關注AI鴻溝：塑造未來勞動世界的全球視角 Mind the AI Divide Shaping a Global Perspective on the Future of Work

MindtheAIDivide

ShapingaGlobalPerspectiveontheFutureofWork

MindtheAIDivide:ShapingaGlobalPerspectiveontheFutureofWork

Copyright?2024UnitedNationsAllrightsreservedworldwide.

Nopartofthispublicationmay,forcommercialpurposes,bereproducedortransmittedinanyformorbyanymeans,electronicormechanical,includingphotocopy,recordingoranyinformationstorageandretrievalsystemnowknownortobeinvented,withoutwrittenpermissionbythepublisher.

RequeststoreproduceexcerptsortophotocopyshouldbeaddressedtotheCopyrightClearanceCenterat.

Allotherqueriesonrightsandlicenses,includingsubsidiaryrights,shouldbeaddressedto:UnitedNationsPublications,405East42ndStreet,S-11FW001,NewYork,NY10017,UnitedStatesofAmerica.Email:permissions@.Website:.

ThedesignationsemployedandthepresentationofthematerialinthispublicationdonotimplytheexpressionofanyopinionwhatsoeveronthepartoftheSecretariatoftheUnitedNationsconcerningthelegalstatusofanycountry.

PDFISBN:9789211066524

Foreword

TheunevenadoptionofArtificialIntelligence(AI)isacriticalissuethatgoesbeyondeconomic

growth.Itimpactsglobalequity,fairnessandthesocialcontractthatisattheheartofsocialjustice.Disparitiesinaccesstorobustinfrastructure,advancedtechnology,qualityeducationandtrainingaredeepeningexistinginequalities.AstheglobaleconomyincreasinglyshiftstowardsAI-driven

productionandinnovation,lessdevelopedcountriesriskbeingleftfurtherbehind,exacerbating

economicandsocialdivides.Withouttargetedandconcertedeffortstobridgethisdigitaldivide,

AI’spotentialtofostersustainabledevelopmentandalleviatepovertywillremainunrealized,leavingsignificantportionsoftheglobalpopulationdisadvantagedintherapidlyevolvingtechnological

landscape.

DuringtheconsultationsheldbytheSecretary-General’sHigh-levelAdvisoryBodyonArtificial

Intelligence,ithasbecomeclearthattheworldofworkisattheheartoftheadoptionofAI.Itis

thuscriticaltounderstandthepotentialforAItoaffectlabourdemandandtransformoccupations.Itisattheworkplacewherethepotentialforproductivitygainsandimprovedworkingconditionsforthebenefitofworkers,theirfamilies,andsocietiesatlarge,canberealized.Butsuchbenefitswillnothappenontheirown;theywillonlybeachievediftherightconditionsareinplace,includingtheavailabilityofdigitalinfrastructureandskills,butalsoacultureofsocialdialoguethatfostersapositiveintegrationoftechnology.

PromotinginclusivegrowthrequiresproactivestrategiestosupportAIdevelopmentincountriesonthewrongsideoftheAIdivide.Thisinvolvesenhancingdigitalinfrastructure,promotingtechnologytransfer,buildingAIskills,andensuringthatalljobsalongtheAIvaluechainareofgoodqualityandimprovethelivesofworkingpeople.ByprioritizinginternationalcollaborationinAIcapacitybuilding,wecancreateamoreequitableandresilientAIecosystem,unlockingopportunitiesforshared

prosperityandhumanadvancementworldwide.

WelookforwardtocontinuingourcollaborativeeffortstoshapetheglobalgovernanceofAI,upholdhumandignityandlaborstandards,andexpandeconomicopportunityforall.

