國(guó)際勞工組織-ChatGPT等生成式AI能增強(qiáng)就業(yè)不會(huì)取代崗位（英）

ILO

WorkingPaper

96August

/

2023Generative

AI

and

Jobs:

A

globalanalysis

of

potential

effects

on

jobquantity

and

qualityXAuthors

/

Pawe?

Gmyrek,

Janine

Berg,

David

BescondCopyright

?

InternationalLabourOrganization

2023This

is

an

open

access

work

distributed

under

the

Creative

Commons

Attribution

4.0

InternationalLicense

(/licenses/by/4.0/).

Users

can

reuse,

share,

adapt

and

buildupon

the

original

work,

as

detailed

in

the

License.

The

ILO

must

be

clearly

credited

as

the

own-er

of

the

original

work.

The

use

of

the

emblem

of

the

ILO

is

not

permitted

in

connection

withusers’work.Attribution

–

The

work

must

be

cited

as

follows:

Gmyrek,

P.

,

Berg,

J.,

Bescond,

D.

Generative

AIand

Jobs:

A

global

analysis

of

potential

e?ects

on

job

quantity

and

quality.

ILO

Working

Paper

96.Geneva:

InternationalLabourO?ce,2023.Translations–

In

case

of

a

translation

of

this

work,

the

following

disclaimer

must

be

added

alongwith

the

attribution:

This

translation

was

not

created

by

the

International

Labour

Organization

(ILO)and

should

not

be

considered

an

o?cial

ILO

translation.

The

ILO

is

not

responsible

for

the

content

oraccuracy

of

this

translation.Adaptations

–

In

case

of

an

adaptation

of

this

work,

the

following

disclaimer

must

be

addedalong

with

the

attribution:

This

is

an

adaptation

of

an

original

work

by

the

International

LabourOrganization

(ILO).

Responsibility

for

the

views

and

opinions

expressed

in

the

adaptation

rests

solelywith

the

author

or

authors

of

the

adaptation

and

are

not

endorsed

by

the

ILO.This

CC

license

does

not

apply

to

non-ILO

copyright

materials

included

in

this

publication.

If

thematerial

is

attributed

to

a

third

party,the

user

of

such

material

is

solely

responsible

for

clearingtherightswiththerightholder.Any

dispute

arising

under

this

license

that

cannot

be

settled

amicably

shall

be

referred

to

arbitra-tion

in

accordance

with

the

Arbitration

Rules

of

the

United

Nations

Commission

on

InternationalTrade

Law

(UNCITRAL).

The

parties

shall

be

bound

by

any

arbitration

award

rendered

as

a

resultofsucharbitration

asthe?naladjudicationofsucha

dispute.All

queries

on

rights

and

licensing

should

be

addressed

to

the

ILO

Publishing

Unit

(Rights

andLicensing),1211Geneva

22,Switzerland,orby

emailto

rights@.ISBN

9789220395356

(print),

ISBN

9789220395363

(web

PDF),

ISBN

9789220395370

(epub),

ISBN9789220395387

(mobi),

ISBN

9789220395394

(html).

ISSN

2708-3438

(print),

ISSN

2708-3446(digital)/10.54394/FHEM8239The

designations

employed

in

ILO

publications,

which

are

in

conformity

with

United

Nationspractice,

and

the

presentation

of

material

therein

do

not

imply

the

expression

of

any

opinionwhatsoever

on

the

part

of

the

ILO

concerning

the

legal

status

of

any

country,

area

or

territoryorofitsauthorities,orconcerningthedelimitationofitsfrontiers.The

responsibility

for

opinions

expressed

in

signed

articles,

studies

and

other

contributions

restssolely

with

their

authors,

and

publication

does

not

constitute

an

endorsement

by

the

ILO

of

theopinionsexpressed

inthem.Reference

to

names

of

?rms

and

commercial

products

and

processes

does

not

imply

their

en-dorsement

by

the

ILO,

and

any

failure

to

mention

a

particular

?rm,

commercial

product

or

pro-cessisnota

signofdisapproval.InformationonILO

publicationsanddigitalproducts

canbefoundat:/publnsILO

Working

Papers

summarize

the

results

of

ILO

research

in

progress,

and

seek

to

stimulatediscussion

of

a

range

of

issues

related

to

the

world

of

work.

