Özkizil an, Didem; Landini, Fabio
A icle — Published Ve sion
T us wo hy and human-cen ic? The new go e nance
o wo kplace AI echnologies unde he EU’s A i icial
In elligence Ac
T ans e : Eu opean Re iew o Labou and Resea ch
P o ided in Coope a ion wi h:
WZB Be lin Social Science Cen e
Sugges ed Ci a ion: Özkizil an, Didem; Landini, Fabio (2025) : T us wo hy and human-cen ic? The
new go e nance o wo kplace AI echnologies unde he EU’s A i icial In elligence Ac , T ans e :
Eu opean Re iew o Labou and Resea ch, ISSN 1996-7284, SAGE, London, Vol. Online i s , pp. 1-15,
h ps://doi.o g/10.1177/10242589251336193
This Ve sion is a ailable a :
h ps://hdl.handle.ne /10419/320729
S anda d-Nu zungsbedingungen:
Die Dokumen e au EconS o dü en zu eigenen wissenscha lichen
Zwecken und zum P i a geb auch gespeiche und kopie we den.
Sie dü en die Dokumen e nich ü ö en liche ode komme zielle
Zwecke e iel äl igen, ö en lich auss ellen, ö en lich zugänglich
machen, e eiben ode ande wei ig nu zen.
So e n die Ve asse die Dokumen e un e Open-Con en -Lizenzen
(insbesonde e CC-Lizenzen) zu Ve ügung ges ell haben soll en,
gel en abweichend on diesen Nu zungsbedingungen die in de do
genann en Lizenz gewäh en Nu zungs ech e.
Te ms o use:
Documen s in EconS o may be sa ed and copied o you pe sonal
and schola ly pu poses.
You a e no o copy documen s o public o comme cial pu poses, o
exhibi he documen s publicly, o make hem publicly a ailable on he
in e ne , o o dis ibu e o o he wise use he documen s in public.
I he documen s ha e been made a ailable unde an Open Con en
Licence (especially C ea i e Commons Licences), you may exe cise
u he usage igh s as speci ied in he indica ed licence.
h ps://c ea i ecommons.o g/licenses/by/4.0/
h ps://doi.o g/10.1177/10242589251336193
T ans e
1 –15
© The Au ho (s) 2025
A icle euse guidelines:
sagepub.com/jou nals-pe missions
DOI: 10.1177/10242589251336193
jou nals.sagepub.com/home/ s
T us wo hy and human-
cen ic? The new go e nance
o wo kplace AI echnologies
unde he EU’s A i icial
In elligence Ac
Didem Özkizil an
Wissenscha szen um Be lin ü Sozial o schung (WZB), Ge many
Fabio Landini
Uni e si y o Pa ma, I aly
Summa y
The inc easing use o AI echnologies in he wo kplace aises signi ican conce ns abou hei
implica ions o wo ke s’ igh s and he u u e o wo k. The EU’s A i icial In elligence Ac aims
o ensu e ha inno a ion is no a he expense o undamen al igh s. While he Ac seeks o
p omo e human-cen ic and us wo hy AI applica ions, his a icle highligh s se e al egula o y
gaps ha a e likely o os e wo key ends shaping he u u e o wo k in Eu ope: he inc easing
au onomy o AI ech companies o e he design o wo kplace AI and he ein o cemen o
powe imbalances in he wo kplace. These ends lea e EU wo kplaces ulne able o biased and
in usi e AI, he eby hinde ing he ai in eg a ion o AI in o he wo k o he u u e. This a icle
c i iques he Ac ’s sho comings in egula ing AI a wo k and ad oca es a dedica ed EU di ec i e
o mi iga e ad e se impac s o AI in he wo kplace and ein o ce human-cen ic AI p inciples o
socie al bene i .
Résumé
Le ecou s oujou s plus impo an aux echnologies de l'IA su le lieu de a ail susci e de
i es inquié udes quan à leu s implica ions pou les d oi s des a ailleu s e l'a eni du a ail.
Le èglemen de l'UE su l'in elligence a i icielle ise à ga an i que l'inno a ion n’es pas mise
en œu e au dé imen des d oi s ondamen aux. Même si le èglemen en end p omou oi
des applica ions d'IA cen ées su l'humain e dignes de con iance, ce a icle me en é idence
l'exis ence de plusieu s lacunes églemen ai es suscep ibles de a o ise deux endances clés pou
l'a eni du a ail en Eu ope : l'au onomie c oissan e des en ep ises de echnologie de l'IA dans
Co esponding au ho :
Didem Özkizil an, Wissenscha szen um Be lin ü Sozial o schung (WZB), Reichpie schu e 50, Be lin, 10785,
Ge many.
