、

AIinhumanresourcemanagement

ThelimitsofempiricismAuthors

/

JanineBerg,Hannah

Johnston ILO

WorkingPaper154November

/2025Attribution4.0

International(CCBY4.0)Thiswork

is

licensed

underthe

Creative

CommonsAttribution4.0International.See:

https://

creativecommons.org/licenses/by/4.0/.The

user

is

allowed

to

reuse,share(copy

and

redistrib-

ute),adapt(remix,transformandbuildupontheoriginalwork)

as

detailed

inthe

licence.The

usermustclearlycredit

the

ILO

as

the

sourceof

thematerial

andindicate

if

changes

weremade

totheoriginalcontent.Useof

theemblem,nameandlogo

of

theILO

is

not

permitted

in

con-

nection

with

translations,adaptationsorotherderivative

works.Attribution–Theusermustindicateifchangesweremadeandmustcitetheworkasfollows:

Berg,J.,Johnston,

H.AI

inhuman

resource

management:The

limits

of

empiricism.

ILO

Working

Paper154.Geneva:InternationalLabourOffice,2025.?ILO.Translations

–In

case

ofatranslation

ofthiswork,the

following

disclaimer

must

be

added

alongwiththeattribution:Thisisatranslation

of

a

copyrighted

work

of

theInternational

Labour

Organization

(ILO).This

translation

has

not

been

prepared,

reviewed

or

endorsed

by

the

ILO

and

should

not

beconsidered

anofficial

ILOtranslation.The

ILOdisclaimsall

responsibility

for

itscontent

and

ac-

curacy.Responsibility

restssolelywiththeauthor(s)ofthetranslation.Adaptations

–Incaseofanadaptationofthiswork,thefollowingdisclaimermust

be

added

alongwiththeattribution:Thisisanadaptationofacopyrightedworkofthe

InternationalLabour

Organization

(ILO).This

adaptation

has

not

been

prepared,

reviewed

or

endorsed

by

the

ILO

and

should

not

beconsidered

anofficial

ILOadaptation.The

ILOdisclaimsall

responsibility

for

itscontent

and

ac-

curacy.Responsibility

restssolelywiththeauthor(s)oftheadaptation.Third-partymaterials

–

This

Creative

Commons

licence

does

not

apply

to

non-ILO

copyright

ma-

terialsincludedinthispublication.If

thematerialisattributedtoathirdparty,theuserofsuch

materialissolelyresponsible

forclearing

therights

with

therightsholderand

foranyclaimsof

infringement.Any

dispute

arising

under

this

licence

that

cannot

be

settled

amicably

shall

be

referred

toarbitra-

tionin

accordance

with

the

ArbitrationRules

of

theUnitedNations

Commission

on

International

TradeLaw

(UNCITRAL).

Thepartiesshallbeboundby

any

arbitration

awardrendered

as

aresult

of

sucharbitrationas

the

finaladjudicationof

suchadispute.Fordetailsonrightsandlicensing,contact:rights@.

Fordetailson

ILO

publications

anddigitalproducts,

visit:

www.ilo.org/publns.ISBN

9789220428610

(print),

ISBN

9789220428627

(webPDF),

ISBN

9789220428634

(epub),

ISBN

9789220428641(html).ISSN2708-3438(print),ISSN2708-3446(digital)/10.54394/NMSH7611The

designations

employed

inILO

publications,whichare

in

conformitywith

United

Nations

practice,andthe

presentationofmaterialthereindo

notimply

the

expression

ofany

opinion

whatsoeveronthepartof

theILOconcerningthelegalstatusofanycountry,areaorterritory?InternationalLabourOrganization2025orofitsauthorities,orconcerningthedelimitationofitsfrontiersorboundaries.See:www.ilo.

org/disclaimer.Theopinionsand

viewsexpressedin

thispublicationare

thoseof

theauthor(s)anddonotnec-

essarilyreflect

theopinions,

viewsorpoliciesof

theILO.Referencetonamesof

firmsandcommercialproductsandprocessesdoesnotimplytheiren-

dorsementby

theILO,andany

failure

tomentionaparticular

firm,commercialproductorpro-

cessisnotasignof

disapproval.Information

on

ILO

publications

and

digital

products

can

befound

at:/research-

and-publicationsILOWorking

PaperssummarizetheresultsofILOresearchinprogress,

and

seekto

stimulate

discussion

of

a

range

of

issues

related

to

the

world

of

work.