Comments

on

this

ILO

Working

Paperare

welcomeandcanbesenttoRESEARCH@,

berg@.Authorizationforpublication:Richard

Samans,Director

RESEARCHILO

Working

Paperscanbefoundat:/global/publications/working-papersSuggested

citation:Gmyrek,

P.,

Berg,

J.,

Bescond,

D.

2023.

Generative

AI

and

Jobs:A

global

analysis

of

potential

ef-fects

on

job

quantity

and

quality,

ILO

Working

Paper

96

(Geneva,

ILO)./10.54394/FHEM823901ILOWorkingPaper96AbstractThis

study

presents

a

global

analysis

of

the

potential

exposure

of

occupations

and

tasks

toGenerative

AI,

and

speci?cally

to

Generative

Pre-Trained

Transformers

(GPTs),

and

the

possibleimplications

of

such

exposure

for

job

quantity

and

quality.

It

uses

the

GPT-4

model

to

estimatetask-level

scores

of

potential

exposure

and

then

estimates

potential

employment

e?ects

at

theglobal

level

as

well

as

by

country

income

group.

Despite

representing

an

upper-bound

estimateof

exposure,

we

?nd

that

only

the

broad

occupation

of

clerical

work

is

highly

exposed

to

the

tech-nology

with

24

per

cent

of

clerical

tasks

considered

highly

exposed

and

an

additional

58

percentwith

medium-level

exposure.

For

the

other

occupational

groups,

the

greatest

share

of

highly

ex-posed

tasks

oscillates

between1

and4

per

cent,

and

medium

exposed

tasks

do

not

exceed

25per

cent.

As

a

result,

the

most

important

impact

of

the

technology

is

likely

to

be

of

augmentingwork

–

automating

some

tasks

within

an

occupation

while

leaving

time

for

other

duties

–

as

op-posedtofullyautomatingoccupations.The

potential

employment

e?ects,

whether

augmenting

or

automating,

vary

widely

across

coun-try

income

groups,

due

to

di?erent

occupational

structures.

In

low-income

countries,

only

0.4

percent

of

total

employment

is

potentially

exposed

to

automation

e?ects,

whereas

in

high-incomecountries

the

share

rises

to

5.5

percent.

The

e?ects

are

highly

gendered,

with

more

than

doublethe

share

of

women

potentially

a?ected

by

automation.

The

greater

impact

is

from

augmenta-tion,

which

has

the

potential

to

a?ect

10.4

percent

of

employment

in

low-income

countries

and13.4

percent

of

employment

in

high-income

countries.

However,

such

e?ects

do

not

considerinfrastructure

constraints,

which

will

impede

the

possibility

for

use

in

lower-income

countriesandlikelyincrease

theproductivity

gap.We

stress

that

the

primary

value

of

this

analysis

is

not

the

precise

estimates,

but

rather

the

in-sights

that

the

overall

distribution

of

such

scores

provides

about

the

nature

of

possible

changes.Such

insights

can

encourage

governments

and

social

partners

to

proactively

design

policies

thatsupport

orderly,

fair,

and

consultative

transitions,

rather

than

dealing

with

change

in

a

reactivemanner.

Moreover,

the

likely

rami?cations

on

job

quality

might

be

of

greater

consequence

thanthe

quantitative

impacts,

both

with

respect

to

the

new

jobs

created

because

of

the

technology,butalsothepotentiale?ectsonworkintensityandautonomywhenthetechnologyisintegrat-ed

into

the

workplace.

For

this

reason,

we

also

emphasize

the

need

for

social

dialogue

and

reg-ulationtosupportqualityemployment.About

the

authorsPawe?