Email: [email p o ec ed]
1336193TRS0010.1177/10242589251336193T ans e Özkizil an and Landini
esea ch-a icle2025
Main A icle
2 T ans e 00(0)
la concep ion de l'IA su le lieu de a ail e l'accen ua ion des déséquilib es de pou oi su ce
lieu de a ail. Ces endances exposen les lieux de a ail dans l'UE au isque d'une IA biaisée e
in usi e, empêchan une in ég a ion équi able de l'IA dans le a ail de demain. L'a icle souligne
les lacunes du èglemen su l'IA au a ail e p éconise l'adop ion d'une di ec i e eu opéenne
spéci ique isan à a énue les e e s néga i s de l'IA su le lieu de a ail e à en o ce , dans
l'in é ê de la collec i i é, les p incipes d'une IA cen ée su l'ê e humain.
Zusammen assung
Wi s ellen die These au , dass Bildungsungleichhei en, die sich au das Ausmaß de poli ischen
Mi wi kung on Bü ge innen und Bü ge n auswi ken, eben alls Ein luss au die Pa izipa ion
on Beschä ig en am A bei spla z haben – Menschen mi einem höhe en Bildungsabschluss
zeigen sowohl in de Poli ik als auch im Be ieb ein s ä ke es Engagemen . Wi sind eben alls
de Au assung, dass die A bei nehme pa izipa ion das im poli ischen Engagemen e kennba e
Bildungsge älle wei e e schä . Wi un e suchen die Auswi kungen de Pa izipa ion am
A bei spla z, poli ische Diskussionen und de Aneignung on Sozialkapi al am A bei spla z.
Anhand de Da en on 3.037 niede ländischen A bei nehme innen und A bei nehme n zeigen wi ,
dass Menschen mi einem höhe en Bildungsabschluss einen besse en Zugang zu Möglichkei en de
poli ischen Sozialisie ung am A bei spla z haben als Menschen aus bildungs e ne en Milieus. Die
poli ische Sozialisa ion am A bei spla z s eh wiede um in einem posi i en Zusammenhang mi de
poli ischen Be eiligung. Diese E gebnisse deu en da au hin, dass poli ische Ungleichhei en, de en
U sache in Bildungsun e schieden lieg , du ch E ah ungen am A bei spla z e s ä k we den.
Keywo ds
A i icial in elligence, EU AI Ac , he wo ld o wo k, wo ke s’ igh s, Cha e o Fundamen al
Righ s
In oduc ion
The Eu opean Union’s A i icial In elligence Ac (EU, 2024) de ines an ‘a i icial in elligence’
(AI) sys em as:
a machine-based sys em ha is designed o ope a e wi h a ying le els o au onomy and ha may exhibi
adap i eness a e deploymen , and ha , o explici o implici objec i es, in e s, om he inpu i ecei es,
how o gene a e ou pu s such as p edic ions, con en , ecommenda ions, o decisions ha can in luence
physical o i ual en i onmen s.
This comp ehensi e de ini ion unde sco es AI sys ems’ e sa ili y and ans o ma i e po en ial.
They ha e apidly e ol ed in ecen yea s and ha e ound applica ions ac oss di e se domains. In
pa icula , ‘wo kplace a i icial in elligence echnologies’ (WAITs) exempli y his e olu ion, pe -
o ming an inc easing a ay o bo h ou ine and non- ou ine men al and physical asks. This o e s
companies he p ospec o combining he unique s eng hs o humans and machines o comple e
asks mo e e icien ly, accu a ely and apidly (Özkizil an and Hassel, 2021). Al hough he numbe
o o ganisa ions deploying WAITs emains con es ed, a clea end is e iden : Eu opean companies
a e inc easingly emb acing AI echnologies (EU-OSHA, 2022a). Fu he mo e, some schola s p e-
dic ha his end will con inue in he coming yea s (A n z e al., 2016; B ynjol son and McA ee,
2016; F ey and Osbo ne, 2017; Nedelkoska and Quin ini, 2018).
Özkizil an and Landini 3
The g ow h in WAITs deploymen has ins iga ed ib an schola ly discussions ega ding hei
u u e impac s on wo k and employmen ela ions. These deba es ocus p ima ily on he po en ial
o la ge-scale eplacemen o human labou wi h WAITs (A n z e al., 2016; F ey and Osbo ne,
2017; Nedelkoska and Quin ini, 2018) and he changing na u e o wo k (Mu o e al., 2019; Se oz,
2019). A key conclusion om hese discussions is ha WAITs a e likely o exace ba e exis ing
powe asymme ies (Kellogg e al., 2020; Özkizil an and Hassel, 2021) and agg a a e un ai and
exploi a i e p ac ices a wo k (De S e ano, 2019; Ma eescu and Nguyen, 2019; Özkizil an and
Hassel, 2021).
