AI對英國就業(yè)和培訓(xùn)的影響

TheimpactofAIon

UKjobsandtraining

November2023

2

Contents

Acknowledgements3

TheimpactofAIonUKjobsandtraining4

Introduction4

Summary5

1Methodology6

1.1SelectionofAIapplications6

1.2Mappinghumanabilitiestojobroles7

1.3AssessingAIapplicationsagainsthumanabilities7

1.4Calculatingoccupationalexposure8

1.5Mappingoccupationstotrainingpathways9

1.6Datasources9

1.7ResearchbyInternationalMonetaryFund10

2OccupationalexposuretoAI11

2.1OccupationsmostexposedtoAI11

2.2ExposuretoAIbyskilllevelofoccupation13

3ExposuretoAIacrossindustriesandgeography16

3.1ExposuretoAIacrossindustry16

3.2ExposuretoAIbygeography17

4ExposuretoAIbyqualification18

4.1Trainingroutes18

4.2Subjectareas19

Annex1:Apprenticeships22

Annex2:Augmentationversussubstitution24

Annex3:ComparisontofindingsfromthePewResearchCenter26

Annex4:Furtheranalysisforoccupationsexposedtolargelanguagemodelling28

ExposuretoLLMacrossindustries28

ExposuretoLLMbygeography30

ExposuretoLLMbyqualification31

Trainingroutes31

Subjectareas32

3

Acknowledgements

TheauthorswouldliketoextendtheirthankstoEdwardFelten,RobertSeamansand

ManavRajwhopublishedtheresourcesfortheirresearch,allowingittobereusedfor

thisreport.TheywouldalsoliketothanktheNationalFoundationforEducationResearchandresearchersattheUniversityofSheffieldandtheUniversityofWarwickformakingavailablethelatestmappingbetweenSOC2020andO*NET.

4

TheimpactofAIonUKjobsandtraining

Introduction

AdvancesinArtificialIntelligence(AI)arewidelyexpectedtohaveaprofoundand

widespreadeffectontheUKeconomyandsociety,thoughtheprecisenatureandspeedofthiseffectisuncertain.Itisestimatedthat10-30%ofjobsareautomatablewithAI

havingthepotentialtoincreaseproductivityandcreatenewhighvaluejobsintheUKeconomy

.1,

2

TheUKeducationsystemandemployerswillneedtoadapttoensurethatindividualsintheworkforcehavetheskillstheyneedtomakethemostofthepotentialbenefits,advancesinAIwillbring.

Thisreport,producedbytheUnitforFutureSkills

3

intheDepartmentforEducation,is

oneofthefirstattemptstoquantifytheimpactofAIontheUKjobmarket(separateto

automationmoregenerally).TheresearchtakesamethodologyfromaUSbasedstudydevelopedbyFelteneta

l4

andappliesitforaUKcontext.Theapproachconsiderstheabilitiesneededtoperformdifferentjobroles,andtheextenttowhichthesecanbeaidedbyaselectionof10commonAIapplications

5.

Themethodologyisextendedfurtherto

considerthelinkbetweentrainingandjobsimpactedbyAI,usinganoveldatasetthatincludesinformationonthequalificationsheldbyyoungpeopleinemployment.

Resultsshouldbeinterpretedwithcaution

TheestimatesofwhichjobsaremoreexposedtoAIarebasedonanumberof

uncertainassumptionssotheresultsshouldbeinterpretedwithcaution.Quantifyingoccupationsintermsofabilitiestoperformajobrolewillneverfullydescribeallrolesandalevelofjudgementisrequiredwheninterpretingtheresults.Further,theextenttowhichoccupationsareexposedtoAIwillchangeduetothepaceatwhichAI

technologiesaredevelopingandasnewdatabecomesavailable.

However,thethemeshighlightedbytheanalysisareexpectedtocontinueandprovideagoodbasisforconsideringtherelativeimpactofAIacrossdifferentpartsofthe

labourmarket.

1

PwC,Willrobotsreallystealourjobs?

2

TheBritishInstituteAcademy,Theimpactofartificialintelligenceonwork

3

.uk/government/groups/unit-for-future-skills

4

FeltenE,RajM,SeamansR(2023)‘HowwillLanguageModelerslikeChatGPTAffectOccupationsand

Industries?’