Comments

on

this

ILO

WorkingPaper

are

welcomeandcanbesent

to

berg@.Authorization

forpublication:CarolineFredrickson,Director,ResearchDepartmentILO

WorkingPaperscanbe

foundat:

www.ilo.org/research-and-publications/working-papersSuggested

citation:Berg,

J.,

Johnston,

H.2025.

AIinhumanresourcemanagement:

The

limits

of

empiricism,ILOWorkingPaper154(Geneva,ILO).

/10.54394/NMSH7611Therapidintegrationofartificialintelligence(AI)into

Human

Resource

Management(HRM)

is

transforminghoworganizationsrecruit,manage,andevaluatetheirworkforces.Whilepropo-

nentschampionAIasameanstoenhanceefficiency,reducebias,andalign

HR

practiceswith

strategic

business

goals,

this

paper

argues

that

such

optimism

is

misplaced.Drawing

on

a

critical

review

of

AI's

application

across

four

coreHRM

functions—recruitment,

compensation,

schedul-

ing,andperformancemanagement—thispaperidentifiessignificantrisksandlimitationsaris-

ing

from

the

fundamentalstructureof

AIsystems.Central

to

the

analysisis

a

three-parameter

framework

for

assessing

AI

tools:

theirobjective,

the

datatheyrelyupon,andhow

theyare

programmed.

Thepapershows

thatacrossHR

functions,

AIsystemsfrequentlyoperationalizereductiveorpoorlyalignedobjectives,relyonlow-quality

orbiaseddata,

and

areprogrammedinnon-transparent

ways

thatundermine

theirusefulness.

Thesestructuralshortcomingsnotonlyunderminetheeffectivenessof

AIsystemsbutalsoin-

troducelegal,ethical,andpracticalrisks

for

firmsand

their

workers.Keywords:

artificial

intelligence,

human

resource

management,

data

analytics,

algorithmic

man-

agementAbouttheauthorsJanineBergis

SeniorEconomist

andHead

of

theEffective

Labour

Institutions

Unit

in

the

Research

Department

of

the

ILO.

Since

joining

the

ILO

in

2002,

she

has

conducted

research

on

the

econom-

icandsocialeffectsoflabourlawsaswellasprovidedtechnicalassistanceonpoliciesforgen-

erating

jobsandimproving

workingconditions.

Sheis

theauthorof

severalbooksandnumer-

ousarticlesonemployment,labourmarketinstitutionsand

thedigital

transformationof

work.HannahJohnston

isanAssistant

Professor

intheSchoolofHuman

Resources

Management

at

YorkUniversityinToronto,Canada,specializingonthedigitalizationof

work.Priorto

joining

York,HannahwasapostdoctoralfellowatNortheasternUniversityinBostonandalsoworked

professionallyat

theInternationalLabourOrganization.Hannahhasalongstandinginterestin

theplatform

economy

andis

a

collaborator

with

OxfordUniversity’sFairworkProject.Her

recent

publications

canbe

foundin

journalsincluding

Industrial

andLaborRelationsReview,

Work

and

Occupations,and

theInternationalLabourReview.01

ILOWorkingPaper

154AbstractAbstract01Abouttheauthors01、Introduction04、

1Howdidwegethere?Theriseof

“peopleanalytics”andmanagingwork-

ersthroughdata05、

2TheworkingsofAIsystems:Objective,dataandprogramming

RecruitmentSourcingScreening,InterviewingandSelectionCompensationSchedulingworkPerformancemanagement08091011131619、

3AIinHRM:

Unbridled

optimism23、

4WhatisanHR

managertodowithAI?26、Conclusion28References

29Acknowledgements

36Tableofcontents02

ILOWorkingPaper

154Figure1.McKinsey

infographic

on

the

benefits

of

people

analytics

06Figure2.The

workings

of

AI

systems

09ListofFigures03

ILOWorkingPaper154Motivated

byadesireto

moreefficientlyand

effectively

manage

people

in

organizations,

HR

managersareusing

theprogrammingandanalytical

capacities

of

AI

to

fulfillkeyHR

functions

–

includingselectionandrecruitmentofpersonnel,compensationdeterminationandstructure,

performancereviewandevaluation,and

theorganizationof

working

time.Largely

absent

from

therush

tointegrate

AI,however,hasbeen

a

comprehensive

assessment

of

whether,

and

under

what

circumstances,

AI

is

useful

for

themanagement

of

people

within

organizations.

