcerns(Lele,

2019

;Taeihagh&Lim,

2019

).Forinstance,governmentsinJapanandSingaporehavesupportedtheuseofassistiveandsurgicalrobotsinhealthcareandautonomousvehiclesforpublictranspor-tationtomeetlabourshortagesandtackleageingpopulations(Inagaki,

2019

;SNDGO

2019

;Taeihagh&Lim,

2019

;Tan&Taeihagh,

2021

,

2021b

).Cost-savingsandincreasedproductivityarethemainmotivationsforAIadoptioninvarioussectors,whichisalreadytransformingthemanufacturing,logistic,service,andmaritimeindustries(WorldEconomicForum,

2018

).AI-basedtechnologiesarealsoastrategicmilitaryassetforcountriessuchasChina,US,andRussia,whosegovernmentshavemadesignificantinvestmentsinrobots,dronesandfullyautonomousweaponsystemsfornationaldefenceandgeopoliticalinfluence(Allen,

2019

;Lele,

2019

).

3.UnderstandingtherisksofAI

ManyscholarshighlightthesafetyissuesthatcanarisefromdeployingAIinvariousdomains.AmajorchallengefacedbymostAIapplicationstodatestemsfromtheirlackofgeneralizabilitytodifferentcontexts,inwhichtheycanfaceunexpectedsituationswidelyreferredtoas‘cornercases’thatthesystemhadnotbeentrainedtohandle(Bostrom&Ludkowsky

2014

;Lim&Taeihagh,

2019

;Pei,Cao,Yang,&Jana,

2017

).Forinstance,fatalcrasheshavealreadyresultedfromtrialsofTesla’spartiallyautono-mousvehiclesduetothesystem’smisinterpretationofuniqueenvironmentalconditionsthatithadnotpreviouslyexperiencedduringtesting.Whilevariousmeansofdetectingthesecornercasesinadvancehavebeendevised,suchassimulatingdataonmanypossibledrivingsituationsforautonomousvehicles,notallscenarioscanbecoveredorevenenvisionedbythehumandesigners(Bolte,Bar,Lipinski,&Fingscheidt,

2019

;Peietal.,

2017

).DuetothecomplexityandadaptivenatureofMLprocesses,itisdifficultforhumanstoarticulateorunderstandwhyandhowadecisionwasmade,whichhinderstheidentificationofcornercasebehavioursinadvance(Mittelstadtetal.,

2016

).AsMLdecisionsarehighlydata-drivenandunpredictable,thesystemcanexhibitvastlydifferentbehavioursinresponsetoalmostidenticalinputsthatmakeitdifficulttospecify‘correct’behavioursandverifytheirsafetyinadvance(Koopman&Wagner,

2016

).Inparticular,scholarspointoutpotentialsafetyhazardsthatcanalsoarisefromtheinteractionbetweenAIsystemsandtheirusersduetotheproblemofautomationbias,wherehumansaffordmorecredibilitytoautomateddecisionsduetothelatter’sseeminglyobjectivenatureand,thus,growcomplacentanddisplaylesscautiousbehaviourwhileusingAIsystems(Osoba&Welser,

2017

;Taeihagh&Lim,

2019

).Thus,human-machineinterfacessignificantlyshapethedegreeofsafety,particularlyinsocialsettingsthat

POLICYANDSOCIETY141

involvefrequentinteractionswithuserssuchasrobotsforpersonalcare,autonomousvehicles,andserviceproviders.

Thedecision-makingautonomyofAIsignificantlyreduceshumancontrolovertheirdecisions,creatingnewchallengesforascribingresponsibilityandlegalliabilityfortheharmsimposedbyAIonothers.Existinglegalframeworksfortheascribingofrespon-sibilityandliabilityformachineoperationtreatmachinesastoolsthatarecontrolledbytheirhumanoperatorbasedontheassumptionthathumanshaveacertaindegreeofcontroloverthemachine’sspecification(Matthias

2004

;Leenes&Lucivero,

2014

).However,asAIrelieslargelyonMLprocessesthatlearnandadapttheirownrules,humansarenolongerincontroland,thus,cannotbeexpectedtoalwaysbearrespon-sibilityforAI’sbehaviour.Understrictproductliability,manufacturersandsoftwaredesignerscouldbesubjecttoliabilityformanufacturingdefectsanddesigndefects,buttheunpredictabilityofMLdecisionsimpliesthatmanyerroneousdecisionsmadebyAIarebeyondthecontrolofandcannotbeanticipatedbytheseparties(Butcher&Beridze,

2019

;Kimetal.