AmandeepSinghGill

UnitedNationsSecretary-General’sEnvoyonTechnology

GilbertF.Houngbo

Director-GeneraloftheInternationalLabourOrganization

MindtheAIDivide:ShapingaGlobalPerspectiveontheFutureofWork|3

Contents

Foreword3

Section1.Introduction5

Section2.Unevenground:UnderstandingAI’sroleinreshapinglabourmarkets6

Ensuringjobqualityunderaugmentation10

Section3.TheAIvaluechainandthedemandforskills11

AdaptingskillsfortheAIlandscape14

Section4.Movingforward:Strengtheninginternationalcooperation,building17

nationalcapacity,andaddressingAIintheworldofwork

StrengthenedinternationalcooperationonAI17

BuildingnationalAIcapacity18

TowardsapositiveintegrationofAIintheworldofwork18

Acknowledgments20

References21

4|MindtheAIDivide:ShapingaGlobalPerspectiveontheFutureofWork

Section1

Introduction

TherapidadvancementofArtificialIntelligence(AI)promiseswidespreadtransformations

foroursocieties,oureconomiesandthe

worldofwork.Whilesuchadvancesoffer

tremendousopportunitiesforinnovationand

productivity,theunevenratesofinvestment,

adoptionanduseamongcountriesrisks

exacerbatingthealreadywidedisparities

inincomeandqualityoflife.Thereisa

pronounced“AIdivide”emerging,wherehighincomenationsdisproportionatelybenefitfromAIadvancements,whilelow-andmedium-

incomecountries,particularlyinAfrica,lag

behind.Worse,thisdividewillgrowunless

concertedactionistakentofosterinternationalcooperationinsupportofdevelopingcountries.Theabsenceofsuchpolicieswillnotonly

widenglobalinequalities,butriskssquanderingthepotentialofAItoserveasacatalystfor

widespreadsocialandeconomicprogress.

WhileAIwillpotentiallyaffectmanyaspects

ofourdailylives,itsimpactislikelytobemostacuteintheworkplace.Wealthiercountries

aremoreexposedtothepotentialautomatingeffectsofAIintheworldofwork,buttheyarealsobetterpositionedtorealizetheproductivitygainsitoffers.Developingcountries,onthe

otherhand,maybetemporarilybuffered

becauseofalackofdigitalinfrastructure,butthisbufferrisksturningintoabottleneckforproductivitygrowth,andmoreimportantly,forthefutureprosperityoftheirpopulations.

Ensuringinclusivegrowthinthefuture

requiresproactivemeasurestoempowerAI

developmentincountriesatthedisadvantagedreceivingendofthedigitaldivide,fostering

digitalinfrastructureaswellasAIskills,and

promotingtechnologytransferandabsorption.Suchdigitalskillscanalsoenableamore

positiveintegrationofAIintheworkplace,particularlywhencombinedwithsocial

dialogue.Socialdialogueonthedesign,

implementationanduseoftechnologyattheworkplace,aswellasinthedevelopmentofregulationsessentialforensuringrespect

ofworkers’fundamentalrights,isneeded.

Indeed,whethertheintegrationoftechnologyintoworkprocessesspursproductivitygrowthorimprovesworkingconditionsinsupport

ofsocialjusticedependsinlargepartonthestrengthofsuchcollaborationanddialogue.

SovereigneffortsplayacrucialroleinshapingAIcapacitybuildingascountriesassert

theirautonomyindevelopingstrategies

andpoliciestailoredtotheiruniquesocio-

economiccontexts.Localprocesses,driven

byculturalvalues,politicaleconomies,and

societalneeds,cansignificantlyimpactthe

effectivenessandsustainabilityofAIinitiatives.However,disparitiesinresourcesandexpertiseremainandcanhinderAIdevelopmentinthe

GlobalSouth.Inresponse,thereisagrowing

recognitionoftheresponsibilityofdevelopedcountriestosupportcapacitybuildingeffortsinresourcescarcecountries.Asoutlined

intherecentInterimReportoftheUnited

NationsSecretary-General’sHigh-levelAdvisoryBodyonAI1,thisrecognitionextendsbeyond

financialaidtoincludeknowledgesharing,

skillsdevelopment,technologytransfer,aswellascollaborativeresearchanddevelopment

partnerships.Byleveragingtheiradvanced

AIecosystems,theGlobalNorthcanhelp

bridgethegapandempowercountriesinthe

GlobalSouthtoharnessAIforsustainable

development,whilerespectingtheirsovereigntyandpromotinglocalinnovationecosystems.ByprioritizingglobalcollaborationforAIcapacitybuilding,theinternationalcommunitycan

nurtureamoreequitableandresilientglobalAIecosystem,unlockingopportunitiesforsharedprosperityandhumanflourishingacrossthe

world.