Gmyrek

isSeniorResearcher

intheResearch

DepartmentoftheILO.Janine

Berg

isSeniorEconomistintheResearch

DepartmentoftheILO.David

Bescond

isDataScientistintheILO’s

DepartmentofStatistics.02ILOWorkingPaper96Table

ofcontentsAbstract010105AbouttheauthorsAcronymsXIntroduction07X

1

MethodsandData1011121.1.ISCO

dataonoccupationsandtasks1.2.Prompt

designandsequenceX

2

AssessmentofthePredictions,

RobustnessTests

andtheBoundsforAnalysis17X

3

Results20243.1.Automationvsaugmentation:distributionofscores

across

tasksandoccupationsX

4

Exposedoccupationsasashare

ofemployment:

globalandincome-basedestimates303032364.1.AugmentationvsAutomation:ILO

microdata4.2.AugmentationvsAutomation:globalestimate4.3.ThebigunknownX

5

Managingthetransition:

Policiestoaddress

automation,augmentationandthegrowing

digitaldivide383839405.1Mitigatingthenegativee?ectsofautomation5.2Ensuringjobqualityunderaugmentation5.3AddressingthedigitaldivideXConclusion43Appendix1.CountrieswithmissingISCO-08

4-digitdata:estimationprocedure454751ReferencesAcknowledgementsanduseofGPT03ILOWorkingPaper96List

of

FiguresFigure

1.Meanautomationscores

by

occupation,basedonISCO

andGPTtasks212425272829Figure

2.Tasks

withmediumandhighGPT-exposure,

by

occupationalcategory(ISCO

1-digit)Figure

3.Box

plotoftask-levelscores

by

ISCO

4d,grouped

by

ISCO

1dFigure

4.AugmentationvsautomationpotentialatoccupationallevelFigure

5.OccupationswithhighautomationpotentialFigure

6.OccupationswithhighaugmentationpotentialFigure

7a.Automationvsaugmentationpotential:shares

oftotalemployment,microdatafor59countries30Figure

7b.Automationvsaugmentationpotential:shares

oftotalemploymentineachsex(ILO

microdata)3133Figure

8.Countrycoverage

basedonthelevel

ofdigitsinISCO-08

(ILO

data)Figure

9a.Globalestimates:jobswithaugmentationandautomationpotentialasshare

oftotalemployment34Figure

9b.Automationvsaugmentationpotential:shares

oftotalemploymentforeachsex(globalestimate)3536Figure

10.Occupationswithhighautomationpotential,by

ISCO

4-digitandincomegroupFigure

11a.The“BigUnknown”:

occupationsbetweenaugmentationandautomationpotential

37Figure

11b.The“BigUnknown”:

share

oftotalemployment,by

incomegroup

(globalestimate)Figure

11.Share

ofpopulationnotusingtheinternet374142Figure

12.Aclassicgrowth

path:incomeandoccupationaldiversi?cation04ILOWorkingPaper96List

of

TablesTable

1.ISCO-08

Structureofoccupationsandtasksusedinthestudy1114151722Table

2.Sampleoftasksandde?nitionsfrom

ISCO

andpredictedby

GPT-4Table3.Sampleoftask-levelscores(high-incomecountrycontext)Table4.aTestofscoreconsistency(100task-levelpredictions)Table4.bTaskswithhighautomationpotentialclusteredintothematicgroups*2632Table5.Groupingofoccupationsbasedontask-levelscoresTable6.MicrodatacoveragebylevelsISCO-08:numberofcountries05ILOWorkingPaper96Acronyms3GThird

Generation

(referring

to

a

generation

of

standards

for

mobile

telecom-munications)AdaA

languagemodelby

OpenAIusedtogenerate

embeddingsArti?cialGeneral

IntelligenceAGIAIArti?cialIntelligenceANNAPIArti?cialNeural

NetworkApplicationProgramming

InterfaceAutomatedTeller

MachinesATMsCPUDLCentral

Processing

UnitDeepLearningDOLEESCOGPTsGPT-4GPUHICDepartmentofLaborandEmploymentEuropean

Skills,Competences,Quali?cationsandOccupationsGenerative

Pre-Trained

TransformersGenerative

Pre-Trained

Transformer

4Graphics

Processing

UnitHigh-IncomeCountriesICTInformationandCommunicationsTechnologyInternationalLabourOrganizationInternationalStandard

Classi?cationofOccupationsInternationalStandard

Classi?cationofOccupations2008K-MeansClusteringAlgorithmILOISCOISCO-08K-MeansLFSLabourForce