2017

;Lim&Taeihagh,

2019

).ThisraisescriticalquestionsregardingtheextenttowhichdifferentpartiesintheAIsupplychainwillbeheldliableindifferentaccidentscenariosandthedegreeofautonomythatissufficientto‘limit’theresponsi-bilityofthesepartiesforsuchunanticipatedaccidents(Osoba&Welser,

2017

;Wirtz,Weyerer,&Sturm,

2020

).Itisalsowidelyrecognisedthatexcessiveliabilityriskscanhinderlong-runinnovationandimprovementstothetechnology,whichhighlightsamajorissueregardinghowgovernmentscanstructurenewliabilityframeworksthatbalancethebenefitsofpromotinginnovationwiththemoralimperativeofprotectingsocietyfromtherisksofemergingtechnologies(Leenesetal.,

2017

).

Giventhevalue-ladennatureofthedecisionsautomatedbyalgorithmsinvariousaspectsofsociety,AIsystemscanpotentiallyexhibitbehavioursthatconflictwithsocietalvaluesandnorms,promptingconcernsregardingtheethicalissuesthatcanarisefromAI’srapidadoption.Oneofthemostintensivelydiscussedissuesacrossindustryandacademiaisthepotentialforalgorithmicdecisionstobebiasedanddiscriminatory.AsMLalgorithmscanlearnfromdatagatheredfromsocietytomakedecisions,theycouldnotonlyconflictwiththeoriginalethicalrulestheywereprogrammedwithbutalsoreproducetheinequalityanddiscriminatorypatternsofsocietythatiscontainedinsuchdata(Goodman&Flaxman,

2017

;Osoba&Welser,

2017

;Piano,

2020

).Ifsensitivepersonalcharacteristicssuchasgenderorraceinthedataareusedtoclassifyindividuals,andsomecharacteristicsarefoundtonegativelycorrelatewiththeoutcomethatthealgorithmisdesignedtooptimise,theindividualscategorisedwiththesetraitswillbepenalisedoverotherswithdifferentgroupcharacteristics(Liu2018).Thiscouldyielddisparateoutcomesintermsofriskexposureandaccesstosocialandeconomicbenefits.Biascanalsobeintroducedthroughthehumandesignerinconstructingthealgorithm,andevenifsensitiveattributesareremovedfromthedata,therearetechniquesforMLalgorithmstouse‘probabilisticallyinferred’variablesasaproxyforsensitiveattributes,whichismuchhardertoregulate(Krolletal.,

2016

;Osoba&Welser,

2017

).TheriskofbiasanddiscriminationstemmingfromtheoptimisationprocessinAIalgorithmsreflectsadominantconcernsurroundingdiscussionsoffairnessinAIgovernance–thetrade-offbetweenequityandefficiencyinalgorithmicdecision-making–(S?tra,

2020

)andhowabalancecanbestrucktoproducesociallydesirableoutcomescateringtothedifferentgroups’ethicalpreferencesremainssubjecttodebate.

142A.TAEIHAGH

AvastbodyofliteratureandgovernmentreportshavehighlightedissuesofdataprivacyandsurveillancethatcanarisefromAIapplications.AsalgorithmsinAIsystemsutilisesensorstocollectdataandbigdatatechnologiestostore,processandtransmitdatathroughexternalcommunicationnetworks,therehavebeenconcernsregardingthepotentialmisuseofpersonaldatabythirdpartiesandincreasingcallsformoreholisticdatagovernanceframeworkstoensurereliablesharingofdatawithinandbetweenorganisations(Gasser&Almeida,

2017

;Janssen,Brous,Estevez,Barbosa,&Janowski,

2020

).AIsystemsstoreextensivepersonalinformationabouttheirusersthatcanbetransmittedtothirdpartiestoprofileindividuals’preferences,suchasusingpasttraveldatacollectedinautonomousvehiclestotailoradvertisementstopassengers(Chenetal.,

2020

;Lim&Taeihagh,

2018

),usingpersonalandmedicalinformationcollectedbypersonalcarerobotsandnetworkedmedicaldevicesforthesurveillanceofindividuals(Guihotetal.,

2017

;Leenesetal.,

2017

;Tan,Taeihagh,&Tripathi,

2021

).TheownershipofsuchdataandhowAIsystemdevelopersshoulddesigntheserobotstoadheretoprivacylawsarekeyconcernsthatremaintobeaddressed(Chenetal.,

2020

;Leenesetal.,

2017

).SurveillanceisalsoakeyconcernovertheuseofAIinmanydomains,suchassurveillancerobotsintheworkplacethatmonitoremployeeperformanceandgovern-mentagenciespotentiallyusingautonomousvehiclestotrackpassengermovementswithnegativeimplicationsfordemocraticfreedomsandpersonalautonomy(Leenesetal.,

2017

;Lim&Taeihagh,

2018

).