5Abstractstrategygames;real-timevideogames;imagerecognition;visualquestionanswering;imagegeneration;readingcomprehension;languagemodelling;translation;speechrecognition;instrumentaltrackrecognition.

5

Summary

Thisreportshowstheoccupations,sectorsandareaswithintheUKlabourmarketthatareexpectedtobemostimpactedbyAIandlargelanguagemodelsspecifically.Italsoshowsthequalificationsandtrainingroutesthatmostcommonlyleadtothesehighlyimpactedjobs.Themainfindingsare:

?ProfessionaloccupationsaremoreexposedtoAI,particularlythose

associatedwithmoreclericalworkandacrossfinance,lawandbusiness

managementroles.Thisincludesmanagementconsultantsandbusiness

analysts;accountants;andpsychologists.TeachingoccupationsalsoshowhigherexposuretoAI,wheretheapplicationoflargelanguagemodelsisparticularly

relevant.

?Thefinance&insurancesectorismoreexposedtoAIthananyothersector.TheothersectorsmostexposedtoAIareinformation&communication;

professional,scientific&technical;property;publicadministration&defence;andeducation.

?WorkersinLondonandtheSouthEasthavethehighestexposuretoAI,

reflectingthegreaterconcentrationofprofessionaloccupationsinthoseareas.

WorkersintheNorthEastareinjobswiththeleastexposuretoAIacrosstheUK.However,overallthevariationinexposuretoAIacrossthegeographicalareasismuchsmallerthanthevariationobservedacrossoccupationsorindustries.

?Employeeswithhigherlevelsofachievementaretypicallyinjobsmore

exposedtoAI.Forexample,employeeswithalevel6qualification(equivalenttoadegree)aremorelikelytoworkinajobwithhigherexposuretoAIthan

employeeswithalevel3qualification(equivalenttoA-Levels).

?EmployeeswithqualificationsinaccountingandfinancethroughFurtherEducationorapprenticeships,andeconomicsandmathematicsthrough

HigherEducationaretypicallyinjobsmoreexposedtoAI.Employeeswithqualificationsatlevel3orbelowinbuildingandconstruction,manufacturing

technologies,andtransportationoperationsandmaintenanceareinjobsthatareleastexposedtoAI.

TheanalysismeasurestheexposureofjobstoAI,ratherthandistinguishingwhetherajobwillbeaugmented(aided)orreplaced(substituted)byAI.Researchbythe

InternationalLaborOrganization(ILO

)6

suggeststhatmostjobsandindustriesareonlypartlyexposedtoautomationandaremorelikelytobecomplementedratherthan

substitutedbygenerativeAIlikeChatGPT.Annex2mapsthejobshighlightedinthatreporttotheUKjobmarket,andgenerallyincludecustomerserviceandadministrative

occupations,includingcallandcontactcentreandunclassifiedadministrativeoccupations.

6

GenerativeAIandjobs:Aglobalanalysisofpotentialeffectsonjobquantityandquality()

6

1Methodology

ThemethodologybroadlyfollowstheapproachdescribedbyFelteneta

l7

tocreateanAIOccupationalExposure(AIOE)score,withsomeadaptationstomakeitsuitableforaUKcontext.

1.1SelectionofAIapplications

TheAIOEisconstructedbasedonassumptionsaroundtheuseofadefinedsetof

commonAIapplications.The10AIapplicationsselectedarebasedonthosewheretheElectronicFrontierFoundation(EFF)hasrecordedscientificactivityandprogressinthetechnologyfrom2010onwards.

Table1:AIapplications

AIapplication

Definition

Abstractstrategygames

Theabilitytoplayabstractgamesinvolving

sometimescomplexstrategyandreasoningability,suchaschess,go,orcheckers,atahighlevel.

Real-timevideogames

Theabilitytoplayavarietyofreal-timevideogamesofincreasingcomplexityatahighlevel.

Imagerecognition

Thedeterminationofwhatobjectsarepresentinastillimage.

Visualquestionanswering

Therecognitionofevents,relationships,andcontextfromastillimage.

Imagegeneration

Thecreationofcompleximages.

Readingcomprehension

Theabilitytoanswersimplereasoningquestionsbasedonanunderstandingoftext.

Languagemodelling

Theabilitytomodel,predict,ormimichumanlanguage.