1

/ai-advisory-body

MindtheAIDivide:ShapingaGlobalPerspectiveontheFutureofWork|5

Section2

Unevenground

UnderstandingAI’sroleinreshapinglabourmarkets

ResearchonthepossibleeffectsofgenerativeAIonemploymentacrosstheworldsuggeststhatwhiletherearelikelytobeimportant

transformativeeffectsonsomeoccupations,impactsintermsofjoblossesaremuchlessthanheadlinefiguresappearinginthemedia,andcertainlydonotpointtoajoblessfuture.AccordingtoananalysisundertakenbytheInternationalLabourOrganizationonthe

potentialexposureoftaskstogenerativeAItechnology,clericalsupportworkersarethemostexposedoccupationalgroupwith24

percentofthetasksinthesejobsassociatedwithhighlevelofexposuretoautomation

andanother58percentwithmedium-levelexposure(seeFigure1).2Otheroccupationalgroupsarelessexposed,withonly1to4

percentoftasksconsideredashavinghigh

automationpotential,andmedium-exposedtasksnotexceeding25percent.Thismeansthat,whilecertaintasksintheseoccupationscouldpotentiallybeautomated,mosttasks

stillrequirehumanintervention.Suchpartialautomationcouldenableefficiencygains,byallowinghumanstospendmoretimeonotherareasofwork.

Importantly,taskautomationdoesnot

necessarilyimplyredundancies,asthe

technologycanalsocomplementoraugmenthumanlabourwhenonlycertaintasksare

automated.Whethertheadoptionofthe

technologyleadstoautomation(jobloss)oraugmentation(jobcomplementarity)dependsonthecentralityoftheautomatedtasktotheoccupation,howthetechnologyisintegrated

Figure1:Taskswithmediumandhigh-levelexposuretogenerativeAItechnologybymajoroccupationalgroup(ISCO1-digit)

Source:Gmyreketal.,2023.

2Thestudyanalysesthepotentialforautomationwiththe436internationallystandardizedISCO-08occupationsandthenclassifiestheoccupationbasedonthemeanandstandarddeviationofthescore.Formoredetailssee[1].

6|MindtheAIDivide:ShapingaGlobalPerspectiveontheFutureofWork

intoworkprocessesandmanagement’s

desiretoretainhumanstoperformoroverseesomeofthetasks,despitethepotentialof

automation.

TheILOanalysisusesoccupationalexposurescores(themeanexposureofeachofthe

taskswithinanoccupation)andappliesthesescorestoemploymentdatafromlabourforcesurveysofmorethan140countriestoassesspotentialemploymentimpactattheglobal

andregionallevel.Withrespecttoautomation,theshareofemploymentthatisexposed

ishighestinEuropeandNorthernAmerica,

reflectingthegreatereconomicandlabour

marketdiversificationoftheseregions.In

LatinAmerica,AsiaandAfrica,theshareofemploymentpotentialexposedtoautomationismuchsmaller,duetothegreatershareofworkersemployedinoccupationsthatwouldnotbeexposedtogenerativeAItechnologysuchasinagriculture,transportorfood

vending.

Nevertheless,women’spotentialexposuretotheautomatingeffectsofgenerative

AItechnologyismuchhigher,duetotheir

over-representationinclericaloccupations

(seefigure2).Inmostregions,thepotential

exposureofwomenismorethandoublethatofmen’sexposure.Someofthisemploymentisinbusinessprocessoutsourcing,suchascontactorcallcenterwork,whichisanimportantpartoftheeconomyofseveraldevelopingcountries,includingIndiaandthePhilippines.Theindustryisanimportantsourceofformalandrelativelywell-paidemployment,particularlyforwomen.Whilepotentialexposuredoesnotnecessarilytranslatetodisplacement,itisclearthatthe

advancesintechnologymayputsomeofthesejobsatrisk.3

Anotherfindingisthatasignificantlylarger

shareoftotalemploymentisinoccupations

withhighaugmentationpotential,andthis

holdsacrossregions,from10.2percent

inSub-SaharanAfricato16.1percentin

SoutheasternAsiaandthePacific(Seefigure3).Thus,thepotentialforoccupationsto

benefitfromtheproductivity-enhancingeffectsofthetechnologyisrelativelysimilaracross

countries.Inpractice,however,itislesslikely

Figure2:Potentialexposuretoautomationbyglobalsub-region

3‘AICouldKilloffMostCallCentres,SaysTataConsultancyServicesHead’,April25,2024.

MindtheAIDivide:ShapingaGlobalPerspectiveontheFutureofWork|7

Figure3:Potentialexposuretoaugmentationbyglobalsub-region

toberealizedduetoconstraintsinphysical

infrastructure(electricityaccess,broadband)aswellasdigitalskills.Indeed,subsequent

researchthatincorporatesdataoncomputeruseatwork[2]revealsthatmanyofthe

occupationswithpotentialforaugmentationhaverelativelylowusageofcomputeratwork,

suggestingthattheconditionsarenotinplaceforrealizingthepotentialproductivitygains.