TheautonomyassumedbyAIsystemstomakedecisionsinplaceofhumanscanintroduceethicalconcernsintheirapplicationacrossvarioussectors.Studieshaveunderlinedthepotentialforpersonalisationalgorithmsusedbydigitalplatformstounderminethedecision-makingautonomyofdatasubjectsbyfilteringinformationpresentedtousersbasedontheirpreferencesandinfluencingtheirchoices.Byexertingcontroloveranindividual’sdecisionandreducingthe‘diversityofinformation’pro-vided,personalisationalgorithmscanreducepersonalautonomyand,thus,beconstruedasunethical(Mittelstadtetal.,

2016

).Inhealthcare,theuseofrobotstoprovidepersonalcareserviceshaspromptedconcernsoverthepotentiallossofautonomyanddignityofcarerecipientsifrobotsexcessivelyrestrictpatients’mobilitytoavoiddangeroussitua-tions(Leenesetal.,

2017

;Tanetal.,

2021

).Studieshaveyettoexaminehowtheseriskscanbebalancedagainsttheirpotentialbenefitsforautonomyinotherscenarios,suchasautonomousvehiclesincreasingmobilityforthedisabledandelderly(Lim&Taeihagh,

2018

),andpersonalcarerobotsofferingpatientsgreaterfreedomofmovementwiththeassuranceofbeingmonitored(Leenesetal.,

2017

).Inthemilitary,autonomousweaponsystemssuchasdronesandunmannedaerialvehicleshavebeendevelopedtoimprovetheprecisionandreliabilityofmilitarycombat,planningandstrategy,buttherehasbeenincreasingmomentumacrossindustryandacademia,includingprominentfigures,high-lightingtheirethicalandlegalunacceptability(Lele,

2019

;Roff,

2014

).Centraltotheseconcernsisthedelegationofauthoritytoamachinetoexertlethalforce‘independentlyofhumandeterminationsofitsmoralandlegallegitimacy’andthelackofcontrollabilityovertheseadaptivesystemsthatcouldamplifytheconsequencesoffailure,promptingfearsofadystopianfuturewheresuchweaponsinflictcasualtiesandescalatecrisesatamuchlargerscale(Firlej&Taeihagh,

2021

;Scharre,

2016

;Solovyeva&Hynek,

2018

). Unemploymentandsocialinstabilityresultingfromtheautomationofroutinecog-nitivetasksremainsoneofthemostpubliclydebatedissuesconcerningAIadoption

POLICYANDSOCIETY143

(Frey&Osborne,

2017

;Linkovetal.,

2018

).Theeffectsofautomationarealreadyfeltinindustriessuchasthemanufacturing,entertainment,healthcare,finance,andtransport

sectorsascompaniesincreasinglyinvestinAItoreducelabourcostsandboostefficiency(Linkovetal.,

2018

).Whiletechnologicaladvancementshavehistoricallycreatednewjobsaswell,thereareconcernsthatthedistributionofemploymentopportunitiesisunevenacrosssectorsandskilllevels.Studiesshowthathighlyroutineandcognitivetasksthatcharacterisemanymiddle-skilledjobsareatahighriskofautomation.Incontrast,taskswithrelativelylowerrisksofautomationarethosethatmachinescannoteasilyreplicate–thisincludesmanualtasksinlow-skilled,serviceoccupationsthatrequireflexibilityand‘physicaladaptability’,aswellashigh-skilledoccupationsinengineeringandsciencethatrequirecreativeintelligence(Frey&Osborne,

2017

;WorldEconomic;Forum,

2018

).Ashigh-andlow-skilledoccupationsbenefitfromincreasedwagepremiumsandmiddle-skilledjobsarebeingphasedout,automationcouldexacerbateincomeandsocialinequalities(Alonsoetal.

2018

).

4.GoverningAI

4.1WhyAIgovernanceisimportant

UnderstandingandmanagingtherisksposedbyAIiscrucialtorealisethebenefitsofthetechnology.Increasedefficiencyandqualityinthedeliveryofgoodsandservices,greaterautonomyandmobilityfortheelderlyanddisabled,andimprovedsafetyfromusingAIinsafety-criticaloperationssuchasinhealthcare,transportandemergencyresponsearethemanysocio-economicbenefitsarisingfromAIthatcanpropelsmartandsustainabledevelopment(Agarwal,Gurjar,Agarwal,&Birla,

2015

;Lim&Taeihagh,

2018

;Yigitcanlaretal.,

2018

).Thus,asAIsystemsdevelopandincreaseincomplexity,theirrisksandinterconnectivitywithothersmartdevicesandsystemswillalsoincrease,necessitatingthecreationofbothspecificgovernancemechanisms,suchasforhealth-care,transportandautonomousweapons,aswellasabroaderglobalgovernanceframe-workforAI(Butcher&Beridze,

2019

).

4.2ChallengestoAIgovernance

ThehighdegreeofuncertaintyandcomplexityoftheAIlandscapeimposesmanychallengesforgovernmentsindesigningandimplementingeffectivepoliciestogovernAI.ManychallengesposedbyAIstemfromthenatureoftheproblem,whicharehighlyunpredictable,intractableandnonlinear,makingitdif