Translation

Thetranslationofwordsortextfromonelanguageintoanother.

Speechrecognition

Therecognitionofspokenlanguageintotext.

Instrumentaltrackrecognition

Therecognitionofinstrumentalmusicaltracks.

ThissetofapplicationsdoesnotcomprehensivelycoverthesetofapplicationsforwhichAIcouldultimatelybeused;however,basedonfurtherworkconductedbyFeltenetalwithfieldexperts,itisbelievedthattheserepresentfundamentalapplicationsofAIthat

7FeltenE,RajM,SeamansR(2023)HowwillLanguageModelerslikeChatGPTAffectOccupationsandIndustries?

7

arelikelytohaveimplicationsfortheworkforceandareapplicationsthatcoverthemostlikelyandmostcommonusesofAI.

1.2Mappinghumanabilitiestojobroles

ThemethodologybyFeltenetalusestheOccupationalInformationNetwork(O*NET)

databaseofoccupationalcharacteristicsandworkerrequirementsinformationacrosstheUSeconomy

.8

ThereiscurrentlynoequivalentdatabaseforUKoccupation

s9

sothe

O*NETdataismappedtotheUKusingacrosswalkbetweenO*NEToccupationsandSOC2010.

TheO*NETsystemuses52distinctabilitiestodescribetheworkplaceactivitiesofeachoccupation,eachwithaseparatescorefor‘level’and‘importance’.Abilitiesaregroupedunderfourcategories:cognitive,physical,psychometerandsensory.Examplesof

abilitiesareoralcomprehension,writtenexpression,mathematicalreasoning,manualdexterity,andstamina

.10

SOC2010wasusedinsteadofSOC2020toalignwithinformationontrainingpathwaysandduetoknownissueswithSOC2020

11

.UpdatingtheanalysistoSOC2020willleadtosmallchangesintheorderingofAIOEscoresbutnottheoverallfindings.

1.3AssessingAIapplicationsagainsthumanabilities

AIapplicationsarelinkedtoworkplaceabilitiesusingacrowd-sourceddatasetcollectedbyFeltenetal,andconstructedusingsurveyresponsesof“gigworkers”fromAmazon'sMechanicalTurk(mTurk)webservice.Thedatahasameasureofapplication-ability

relatednessforeachcombinationboundbetween0and1.Thismeasureofapplication-abilityrelatednessisthenorganisedintoamatrixthatconnectsthe10AIapplicationstothe52O*NEToccupationalabilities.Anability-levelexposureiscalculatedasfollows:

Aij=xij

(1)

Inthisequation,iindexestheAIapplicationandjindexestheoccupationalability.The

ability-levelexposure,A,iscalculatedasthesumofthe10application-abilityrelatednessscores,x,asconstructedusingmTurksurveydata.Bycalculatingtheability-levelAI

8FeltenE,RajM,SeamansR(2023)HowwillLanguageModelerslikeChatGPTAffectOccupationsandIndustries?

9

.uk/government/publications/a-skills-classification-for-the-uk

10

O*NET28.0DatabaseatO*NETResourceCenter()

11

RevisionofmiscodedoccupationaldataintheONSLabourForceSurvey,UK-OfficeforNational

Statistics

8

exposureasasumofalltheAIapplications,allapplicationsareweightedequall

y12.

Thisapproachassumesthateachapplicationhasanindependenteffectonanabilityand

doesnotconsiderinteractionsacrossapplications.

Theestimatesforeachapplicationarethenstandardisedtogivearatingbetween0and1.

1.4Calculatingoccupationalexposure

Foreachoccupation,thevaluesforthelevelandimportanceofeachabilityarecombinedwiththeratingfortherelatednessofeachAIapplicationtocreateanAIOccupational

Exposure(AIOE)score.ThisisdoneoverallforallAIapplications,andindividuallyforeachapplication,e.g.languagemodelling.