SurveysLICLow-Income

CountriesLLMsLarge

LanguageModels06ILOWorkingPaper96LMICLower-Middle-Income

CountriesMachineLearningMLNLPNatural

LanguageProcessingOrganisation

forEconomicCo-operation

andDevelopmentOccupationalInformationNetworkOpenArti?cialIntelligence(organization's

name)High-level

programming

languageReinforcement

LearningOECDO*NETOpenAIPythonRLSDStandard

DeviationSMEsUMICUSSmallandMedium-sizedEnterprisesUpper-Middle-IncomeCountriesUnitedStatesUSDUMICUSUnitedStatesDollarUpper-Middle-IncomeCountriesUnitedStates07ILOWorkingPaper96IntroductionXEach

new

wave

of

technological

progress

intensi?es

debates

on

automation

and

jobs.

Currentdebates

on

Arti?cial

Intelligence

(AI)

and

jobs

recall

those

of

the

early

1900s

with

the

introduc-tion

of

the

moving

assembly

line,

or

even

those

of

the

1950s

and

1960s,

which

followed

the

intro-duction

of

the

early

mainframe

computers.

While

there

have

been

some

nods

to

the

alienationthat

technology

can

bring

by

standardizing

and

controlling

work

processes,

in

most

cases,

thedebates

have

centred

on

two

opposing

viewpoints:

the

optimists,

who

view

new

technology

asthe

means

to

relieve

workers

from

the

most

arduous

tasks,

and

the

pessimists,

who

raise

alarmabouttheimminentthreat

tojobsandtheriskofmassunemployment.What

has

changed

in

debates

on

technology

and

workers,

however,

is

the

types

of

workers

af-fected.

While

the

advances

in

technology

in

the

early,

mid

and

even

late-1900s

were

primarilyfocused

on

manual

workers,

technological

development

since

the

2010s,

in

particular

the

rapidprogress

of

Machine

Learning

(ML),

has

centred

on

the

ability

of

computers

to

perform

non-rou-tine,

cognitive

tasks,

and

by

consequence

potentially

a?ect

white-collar

or

knowledge

workers.In

addition,

these

technological

advancements

have

occurred

in

the

context

of

much

strong-er

interconnectedness

of

economies

across

the

globe,

leading

to

a

potentially

larger

exposurethan

location-based,

factory-level

applications.

Yet

despite

these

developments,

to

an

averageworker,

even

in

the

most

highly

developed

countries,

the

potential

implications

of

AI

have,

untilrecently,

remained

largely

abstract.The

launch

of

ChatGPT

marked

an

important

advance

in

the

public’s

exposure

to

AI

tools.

In

thisnew

wave

of

technological

transformation,

machine

learning

models

have

started

to

leave

thelabs

and

begin

interacting

with

the

public,

demonstrating

their

strengths

and

weaknesses

indaily

use.

The

chat

function

dramatically

shortened

the

distance

between

AI

and

the

end

user,simultaneously

providing

a

platform

for

a

wide

range

of

custom-made

applications

and

inno-vations.

Given

these

signi?cant

advancements,

it

is

not

surprising

that

concerns

over

potentialjoblosshaveresurged.While

it

is

impossible

to

predict

how

generative

AI

will

further

develop,

the

current

capabilitiesand

future

potential

of

this

technology

are

central

to

discussions

of

its

impact

on

jobs.

Scepticstend

to

believe

that

these

machines

are

nothing

more

than

“stochastic

parrots”–

powerful

textsummarizers,

incapable

of

“l(fā)earning”

and

producing

original

content,

with

little

future

for

gen-eral

purpose

use

and

unsustainable

computing

costs

(Bender

et

al.

2021).

On

the

other

hand,more

recent

technical

literature

focused

on

testing

the

limits

of

the

latest

models

suggests

anincreasing

capability

to

carry

out

“novel

and

di?cult

tasks

that

span

mathematics,

coding,

vision,medicine,

law,

psychology

and

more”,

and

a

general

ability

to

produce

responses

exhibiting

someforms

of

early

“reasoning”

(Bubeck

et

al.

2023).

Some

assessments

go

as

far

as

suggesting

thatmachine

learning

models,

especially

those

based

on

large

neural

networks

used

by

GenerativePre-trained

Transformers

(GPT,

see

Text

Box

1),

might

have

the

potential

to

eventually

become

ageneral-purpose

technology

(Goldfarb,

Taska,

and

Teodoridis

2023;

Eloundou

et

al.