AscanbeseeninFigure4,theshareof

workerswithoutaccesstoacomputeratwork(“nocomputer”)exceedsthosewhousea

computerin9ofthe16countrieslisted.As

Figure4:Potentialexposuretoaugmentationandcomputeruseatwork

Source:Gmyrek,WinklerandGarganta,2024.

8|MindtheAIDivide:ShapingaGlobalPerspectiveontheFutureofWork

such,thelikelihoodtorealizeproductivitygainsfromgenerativeAItechnologywillbelimited.

Figure5givesinformationonthe

characteristicsofthosewhomightbeaffectedbyautomationfromgenerativeAItechnologyinLatinAmerica.Asthedatashow,itiseducatedwomenlivinginurbanareasandbelonging

tothetopfifthoftheincomedistributionthataremostexposed.ForLatinAmerica,theseoccupationsareoverwhelminglyinsalaried,formalemploymentandinthesectorsof

finance,professionalservicesandpublicadministration.Inshort,theyaregoodjobs,whoselosswouldhavenegativemultipliereffectsbotheconomicallyandsocially.

Theanalysisdoesnotaddressthepotentialfornewjobcreation.Thus,whilemiddle-incomecountriessuchasIndiaandthePhilippines,

aremoreexposedtotheautomatingeffects

ofgenerativeAItechnologyintheircallcentrework,theirdigitalinfrastructureandskilled

workforcecanalsobeanassetforspawningthegrowthofcomplementaryindustries.

Harnessingsuchpotentialisparamount.

Indeed,withtherightconditionsinplace,a

newwaveoftechnologycouldfuelgrowth

opportunities.Inthepast,technological

advancementshavespurrednewand

successfulindustriesinmanydeveloping

countries.OnesuchexampleistheM-Pesa

moneyservice,whichreliedonthediffusionofmobiletelephonesinKenya.Theservice,

inturn,increasedfinancialinclusionwhich

helpedtopropelthegrowthofSMEsandledtocreationofanetworkof110,000agents,

40timesthenumberofbankATMsinKenya[3];[4].Similarly,astudyofthediffusionof3GcoverageinRwandabetween2002and2019foundthatincreasedmobileinternetcoverage

Figure5:Characteristicsofpersonsholdingoccupationsmostexposedtoautomation,LatinAmerica

Source:Gmyrek,WinklerandGarganta,2024(forthcoming).

MindtheAIDivide:ShapingaGlobalPerspectiveontheFutureofWork|9

waspositivelyassociatedwithemployment

growth,increasingbothskilledandunskilled

occupations[5].Scholars[6]alsofindpositiveemploymenteffects,fromthearrivalofinternetin12Africancountries,albeitwithaslight

biastowardsskilledoccupations.Thesegainsareattributedtoincreasesinproductivityandgrowthofmarketsthatfollowedincreased

connectivity,underliningtheneedforsuch

investments,givenimportantmultipliereffectsontheeconomyandlabourmarkets.

Ensuringjobqualityunderaugmentation

Anotherareaofconcernisabouttheimpact

ofAItechnologyonworkingconditionsand

jobqualitywhenthetechnologyisintegrated

intotheworkplace.Whilesuchintegration

intoworktaskscanpotentiallypromotemoreengagingworkifroutinetasksareautomated,itcanalsobeimplementedinwaysthat

limitsworkers’agencyoraccelerateswork

intensity.ConcernsoverAI’sintegrationat

theworkplacehasfocusedonthegrowthof

algorithmicmanagement,essentiallywork

settingsinwhich“humanjobsareassigned,

optimized,andevaluatedthroughalgorithms

andtrackeddata”[7].Algorithmicmanagementisadefiningfeatureofdigitallabourplatforms,butitisalsopervasiveinofflineindustries

suchasthewarehousingandlogisticssectors.Inwarehousesanautomated,“voice-picking”

systemdirectswarehousestafftopickcertainproductsinthewarehouse,whileusingdata

collectiontomonitorworkersandsetthe

paceofwork[8].Besideslackingautonomytoorganizetheirworkorsetitspace,workersalsohavelittleabilitytoprovidefeedbackordiscusswithmanagementabouttheorganizationof

work[9].TheintegrationofgenerativeAIintootherfieldssuchasbanking,insurance,socialservices,andcustomerservicemorebroadlymayhaveasimilareffect.