Yet

despite

thelackofassessment,the

‘AIforHR’

industryandtheadoptionof

thesetoolsandsystemsby

individual

firmsandorganizationsisburgeoning.Thispaperpresentsaframeworkforunderstandingandevaluatingthepotentialbenefitsand

possiblerisksorharmspresentedby

AI

systemsin

workforcemanagement.Following

a

section

of

the

paper

documenting

the

historical

context

of

theHumanResources

field

that

has

given

rise,

first

topeople

analytics

and

then

to

AI,

thepaperpresents

a

frameworkbasedon

threeinter-re-

lated

parameters

that

can

help

assess

the

quality,

legality,

and

suitability

of

AI

systems

used

in

the

field.

These

are:

(1)

the

system

objective,

(2)

the

dataitisbuilt

on

andrelies

on,

and

(3)

howtheAI

systemisprogrammed.Drawing

on

existingliterature

abouthow

AIisbeingused

for

workforce

management,

thepaperapplies

the

three-parameter

framework

tomap

thecontoursof

AIuse

relative

to

fourkeyHumanResourceManagement

functions

where

adoption

of

AI

technologies

hasbeenprominent:recruitment,compensation,scheduling,andperformancemanagement.

Ourdiscussionsection,

“Unbridledoptimism”

views

thedisciplinaryandoccupationalhistoryof

HR

alongside

the

findingsonHR’semergentuseof

AI.

We

argue

that

thesearch

for

occupation-

allegitimacyby

HRprofessionalshasfostered

a

preoccupationwith

numeracy

and

positivism

that

has

providedfertilegroundforthe

‘evidence-basedsolutions’

thatAIsystems

purportto

offer.Thistendencytowardspositivism,wecontend,islikelytoresult

inwidespread

adoption

of

AIundercircumstancesthatcreaterisksforworkers,liabilitiesforfirms,andcostsforsocie-

ty.Giventheseeminglyinevitabletransformationof

workduetoAI,wearguefortheneedfor

HRprofessionals

toimprove

theirunderstandingof

the

workingsofAIsystemsso

that

they

can

better

judge

theirpotentialandlimitations,and

that

thisisbestachieved

when

theyparticipate

in

thedesignof

systems

thatareimplementedin

their

workplaces.、

Introduction04

ILOWorkingPaper154HumanResourceManagement

emerged

in

the1950s

as

a

distinct

field

of

study

and

practice

con-

cerned

with

themanagement

of

people

in

organizations.

Carved

out

from

thebroader

discipline

of

IndustrialRelations,

whichhashistoricallyexaminedlabourrelationsin

thecontextof

unions

andcollectivebargainingorcollectiveemploymentrelationsHumanResourceManagement

ismostassociatedwithanemployer

srelationswithindividualemployees(Kaufman2001).As

practitioners,HRmanagers

are

typically

engagedin

facilitating

arange

of

personnel

functions

includingrecruitment

and

selection,

compensation,

scheduling

andperformancemanagement

andpromotiontoachieve

theorganizationsgoals.Since

the

1980s,

the

dominant

paradigm

guidingHR

managers

has

beenstrategicHRM

(Paauwe

andBoon

2018).Building

onHR’s

originsand

the

foundational

concept

of

scientific

management,

strategic

HRMaimsto

link

firm

performance

to

the

specific

methods

and

practices

deployed

within

the

firm

to

manage

their

workers.