Moun ing conce ns abou he socio-economic implica ions o he inc easing use o WAITs ha e
con ibu ed o a heigh ened awa eness a he EU le el (EESC, 2017), leading o poli ical ini ia i es,
including he Eu opean AI S a egy (Eu opean Commission, 2018), E hics Guidelines o
T us wo hy AI (AI HLEG, 2019), he 2030 Digi al Compass (Eu opean Commission, 2021), he
In e na ional Ou each o Human-Cen ic A i icial In elligence Ini ia i e (Eu opean Commission,
2022), and he Eu opean Decla a ion on Digi al Righ s and P inciples (EU, 2023). Though di e se
in scope and ocus, hese ini ia i es aim o c ea e human-cen ic AI ha se es people, p o ec s
undamen al igh s and empowe s businesses. This EU-le el human-cen ic app oach o AI also
ex ends o he wo ld o wo k. I seeks o es ic he g ow h o manage ial disc e ion enabled by
AI-led decision-making ools (EU, 2023), as well as o p epa e labou ma ke s o an AI-led u u e
wo ld o wo k (EU, 2023; Eu opean Commission, 2021).
The human-cen ic app oach o AI has also in o med he policy-making p ocess o he EU’s
A i icial In elligence Ac ( he ‘Ac ’). The Ac en e ed in o o ce in Augus 2024 and i is se o be
implemen ed g adually o e ime (Fu u e o Li e Ins i u e, 2024). The in en behind his pionee ing
legisla ion is o s i e o a delica e balance: os e ing echnological inno a ion and compe i i e-
ness wi hin he EU, while sa egua ding i s ci izens’ sa e y and undamen al igh s (Council o he
EU, 2024; Madinie , 2021). The Ac ’s co e p inciple, decla ed in Reci al 6, unequi ocally empha-
sises i s aspi a ion o uphold human-cen ic AI sys ems: ‘As a p e equisi e, AI should be a human
cen ic echnology. I should se e as a ool o people, wi h he ul ima e aim o inc easing human
well-being.’ This guiding p inciple ansla es in o he Ac ’s pu pose: ‘ o p omo e he up ake o
human cen ic and us wo hy a i icial in elligence (AI) while ensu ing a high le el o p o ec ion
o heal h, sa e y, undamen al igh s as ensh ined in he Cha e o Fundamen al Righ s o he
Eu opean Union ( he ‘Cha e ’)’ (Reci al 1).
The Ac is laudable o i s commi men o ad ancing he adop ion o human-cen ic and us -
wo hy AI o imp o e human well-being and i s adhe ence o he p inciples ensh ined in he
Cha e . Howe e , his a icle con ends ha se e al sho comings signi ican ly unde mine he Ac ’s
e ec i eness in p o ec ing wo ke s’ igh s and eedoms, as se ou in he Cha e . These include
he ollowing. The Ac , legally based on A icle 114 o he T ea y on he Func ioning o he
Eu opean Union (TFEU), con ines he p o ec ion o undamen al igh s o a p o ision g an ing he
EU gene al au ho i y o in e nal ma ke ha monisa ion. I elies on p o ide s o high- isk AI sys-
ems o sel -assess and sel -ce i y hei p oduc s’ compliance. I exemp s p i a e companies
deploying high- isk WAITs om conduc ing Fundamen al Righ s Impac Assessmen s and om
he manda o y egis a ion equi emen o hese sys ems. I unde mines he igh o wo ke s o an
explana ion o decisions made by high- isk AI sys ems ha impac hei heal h, sa e y o unda-
men al igh s. I also lea es gene al-pu pose AI unde - egula ed, ca es ou exemp ions o banned
emo ion- ecogni ion sys ems in wo kplaces, and neglec s he b oade socio-economic impac s o
WAITs by ocusing exclusi ely on algo i hmic managemen ools.