Technologicaladvancementsareoftenfeltmoreimmediatelyattheworkplacelevelandareusuallybestaddressedattheworkplace.

Asaresult,whethertheeffectoftechnology

onworkingconditionsispositiveornegativedependsinlargepartonthevoicethatworkershaveinthedesign,implementationanduseoftechnology.Havingsuchagencyreliesinturnontheopportunitiesforworkerparticipation

anddialogue.Thiscantakeplaceeither

throughformalizedsettings,suchasworks

councilsorguidanceprovidedincollective

bargainingagreements,orlessformally,in

workplaceswherethereisahighdegreeof

employeeengagement.StudiesinEurope

haveshownthatitiscountrieswithstrongerandmorecooperativeformsofworkplace

consultation,essentiallytheNordiccountriesandGermany,whereworkersaremoreopentotechnologicaladoptionattheworkplace[10].

10|MindtheAIDivide:ShapingaGlobalPerspectiveontheFutureofWork

Section3

TheAIvaluechainand

thedemandforskills

Liketheproductionofmanygoodsand

servicesintheglobaleconomy,AIhasitsownvaluechain.AsdepictedinFigure6,therearedifferentstagesoftheAIvaluechain,eachwithspecifichumanandsocialinfrastructureneeds.Asistypicalinmostglobalvaluechains,stagesdifferintheamountofvaluereceivedforthe

contributionmade,withlower-valueadded

activitiespredominantinmiddleandlow-

incomecountriesanddesignanddeploymentassociatedwithhigher-incomecountries.

DataisfundamentaltothedevelopmentandoperationofAIsystems.Human-prepared

dataisfedintoAIsystemstohelpthemlearnthenecessaryconnectionsandpatternsfor

functionality.Thesourcesofthisdataare

diverse,dependingonthesystem’spurpose.Publiclyavailabledata,suchasUnitedNationsdocumentsusedfortrainingtranslation

programs,contributedtoadvancesinnaturallanguageprocessing.Proprietarydataisalsocrucial,particularlyinworkplaceapplications,likecallcenterrecordingsusedtotrain

chatbotsforcustomerservice.Withglobal

connectivity,datacollectioncontinuesto

providetheessentialrawmaterialforfutureAIapplications.

Whendataiscollected,itisusually

unstructured.Highlyskilleddataengineers

willpre-processthedataintoausableformat,but‘datalabelers’areneededtolabeland

classifydatasothatitisusable.Labelled

andannotateddatasetsarecriticalforthe

developmentandeffectivenessofmachine

learningmodels.Workersinvolvedindata

enrichmentcarryoutanarrayoftasksthat

includemarkingradiologyscanstoaidin

creatingAIsystemscapableofdetecting

cancer;categorizingtoxicandunsuitable

onlinecontenttoimprovecontentmoderationalgorithmsordiminishthenegativityinlargelanguagemodelresponses;annotating

videofootagefromdrivingsessionstotrain

autonomousvehicles;editinglargelanguagemodeloutputstoboosttheirfunctionality;andmore.4

Contentmoderationistheprocessof

monitoringandfilteringuser-generated

contentondigitalplatforms,suchassocialmedia,forums,andwebsites,toensurethatitcomplieswiththeplatform’sguidelinesandpolicies.Thegoalofcontentmoderationistomaintainasafe,respectful,andpositive

environmentforallusersbyremovingor

Figure6:ValuechainofAI

3

Note:Orangerepresentstheactivitiesthathavelowervalue-added.

Source:Authors’elaboration.

2

4

7

6

5

1

4ValuingDataEnrichmentWorkers:TheCaseforaHuman-CentricApproachtoAIDevelopment|UnitedNations

MindtheAIDivide:ShapingaGlobalPerspectiveontheFutureofWork|11

flaggingcontentthatisinappropriate,offensive,harmful,orillegal.Contentmoderationcanbeperformedmanuallybyhumanmoderators

orautomaticallybyusingalgorithmsand

machinelearningtools.Thetypesofcontentthatmaybesubjecttomoderationcanvary

widely,includingbutnotlimitedtohate

speech,harassment,violence,nudity,andfalseinformation.Evenwiththeuseofalgorithmsandmachinelearningtoolsforcontent

moderation,thereistypicallyalwaysahumaninvolvedintheprocess.Thesetechnologiescanhelpautomateandscalethemoderationprocess,buttheyarenotperfectandcan

sometimesmakemistakesormissnuancesthatahumanmoderatorwouldbeabletopickupon.