When

first

introduced,

this

approach

represented

a

significant

shift

infirmmanagement.Prior

to

this

shift,

organizational‘strategy’

referred

to

a

firm’sperspec-

tiveor

‘worldview’,itsintendedplanandpatternedbehaviour,as

wellasa

firm’suse

of

ploys

to

outwitcompetitors(Mintzberg1987).However,in

the1980s

firmstrategybegan

to

focusmore

oncausality,withresearchers

andpractitioners

attempting

toidentifyinputs

(e.g.management

practices)

and

outputs

(e.g.marketperformance),

quantify

them,

and

derive

causal

connections

betweenthetwo.Operationalizingthisstrategicapproach,as

many

have

argued,

has

largely

been“gearedtowardsspecificnumericaltargets”

(WoodandKispál-Vitai2017).Firmsseeknot

only

to

derive

general

conclusions

or

trends

through

quantification,but

also

to

compare

arange

of

strategies

for

thepurposeof

distinguishingasingular

bestpractice

.This

epistemological

shift

has

fueled

the

development

of

new

metricsand

datapoints

that

can

be

used

to

analyze

the

relationship

between

workforce

management

and

firm

performance.

In

turn,

HRmanagerscanusethesenewsourcesofdataabouttheworkforcetoinformdecision-mak-

ing.ThistypeofEvidence-Based

Managementtechnique(EBM)hasbeenlargely

promoted

in

strategicHRMtohelpovercomepitfalls,suchasrelyingonpersonalexperienceormanagerial

whims,orthetendencytomimicthestrategiesorapproachesoftopperformers(Pfefferand

Sutton2006;Rousseau2006);Reay,Berta,andKohn2009).Theresulthas

beenavastarsenal

ofdigitallyenabledworkplaceandwork-relatedtoolsthatcapture,collectandanalyzeworker

behavior

andperformancedata.

Since

the1980s,

thishasdriven

thegrowing

field

of

people

an-

alytics–definedas

“theuseofmeasurementandanalysistechniquestounderstandandopti-

mizethepeoplesideofbusiness(EnderesandShannon2019)andultimatelyhas

pavedthe

path

towards

theadoptionof

AI

forHRM.Withinthefieldofpeopleanalyticsandthedevelopmentof

AI,moredataandbetterdataare

commonly

viewed

as

precursors

for

robust

and

powerful

systems.People

analytics

data

may

per-

taintoworkers’

demographics,descriptiveinformationaboutthelocationornatureof

the

job,

performance,trainingorprofessionalhistory,ortenure.Whileeveryorganizationhasdataon

its

workforce,advancedanalysis

for

workforceoptimizationandplanningcanonlybeachieved

when

data

are

high

quality,

robust

and

plentiful,

and

when

organizations

have

adequately

trained

staff

toprocess

andmake

sense

of

them.Figure1is

aninfographic

from

aMcKinseypublication

on

the

virtue

of

people

analytics,

stressing

that

the

power

of

such

systems

to

provide

workforce

insightsincreases

with

the

volumeandqualityof

workforcedata(Ledetetal.2020).、

1Howdidwegethere?Theriseof

“peopleanalytics”

andmanagingworkersthroughdata05

ILOWorkingPaper154Source:(Ledetetal.

2020).Whenfirst

introduced,

HR

professionals

used

peopleanalyticstoassess

patternsabouttheir

workforce,informing

a

widerange

of

HR

functionsincludingrecruitment

and

hiring,

promotion,

compensation,andhealthandsafety(Giermindletal.2022).

Withmore

andhigher-quality

“big

data”and

increases

incomputing

power,

peopleanalytics

is

being

propelledfromcorrelation

analysisintotheworldofpattern-based

prediction,and

Human

Resource

professionals,once

responsible

forexecutinga

widerangeof

functions,are,insomeinstances,relinquishing

these

responsibilitiestoalgorithmsandAI.Thisdecades-longstrategicshiftprovidescriticalcontext

forunderstandingwhythefieldofHRMhasembracedtheuseof

AI;theparticularharmsthat

mayemerge

from

theuseof

AI;and

why

the

fieldofHRislargelyblind

to

them.AI

is

distinguished

by

vast

quantities

of

data

and

rapid

quantitative

analysis.

These

new

sources

of

data

have

held

particular

appeal

toa

profession

that

has

oriented

itself

towardsEBM.