2023).1

Thiswould

have

multiplier

e?ects

on

the

economy

and

labour

markets,

as

new

products

and

servic-eswouldlikelyspringfrom

thistechnologicalplatform.As

social

scientists,

we

are

not

in

position

to

take

sides

in

these

technical

debates.

Instead,

wefocus

on

the

already

demonstrated

capabilities

ofGPT-4,including

custom-made

chatbots

withretrieval

of

private

content

(such

as

collections

documents,

e-mails

and

other

material),

natu-ral

language

processing

functions

of

content

extraction,

preparation

of

summaries,

automatedcontent

generation,

semantic

text

searches

and

broader

semantic

analysis

based

on

text

em-beddings.

Large

Language

Models

(LLMs)

can

also

be

combined

with

other

ML

models,

such

as1The

three

main

characteristics

of

general-purpose

technologies

are

pervasiveness,

ability

to

continue

improving

over

time,

and

abil-itytospawnfurtherinnovation

(Jovanovic

andRousseau,2005).08ILOWorkingPaper96speech-to-text

and

text-to-speech

generation,

potentially

expanding

their

interaction

with

dif-ferent

types

of

human

tasks.

Finally,

the

potential

of

interacting

with

live

web

content

throughcustom

agents

and

plugins,

as

well

as

the

multimodal

(not

exclusive

to

text,

but

also

capable

ofreading

and

generating

image)

character

of

GPT-4

makes

it

likely

that

this

type

of

technologywillexpand

intonew

areas,

thereby

increasing

itsimpactonlabour.Departing

from

these

observations,

this

study

seeks

to

add

the

global

perspective

to

the

alreadylively

debate

on

possible

changes

that

may

result

in

the

labour

markets

as

a

consequence

of

therecent

advent

of

generative

AI.

We

stress

the

focus

of

our

work

on

the

concepts

of

“exposure”and

“potential”,

which

does

not

imply

automation,

but

rather

lists

occupations

and

associatedemployment

?gures

for

jobs

that

are

more

likely

to

be

a?ected

by

GPT-4

and

similar

technologiesin

the

coming

years.

The

objective

of

this

exercise

is

not

to

derive

headline

?gures,

but

rather

toanalyse

the

direction

of

possible

changes

in

order

to

facilitate

the

design

of

appropriate

policyresponses,

includingthepossibleconsequencesonjobquality.The

analysis

is

based

on

4-digit

occupational

classi?cations

and

their

corresponding

tasks

in

theISCO-08

standard.

It

uses

the

GPT-4

model

to

estimate

occupational

and

task-level

scores

of

ex-posure

to

GPT

technology

and

subsequently

links

these

scores

to

o?cial

ILO

statistics

to

deriveglobal

employment

estimates.

Wealso

apply

embedding-based

text

analysis

and

semantic

clus-tering

algorithms

to

provide

a

better

understanding

of

the

types

of

tasks

that

have

a

high

auto-mation

potential

and

discuss

how

the

automating

and

augmenting

e?ects

will

strongly

dependona

range

ofadditionalfactorsandspeci?ccountrycontext.We

discuss

the

results

of

this

analysis

in

the

broader

context

of

labour

market

transformations.We

put

particular

focus

on

the

current

disparities

in

digital

access

across

countries

of

di?erentincome

levels,

the

potential

for

this

new

wave

of

technological

transformation

to

aggravate

suchdisparities,

and

the

ensuing

consequences

on

productivity

and

income.

We

also

give

consider-ation

to

jobs

with

highest

automation

and

augmentation

potential

and

discuss

gender-speci?cdi?erences.

The

analysis

does

not

take

into

account

the

new

jobs

that

will

be

created

to

accom-pany

the

technological

advancement.

Twenty

years

ago,

there

wereno

social

media

managers,thirty

years

ago

there

were

few

web

designers,

and

no

amount

of

data

modelling

would

haverendered

apriori

predictions

concerning

avast

array

of

other

occupations

that

have

emergedin

the

past

decades.

As

demonstrated

by

Autor

et

al.