∑1Aij×Ljk×Ijk

∑1Ljk×Ijk

AIOEk=

(2)

Inthisequation,iindexestheAIapplication,jindexestheoccupationalability,andk

indexestheoccupation.Aijrepresentstheability-levelexposurescorecalculatedin

Equation1.Theability-levelAIexposureisweightedbytheability'sprevalence(Ljk)andimportance(Ijk)withineachoccupationasmeasuredbyO*NET(mappedtoSOC2010)bymultiplyingtheability-levelAIexposurebytheprevalenceandimportancescoresforthatabilitywithineachoccupation,scaledsothattheyareequallyweighted.These

prevalenceandimportancescores,accountforthepresenceofdifferentabilitieswithinanoccupation.Abilitiesthatareintegraltoanoccupationhavehighprevalenceand

importancescores,whilethosethatareusedlessoftenorarelessvitalhavelower

prevalenceandimportancescores.Anoccupation'saggregateexposuretoAIis

calculatedbysummingthisweightedability-levelAIexposureacrossallabilitiesinanoccupation.Thescoresarethenstandardisedandrankedfrommosttoleastexposed.Thesescoresareappliedtoemploymentcountsacrossoccupationstogiveaggregateexposurescores,forexampleacrossthegeographicalareas.

IntestingtherobustnessoftheirmethodologyFeltenetalfoundevidencethatAIismostlikelytoaffectcognitiveandsensoryabilities,andtheAIOEscoreswerenotsensitivetoexcludinganyoftheapplicationsinthesample.Therefore,anyAIapplicationsthatmayhavebeenexcludedarealsolikelytoberelatedtoasimilarsetofcognitiveandsensoryabilities.

12Feltenetalcarriedoutfurtheranalysiswhichsuggestedthatweightingtheapplicationsisunlikelytohaveameaningfulimpactonthemeasure.

9

1.5Mappingoccupationstotrainingpathways

RelationshipsbetweenoccupationsandtrainingaretakenfromASHE-LEOdata,anewdataresourceavailableintheDepartmentforEducation.Itbringstogetherthe

longitudinaleducationandlabourmarketinformationintheLongitudinalEducation

Outcomesstudy(LEO

)13

withtheinformationonemploymentandearningsintheAnnualSurveyofHoursandEarnings(ASHE)

.14

Therearearound100,000individualsintheASHE-LEOsampleineachyear.This

represents45-75%oftheoverallASHEsample,withlateryearshavingabettermatchratethanearlieryears,andyoungerageshavingabettermatchratethanolderages.ASHE-LEOisusedhereasanapproximatelyrepresentativesampleofearlycareer

employeesinLEO(employeesaged23-30inthe2018-19taxyear).

Thedataisusedtoidentifythetrainingtakenbyemployeesforeachoccupation.Aseachtrainingroutemaybeassociatedwithmultipleoccupations,aweightedaverageis

calculatedtoarriveatanaverageAIOEscore.

1.6Datasources

Name

Description

AIOEdata1

5

Organisedmeasureofapplication-abilityrelatednessthatconnectsthe10EFFAIapplicationstothe52O*NEToccupationalabilities.

AnnualPopulationSurvey

Aresidencebasedlabourmarketsurvey

encompassingpopulation,economicactivity

(employmentandunemployment),economicinactivityandqualifications.

Apprenticeshipdata

ApprenticeshipsstartsinEnglandreportedforan

academicyearbasedondatareturnedbyproviders.

ASHE-LEO

Educationandlabourmarketinformationinthe

LongitudinalEducationOutcomesstudy(LEO)linkedwiththeinformationonemploymentandearningsintheAnnualSurveyofHoursandEarnings(ASHE)

13

ApplytoaccesstheLongitudinalEducationOutcomes(LEO)dataset-GOV.UK(.uk)

14

AnnualSurveyofHoursandEarnings(ASHE)-OfficeforNationalStatistics(.uk)

15FeltenE,RajM,SeamansR(2021)Occupational,industry,andgeographicexposuretoartificialintelligence:Anoveldatasetanditspotentialuses.StrategicManagementJournal42(12):2195–2217

10

1.7ResearchbyInternationalMonetaryFund

TheInternationalMonetaryFund(IMF)haveconstructedacomplementarityadjustedAIoccupationalexposure(C-AIOE)measure,wheretheexposureofoccupationstoAIaremitigatedbytheirpotentialforcomplementarity

.16

AtahighleveltheauthorsofthisstudymakeanadjustmenttotheFeltenetal

methodologyforAIOE

17

tocapturethepotentialtocomplementorsubstituteforlabourineachoccupation.Theythenapplyboththeoriginalmeasureandthecomplementarity

adjustedmeasurestolabourforcemicrodata(usingISCO-08)from6countriesincludingtheUK,withaparticularfocusonemergingmarkets.