Additionally,

AI

is

popularly

portrayed

as

innovative

and

cutting

edge,

and

the

use

of

AI

technologies

and

tools

isthusseenasawaytoelevatetheHRprofession.Thismotivationisalsounderpinned

by

im-

portantcontextashistoricallymanyhavetendedtoregardHRprofessionalsas

administrative

functionaries,

who–lackingpower–areengagedinmerebureaucratic

service

delivery

without

addinganyreal

value

to

theorganization(Wright2008;Legge1978).Although

thestrategic

focusofHRMin

the1980s(along

withother

activitiessuch

as

the

forma-

tion

of

professional

associations

and

educational

and

training

courses)provided

one

avenue

for

theoccupation

torecastitself

onequal

footing

withothermanagerialprofessions(Legge1978;

Cayrat

andBoxall

2023),criticsmaintain

that

the

fieldhas

stillnotprovidedevidence

of

its

worth

(Wright2008;

Alvesson2008;Kryscynskietal.2018;CayratandBoxall2023;Hammonds2005).

Thus,thispersistentquestforoccupationallegitimacypresentsanotherimportantmotivation

forHR’sembraceof

people

analytics

and

AI:byembracingevidence-basedmanagement

(EBM),、

Figure1.

McKinseyinfographiconthebenefitsofpeopleanalytics06

ILOWorkingPaper154andthedataaccumulatinganddata-intensivetoolsandtechnologies

that

facilitateit,HRprac-

titionersmightprovideevidenceof

their

value

to

the

firm.Thisapproach,however,hasnotbeen

withoutcriticism.Scholarsargue

that

whileempiricalre-

searchlinkingHRpractices

toorganizationalperformancecanat

timesdemonstrateaclearas-

sociation,therelationshipbetweenthese

variablesisunder-theorized(FleetwoodandHesketh

2008).Theempiricalapproachof

HR

scholars

has

proven

adequatefor

generating

predictive

dimensionsoftheory

rooted

in

priorobservationsof

‘what’is

happeningand

‘how’;

however,

robusttheories

shouldalso

be

capable

ofexplainingthe

mechanismsand

reasoning

behind

themechanismsandcausalrelationships(Guest2025;FleetwoodandHesketh2008).Itiswith

respecttothislatterexplanatorydimensionthatHRMtheoryhasfallenshort.Withoutanade-

quateorclearlyarticulatedtheory,researchwill“alsolackanadequaterationaleforthechoice

ofphenomenathatwilleventuallybecomethevariables”usedtodrivefuturepredictivetheo-

rization(ibid);127).Undertheseconditions,ratherthanexplainingwhyorganizationssucceed

or

failand

whatothersystemicconditionsordiscretepracticesmaycontribute

to

thisoutcome

(Guest2025),HRmanagersandresearchersalike

areprone

toreplicatepast

practice

simply

be-

causesuchpracticeshavebeenpreviouslyexamined.HRtheories,inturn,remainpoorlyartic-

ulatedandlackconceptualclarity(Boonetal.2019).Thisriskof

‘measurementwithouttheory’

isamplifiedinthecontextof

AI,

whichbydefinitionisnotaboutbuildinganexplicittheory,but

ratheraboutbuildingpatterns

throughbigdata(ElragalandKlischewski2017).Proponentsof

AIintegration,includingmanyglobalconsulting

firms,promiseexactly

this:

add-

edvaluefromtheuseoftheirnewlydevelopedtechnologiesandtools–and

specifically,that

thepredictivecapacityof

AI

willhelp

toreveal

therecipe

foraneffective

allocation

of

humanre-

sources.

Companies

arerapidly

obliging.

ISG,

an

organizational

changemanagement

company,

foundin2023

thatoneoutof

every

threeorganizations

wasprioritizing

AI

and

analyticsin

their

HRandtechnologystrategy.Thisfindingwasbasedonasurveyofenterpriseleadersatfirms

employingbetween5,000and50,000workers.Asthereportexplained,

“l(fā)eadingHRtechnolo-

gyproviders,suchassuchasOracle,

[IBM],SAP

SuccessFactors

andWorkday,

arefocused

on

embeddingAI,machinelearningandanalyticsinthecoreplatform”(ISG2023).Indeed,

most

of

thesoftwareisprovidedthroughthird-partiesandistypicallyanadd-ontootherservicesal-

readyprovided.ThisfacilitatestheworkofHRprofessionalswhooftenlackadequatetraining

in

whatkindsof

analyticalquestions

toaskorhow

tointerpretquantitative

findings(Giermindl

etal.2022;Kryscynskietal.2018).Yetdespitehighratesof

technologicalintegration,including

AI,

thesame

ISG

survey

found

that

less

than

half

of

organizations

surveyed

realized

business

value

from

their

investments.