This a icle a gues ha hese egula o y gaps will p obably gi e ise o wo key ends shaping
he u u e o wo k in Eu ope: he inc easing au onomy o AI ech companies wi h ega d o he
design o WAITs and he solidi ica ion o powe imbalances in he wo kplace. These ends could
4 T ans e 00(0)
ende EU wo kplaces ulne able o biased and in usi e AI, he eby hinde ing AI’s ai and demo-
c a ic in eg a ion in o he wo k o he u u e. Consequen ly, ega ding issues ela ed o he wo ld o
wo k, he Ac ails o ul il i s p omise o p omo ing he up ake o human-cen ic and us wo hy
AI o enhance human well-being. To p e en such an un a ou able ou come, as Eu opean ade
unions ha e igh ly ad oca ed, an EU di ec i e is needed dedica ed o add essing he unique chal-
lenges posed by WAITs and ec i ying he sho comings o he EU AI Ac (ETUC, 2023; Indus iAll,
2024). Only h ough such a ge ed legisla ion can wo ke s be adequa ely p o ec ed om he po en-
ial ha ms o WAITs, ensu ing ha AI echnology is human-cen ic and us wo hy, and bene i s
EU socie y as a whole.
The a icle is s uc u ed as ollows. I begins wi h a b ie o e iew o he deploymen o WAITs,
ollowed by an analysis o he key ac o s con ibu ing o WAITs ou comes ha a e de imen al o
wo ke s’ in e es s, and con inues wi h an examina ion o he possible impac s o WAITs deploy-
men on wo ke s’ igh s as gua an eed by he Cha e . Nex , i sc u inises he EU AI Ac ’s sho -
comings ela ed o he wo ld o wo k. The a icle concludes wi h a discussion o hese sho comings,
emphasising hei po en ial impac on he u u e wo ld o wo k and ad oca ing o an EU di ec i e
ha mi iga es AI’s ad e se impac s and s eng hens human-cen ic and us wo hy AI p inciples in
he wo kplace.
Deploymen o AI in he wo kplace
A hei cu en le el o de elopmen , WAITs ouch almos e e y ace o he wo ld o wo k, om
ask au oma ion o wo kplace sa e y, om hi ing, i ing and p omo ion decisions o p oduc i i y
moni o ing (Cowo ke .o g, 2021; De S e ano and Wou e s, 2022; Neg ón, 2021b; Özkizil an and
Hassel, 2021). Despi e ongoing deba es on whe he AI will make human labou obsole e (Özkizil an
and Hassel, 2021), he a ailable esea ch e eals ha WAITs adop ion so a has no eplaced
wo ke s in such a way as o inc ease echnology-induced job losses (Acemoglu e al., 2020; Fel en
e al., 2019). Howe e , esea che s p edic an inc ease in he isk o labou displacemen as a esul
o AI-led au oma ion o e he coming wo o h ee decades, a ec ing wo ke s ac oss all skill le els
(Mu o e al., 2019; OECD, 2019; Webb, 2020).
No wi hs anding he possibili y o WAITs se e ely impac ing he ypes o jobs a ailable in he
u u e, many expe s p edic a u u e o wo k in which WAITs design will ocus on new possibili-
ies o human-machine collabo a ion in such a way as o make he bes o he a ailable capabili ies
(B ynjol sson e al., 2018; Eu o ound, 2024; Manyika, 2018). Acco dingly, some WAITs a e likely
o deskill wo ke s, pa icula ly in cases in which machines a e alloca ed o ca y ou in ica e asks
ela ed o in o ma ion p ocessing and decision-making, while wo ke s a e demo ed o ou ine asks
and/o assigned o ollow he ins uc ions o he machines (EPRS, 2021; Ja ahi, 2019; Moo e,
2019). Some WAITs a e also deployed o complemen and augmen human labou , pa icula ly
when hey a e designed o elie e humans om pe o ming ou ine and/o haza dous asks. This
imp o es human capabili ies by allowing wo ke s o concen a e on mo e complex, sa is ying o
c ea i e asks (OECD, 2019; Se oz, 2019).
AI-empowe ed solu ions o wo k o ce managemen , o en e e ed o as ‘algo i hmic manage-
men ools’, also belong o WAITs. These applica ions a e designed o comple ely o pa ially ake
o e manage s’ decision-making oles in issues ela ed o wo k o ce managemen . Algo i hmic
managemen ools a e epo ed o yield opposing ou comes on wo k and employmen ela ion-
ships. Rega ding hei adop ion o bene i bo h wo ke s and employe s, o ins ance, some con-
s uc ion and mining companies use algo i hmic managemen ools o eal- ime wo ke sa e y,
while in he ho el business, panic bu ons a e used o assis wo ke s wo king in isola ion o call o
help in eme gency and/o dange ous si ua ions (Be nha d e al., 2021). When i comes o hei
Özkizil an and Landini 5
deploymen in ways ailo ed o a ou he in e es s o employe s, he schola ly li e a u e is eple e
wi h examples including, bu no limi ed o, employe s’ pe asi e su eillance o wo ke s’ ac i i-
ies, AI-led biased hi ing and i ing decisions, and un ealis ic p oduc i i y goals (De S e ano and
Taes, 2022; EU-OSHA, 2022b; Özkizil an and Hassel, 2021). Some key ac o s con ibu ing o he
nega i e impac o WAITs deploymen on wo ke s’ in e es s a e discussed below.