Theresearchfindsthattherearesubstantialcross-countrydisparitiesinthebaseline

AIOE,withemergingmarketsgenerallydisplayinglowerexposurelevelsthanadvanced

economies.Thisdisparityismainlyduetodifferentemploymentstructures,with

advancedeconomiescharacterisedbylargerproportionsofhigh-skilloccupationssuchasprofessionalsandmanagers.InlinewiththisreportandasoutlinedbyFeltenetal,

theseprofessionsarethemostexposedtoAIduetotheirhighconcentrationofcognitive-basedtasks.However,becausethosehigh-skilloccupationsalsoshowhigherpotentialforAIcomplementarity,thesecross-countrydisparitiesintermsofpotentiallydisruptiveexposurereduceconsiderablyoncecomplementarityisfactoredin.Nevertheless,

advancedeconomiesremainmoreexposedevenundertheC-AIOEmeasure.Emergingmarketswithalargeshareofagriculturalemployment,remainrelativelylessexposed

underbothmeasures,asoccupationsinthissectorhaveverylowbaselineexposuretoAI.Overall,theresultssuggestthattheimpactofAIonlabourmarketsinadvanced

economiesmaybemore“polarised,”astheiremploymentstructurebetterpositionsthemtobenefitfromgrowthopportunitiesbutalsomakesthemmorevulnerabletolikelyjob

displacements.

16

LaborMarketExposuretoAI:Cross-countryDifferencesandDistributionalImplications()

17

FeltenE,RajM,SeamansR(2023)‘HowwillLanguageModelerslikeChatGPTAffectOccupationsand

Industries?’

11

2OccupationalexposuretoAI

ThereisarangeofUKandinternationalresearchonAIandtheimpactthatitwillhaveonjobsandthelabourmarket.Itisverydifficulttomakeanumericalestimateona

technologywhichisnotyetfullyunderstoodandisevolvingatarapidpace.However,aconsensushasbeguntoemergethat10-30%ofjobsintheUKarehighlyautomatableandcouldbesubjecttosomelevelofautomationoverthenexttwodecades.However,theoverallneteffectonemploymentisunclearbutitisoftenassumedthattherewillbeabroadlyneutrallong-termeffectandjobdisplacementwillbematchedbyjobcreation

.18

ThisanalysisassessestherelativeexposureofUKjob

s19

toAIbyuseofanAI

OccupationalExposure(AIOE)score.TheAIOEscoreallowsjobstoberankedto

showwhichjobsaremoreandlesslikelytobeimpactedbyadvancesinAI,basedontheabilitiesrequiredtoperformthejob.AswellasAIgenerally,asimilarexposurescoreiscreatedtoconsiderlargelanguagemodellingspecificallythroughgenerativeAItoolslikeChatGPTandBard.

TheanalysismeasurestheexposureofjobstoAI,ratherthandistinguishingwhetherajobwillbeaugmented(aided)orreplaced(substituted)byAI.Annex2discussesthepotentialforidentifyingUKjobswhichcouldbefullyautomatedasaresultofAIbasedonresearchfromtheInternationalLaborOrganization(ILO).

2.1OccupationsmostexposedtoAI

Table2

showsalistofthetop20occupationsthataremostexposedtoAI,andtolargelanguagemodellingspecifically.Afulllistofalloccupationsispublishedalongsidethisreport.

Theexposurescoreisbasedonanumberofassumptionsincludingtheabilities

consideredimportantforajobatagivenpointintimesorankingsshouldbe

interpretedwithcaution,howeverthethemeshighlightedbytheanalysisareexpectedtocontinu

e20.

TheoccupationsmostexposedtoAIincludemoreprofessionaloccupations,particularlythoseassociatedwithmoreclericalworkandacrossfinance,lawandbusiness

managementroles.Thisincludesmanagementconsultantsandbusinessanalysts,

accountants,andpsychologists.ThiscomparestotheoccupationsleastexposedtoAI,whichincludesportprofessionals,roofersandsteelerectors.