What

has

emergedis

thus

aparadox:

although

trovesof

data

anddevelopmentsin

the

fieldof

AIpromise

toprovideHRprofessionalswiththenecessarydata

and

analysisto

helpthemeffectively

and

efficientlyallocatehumanresources,thesetechnologieshavenotyetdeliveredmuchvalueto

firmsandorganizations.Whynot?Whatdoestheactualevidenceontheimplementationof

AI

inHRMbear?Does

thedesignof

these

AIsystemsallow

them

todeliveraspromised?07

ILOWorkingPaper

154本報(bào)告來(lái)源于三個(gè)皮匠報(bào)告站（）,由用戶(hù)Id:349461下載,文檔Id:988493,下載日期:2025-12-13Toassesshoweffective

AIhasbeenforrealizingtheoverarchinggoalsofHR,itisnecessaryto

delineatewhatthesegoalsare,andtounpacktheworkingsofAIsystem

tobetterunderstand

potentialpitfalls.

AIsystemsarecomposed

of

threeinter-relatedparameters:

(1)

the

system

ob-

jective,(2)thedataitisbuiltonandrelieson,and(3)howit

is

programmed(Seefigure2).The

quality

of

each

of

theseparameters

differentiates

systems

that

work

well

from

those

that

donot.Beginningwiththesystemobjective,whiledefiningsuchanaimmayappearstraightforward,

how

theseaimsareoperationalizedismorecomplicated.

When

AIsystemsare

developed

with

neutralaims–suchasdeterminingtheshortestroutebetweenlocations–

it

is

easyto

relyon

identifiableandrelevantvariablesandthefindingsareeasytointerpret.

But

most

human

re-

source

functionsinvolve

the

“fleshy,messy,indeterminatestuff

of

everydaylife”–stuff

thatare

difficult

to

capture

in

a

discrete

variable

(Katz

2001).

This

can,

as

described

by

Sandy

Gould

(Gould

2024),

create

practical

challenges,

which

she

explains

asthe

need

to,

“accept

resource

constraints

onwhatismeasured,aswellastheontologicallimitations(somethingsarenevergoingtobe

open

to

‘direct’

measurementbyanyconceivablemeans)”

(106).Data,meanwhile,

are

anecessaryinput

to

AI

systems.Limitationsin

datahave

been

the

focus

of

muchcritique,

with

twoparticularlysalientconcerns.First,

thereis

thequestionof

dataquality.

AIsystemsrelyon

trainingdata

to‘learn’

theconnections

andpatterns

thatprovide

the

founda-

tionuponwhichdecisionsaremade(Whangetal.2023).Whenthesedataarepoorquality,AI

systems

yieldpoor

quality

outputs.

Asthe

saying

goes:‘garbagein,

garbage

out’.

A

secondissue

concerns

the

suitability

of

the

data.

Inbespoke

AI

systems

that

are

developedinternallyby

or

for

asingular

firm,

trainingdatamaybeinternal

to

the

firmandconsistofpastoperations(Kresge

2020).

When

systems

are

developed

for‘off

the

shelf’

use,

they

rely

on

more

generalized

datasets

that

can

either

be

purchased

via

a

growing

data

market

(Zuboff

2019)

or

otherwise

compiled

from

availabledata

thataredeemed

toberelevantsources

for

thesystem

(Muldoonet

al.

2023).

The

appropriateness

of

more

generalized

datasets

ought

to

be

evaluated

on

a

case-by-case

basis

but

rarelyis.Equally,whenissuesstemmingfromnon-representativedataarise,acommonlysug-

gestedsolutionisthatmoredataormorerepresentativedataareneeded.However,thesetypes

ofproblemsalsoraisequestionsabout

thesuitabilityof

thedata

tomeet

thestatedobjective.Thosecautiousof

AIhavealsopointedtothepotentialproblemsassociatedwiththethirdpa-

rameter:programming.

Algorithms

are

akeydecision-making

feature

atthecoreof

AI

systems.

They

can

be

defined,

intheir

most

basicsense,asa

set

of

rules

executedthrough

computer

programmingcodewithaparticularaimorobjective.

Algorithmsfunctionwith

variedlevelsof

autonomyandhumaninvolvement;thesecharacteristicsarealso

determined

bytheir