(2022),

some

60

per

cent

of

employment

in2018intheUnitedStateswasinjobsthatdidnotexist

inthe1940s.Indeed,

the

main

value

of

studies

such

as

this

one

is

not

in

the

precise

estimates,

but

rather

inunderstanding

the

possible

direction

of

change.

Such

insights

are

necessary

forproactively

de-signing

policies

that

can

support

orderly,

fair,

and

consultative

transitions,

rather

than

dealingwith

change

in

a

reactive

manner.

For

this

reason,

we

also

emphasize

the

potential

e?ects

oftechnological

change

on

working

conditions

and

job

quality

and

the

need

for

workplace

consul-tation

and

regulation

to

support

the

creation

of

quality

employment

and

to

manage

transitionsinthelabourmarket.We

hope

that

this

research

will

contribute

to

needed

policy

debates

on

digital

transformation

inthe

world

of

work.

While

the

analysis

outlines

potential

implications

for

di?erent

occupationalcategories,

the

outcomes

of

the

technological

transition

are

not

pre-determined.

It

is

humansthat

are

behind

the

decision

to

incorporate

such

technologies

and

it

is

humans

that

need

toguide

the

transition

process.

It

is

our

hope

that

this

information

can

support

the

developmentofpoliciesneededtomanagethesechangesforthebene?tofcurrent

andfuture

societies.Weintend

to

use

this

broad

global

study

as

an

opening

to

more

in-depth

analyses

at

country

level,witha

particularfocusondeveloping

countries.09ILOWorkingPaper96X

Text

Box

1:

What

are

GPTs?Generative

Pre-Trained

Transformers

belongtothefamilyofLarge

LanguageModels–

a

typeofMachineLearningmod-elbasedonneural

networks.The“generative”

partrefers

totheirabilitytoproduce

outputofa

creative

nature,

whichinlanguagemodelscantaketheformofsentences,paragraphs,

orentire

text

structures,

withcharacteristics

oftenun-distinguishablefrom

thatproduced

by

humans.“Pre-trained”

refers

totheinitialtraining

ona

large

corpusoftext

data,typicallythrough

unsupervisedorself-supervisedlearning,duringwhichthemodellearnsaboutthetext

structure

bytemporarily

maskingpartofthecontentandtryingtominimizeerrors

intheprediction

ofthemaskedwords.

Followingpre-training,

suchmodelsare

further?ne-tunedwiththeuseoflabelleddataandso-called“reinforcement

learning”,makingthemmore

suitableforspeci?ctasks.Thispartoftraining

isoftenperceived

asa

specializedjob,executed

bya

handfuloftechnicalexperts.

Inreality,

itislabourintensiveandinvolvesmanyinvisiblecontributors(Dzieza

2023).Itsprerequisite

istheproduction

ofvastamountsoflabelleddata,typicallydoneby

workersoncrowdsourcing

platforms.“Transformers”

refer

totheunderlyingmodelarchitecture,

whichusesnumerous

mechanisms,suchasattentionandself-attentionframeworks,

todevelop

weightsrelated

totheimportanceoftext

elements,suchaswords

ina

sentence,whichare

subsequentlyusedforpredictions

(Vaswani

etal.2017).WhileGPTspeci?callyrefers

tomodelsdeveloped

by

OpenAI(GPT-1,

2,3

and4),thistypeofarchitecture

isusedbymanymore

languagemodelsalready

availablecommercially.

ThelaunchofChatGPTon30November

2022madeGPTsmore

popularamongthepublic,asitmadeitpossibleforindividualswithnoprogramming

knowledge

tointeract

withGPT-3

(andeventually

GPT-4)

through

a

chatbotfunctionwitha

human-liketone.Forresearch

purposesandmore

com-plex

applications,suchlanguagemodelsare

typicallymore

powerful

whenusedthrough

anApplicationProgrammingInterface(API).AnAPIisa

developer

accesspointthatrelies

ona

query-response

protocol

withtheuseofprogrammingsoftware.

Inourcase,werely

ona

PythonscriptbasedonOpenAIlibrary,

designedtoconnecttoGPT-4

model,providea

?ne-tunedprompt

andreceive

a

response,

whichissubsequentlystored

ina

databaseonourserver.

Thisenablesbulkprocessing

oflarge

numbersofrequests

andrelies

ontheGPT-4