ThelistofoccupationsmostexposedtolargelanguagemodellingincludesmanyofthesameoccupationsexposedtoAImoregenerally,withbothlistsincludingsolicitors,

18

Willrobotsreallystealourjobs?(pwc.co.uk)

19Definedby4digitstandardisedoccupationclassification(SOC2010)codes.

20Feltenetal(2021)AppendixC:QuantitativeValidationoftheAIOEandRelatedMeasures

12

psychologistsandmanagementconsultantsandbusinessanalysts.Italsoincludesmoreeducationrelatedoccupations,particularlyforpost-16training.Thisalignswithpublic

statementsaroundthepotentialuseofgenerativeAItoolsbyteachers,forexampleinpreparingteachingmaterial.

Table2:OccupationsmostexposedtoAIandlargelanguagemodelling

ExposuretoallAIapplications

Exposuretolargelanguage

modelling

1

Managementconsultantsandbusinessanalysts*

Telephonesalespersons

2

Financialmanagersanddirectors

Solicitors*

3

Chartedandcertifiedaccountants

Psychologists*

4

Psychologists*

Furthereducationteaching

professionals

5

Purchasingmanagersanddirectors

Marketandstreettradersand

assistants

6

Actuaries,economistsandstatisticians

Legalprofessionalsn.e.c.*

7

Businessandfinancialproject

managementprofessionals

Creditcontrollers*

8

Financeandinvestmentanalystsandadvisers

Humanresourceadministration

occupations*

9

Legalprofessionalsn.e.c.*

Publicrelationsprofessionals

10

Businessandrelatedassociate

professionalsn.e.c.

Managementconsultantandbusinessanalysts*

11

Creditcontrollers*

Marketresearchinterviewers

12

Solicitors*

Localgovernmentadministrativeoccupations

13

Civilengineers

Clergy

14

Educationadvisersandschool

inspectors*

Highereducationteaching

professionals

15

Humanresourcesadministrative

occupations*

Collectorsalespersonsandcreditagents

16

Business,researchandadministrativeprofessionalsn.e.c.

Educationadvisersandschool

inspectors*

17

Financialaccountsmanagers

Humanresourcemanagersand

directors

18

Bookkeepers,payrollmanagersandwagesclerks

Nationalgovernmentadministrativeoccupations*

19

Nationalgovernmentadministrativeoccupations*

Vocationalandindustrialtrainersandinstructors

20

Marketingassociateprofessionals

Socialandhumanitiesscientists

*Occupationsthatappearinbothlistsaremarkedwithanasterisk.

Table3

showsalistoftheoccupationsthatareleastexposedtoAI,andtolargelanguagemodellingspecifically.

TheoccupationsleastexposedtoAIandLLMincludemanyofthesameareas,includingmoremanualworkthatistechnicallydifficult,inunpredictableenvironments,andwith

13

lowerwages(reducingtheincentivetoautomate)–withtheexceptionofsportsplayers.Thisincludes:roofers,rooftilersandslaters;elementaryconstructionoccupations;

plasterers;andsteelerectors.

Table3:OccupationsleastexposedtoAIandlargelanguagemodelling

ExposuretoallAIapplications

Exposuretolargelanguage

modelling

1

Sportsplayers*

Fork-lifttruckdrivers*

2

Roofers,rooftilersandslaters*

Roofers,rooftilersandslaters*

3

Elementaryconstructionoccupations*

Steelerectors*

4

Plasterers*

Vehiclevaletersandcleaners*

5

Steelerectors*

Elementaryconstructionoccupations*

6

Vehiclevaletersandcleaners*

Plasterers*

7

Hospitalporters

Metalplateworkers,andriveters*

8

Cleanersanddomestics

Vehiclepainttechnicians

9

Floorersandwalltilers*

Floorersandwalltilers*

10

Metalplateworkers,andriveters

Mobilemachinedriversandoperativesn.e.c.

11

Launderers,drycleanersandpressers*

Launderers,drycleanersand

pressers*

12

Windowcleaners

Largegoodsvehicledrivers

13

Paintersanddecorators

Roadconstructionoperatives*

14

Fork-lifttruckdrivers*

Railconstructionandmaintenanceoperatives

15

Packers,bottlers,cannersandfillers

Industrialcleaningprocess

occupations

16

Gardenersandlandscapegardeners

Elementaryprocessplantoccupationsn.e.c.