Inmanycases,algorithmsareusedtoflag

orprioritizecontentforreviewbyhuman

moderators,whothenmakethefinaldecisiononwhetherthecontentshouldberemovedorallowedtoremainontheplatform.Additionally,humanmoderatorsmayalsobeinvolvedin

trainingandimprovingthealgorithms,by

providingfeedbackandlabellingdatathatcanbeusedtorefinethesystem’saccuracyand

effectiveness.Individualstaskedwithcontentmoderationdutiesinsocialmediaplatforms

oftensufferfromanxiety,depression,andpost-traumaticstressdisorder,adirectconsequenceoftheircontinuousexposuretodistressing

materialssuchasmurder,suicide,sexualassault,orchildabusevideos.

Theseexamplesdemonstratehowhumansareintegraltotheprovisionofservicesmarketedordescribedas“artificialintelligence”.Indeed,JeffBezosdescribedAmazon’sMechanical

Turk(AMT)platformas“artificial-artificial-

intelligence”asitwashumanintelligence

thatwasprovidingthelabour-intensivework

neededforartificialintelligencesystemsto

operate.AsdescribedontheAMTsite,the

platformprovides“anon-demand,scalable,

humanworkforcetocompletejobsthat

humanscandobetterthancomputers,for

example,recognizingobjectsinphotos”.5

Workersontheplatformareaccessiblethroughanapplicationprogramminginterface(API),

allowingprogrammerstocallonworkerswith

afewsimplelinesofcodewhenworkingonanalgorithm[11].

InadditiontoplatformssuchasAMTand

Appen,datalabelerssometimesworkthroughthird-partycompanieshiredbyleading

techfirms,inasubcontractingrelationship.

Althoughtherearestillmanydatalabelers

workingintheUnitedStatesinEurope,muchoftheworkisbeingdoneindevelopingcountries,giventhelowremunerationassociatedwiththework.Whileprecisefiguresonthenumbersofpersonsworkingasdatalabelersdonotexist,estimatesrangeinthetensofmillions,and

demandforsuchworkislikelytocontinueasAIdatasetsandtrainingneedsgrow[12].ThesizeofthemarketisestimatedatbetweenUS$1-$3billionandlikelytoexperiencedouble-digitgrowthoverthenext5years[13].

Datalabelingworkdoesnotrequiremany

qualifications,besidesliteracy,digitalskills

andaccesstocomputer(ormobiledevice)andinternet.StudiesofearningsofonlineplatformworkersintheUSthatperformthiswork,

regularlyreportmedianearningsofroughly$2-$3perhour,orwellbelowthefederalminimumwageofUS$7.25[14];[11].Giventhelowlevelofpay,itisunsurprisingthatmuchofthisworkhasmovedtodevelopingcountries.

Butevenfromadevelopingcountry

perspective,theearningsarelow,particularlyconsideringtheskillleveloftheworkforce,

withmanyworkersholdinguniversityand

post-graduatedegrees[11].Fortheworkerswhoworkthroughdigitallabourplatforms–andnotbusinessprocessoutsourcingfirms–thereistheaddedconcernthattheyare

hiredasindependentcontractorsandarethusnotcoveredbytheprotectionsandbenefits

emanatingfromastandardemployment

relationship.Moreover,analysesofearningsdifferentialsbetweenworkersinIndiadoingsimilartypesofdataannotationworkrevealedthatplatformworkersearnedtwo-thirds

lessthancomparable,non-platformworker

employees,evenbeforeaccountingforotherbenefitssuchassocialinsurancecontributions[15].

5SeeAmazonMechanicalTurkAPIReference-AmazonMechanicalTurk.Accessedon9June2024.

12|MindtheAIDivide:ShapingaGlobalPerspectiveontheFutureofWork

Butevenamongbusinessprocessoutsourcingfirms,thereareconcernsabouttheworking

conditionsoftheseworkers,withonecase

studyofadataannotationenterprisewith

officesinKenyarevealinglowpay,insecureworkandgender-basedworkplaceviolence[16].Furthermore,thestudyarguedthatthedataannotationskillsusedinthislineofworkwerenotessentiallytransferable,questioningthecareer-enhancingimpactofthislineof

work.

Movingalongthevaluechain,thesubsequentparts–modeldesign,modeltrainingand

tuning,deploymentandmaintenance–representacontrastingpicturewiththe