AI as a h ea o wo ke s’ in e es s
In ecen yea s epo s and announcemen s ha e inc eased conce ning he de elopmen o WAITs
wi h ad anced capabili ies, pu po edly app oaching o e en su passing human abili ies (Neg ón,
2021b; Özkizil an and Hassel, 2021). As a luc a i e p oduc line o AI de elope s, howe e ,
WAITs o en p io i ise co po a e objec i es, such as boos ing p oduc i i y and cu ing cos s
(Neg ón, 2021b; Özkizil an and Hassel, 2021; Rani e al., 2024). I is commonly decla ed in schol-
a ly discussions ha many WAITs p oduce ou comes de imen al o wo ke s’ in e es s (De S e ano
and Taes, 2022; Özkizil an and Hassel, 2021; Se oz, 2019). An examina ion o he ele an li e a-
u e iden i ies h ee unde lying ac o s o pa icula ele ance o his endency.
Fi s , one signi ican ac o is ha o he da a ed in o AI sys ems, which a e conside ed o be
a ely neu al o lawless (EPSC, 2018; Moo e, 2019; Williams e al., 2018). Such da a a e o en
desc ibed as ‘easy o manipula e, may be biased, may e lec cul u al, gende and o he p ejudices
and p e e ences and may con ain e o s’ (EESC, 2017: 6). I has been a gued ha algo i hms ha
ely on biased da a a e highly p one o p oducing hei ou comes based on he pa e ns and p e e -
ences ing ained in ha da a, which could po en ially ad ance he in e es s o employe s (De
S e ano, 2019; EPSC, 2018; Özkizil an and Hassel, 2021). As a esul , algo i hmic decisions could
be used o jus i y disc imina o y p ac ices, a ionalise unequal ea men o pe pe ua e exploi a i e
wo kplace dynamics, ul ima ely unde mining wo ke s’ igh s and in e es s (Özkizil an and Hassel,
2021; Williams e al., 2018).
The second ac o conce ns AI ech wo ke s, namely hose in ol ed in c ea ing, de eloping and
main aining AI-enabled echnology p oduc s and se ices. The numbe o AI ech wo ke s is
epo ed o be e y small, e lec ing he wo ldwide sca ci y o AI alen (McKinsey, 2022; Yuan
e al., 2020). Mos o hese wo ke s lack di e si y in e ms o ace, geog aphy, social class and
gende (De S e ano, 2019; EPSC, 2018; Linkedin, 2019; Moss and Me cal , 2020). While his lack
o di e si y is a global issue wi hin he AI alen pool, i is pa icula ly conce ning o AI p oduc s
de eloped by US-based ech companies. Obse e s ha e no ed ha in many US-based AI i ms, AI
ech wo ke s enjoy a subs an ial deg ee o independence om managemen in he pu sui o hei
wo k (Cihon e al., 2021; Me cal e al., 2019). Acco dingly, AI ech wo ke s, gi en hei abo e-
men ioned cha ac e is ics and p i ileges, a e likely o de ise AI sys ems, including hose used in
he wo ld o wo k, ha acco d wi h hei alues and no ms, which a e claimed o e ol e a ound
concep s such as p oduc i i y maximisa ion, compe ence de elopmen , success sus enance and
cul u al compa ibili y (C aw o d e al., 2019; De S e ano, 2019). As he a gumen p oceeds, hese
may con lic wi h he p inciples upheld by indi iduals om a ious social, economic, e hnic, o
cul u al backg ounds when hey encoun e AI-d i en echnologies (EPSC, 2018; Linkedin, 2019;
Se oz, 2019).
The hi d ac o ha ende s WAITs ou comes de imen al o wo ke s’ in e es s is he lack o
obus e hical p ac ices among AI companies. As highligh ed by a Deloi e (2022) su ey, his issue
is a global conce n as he p ima y ocus o mos companies de eloping and deploying eme ging
echnologies ends o be on echnical equi emen s and minimal compliance a he han on obus
e hical conside a ions designed o maximise he bene i s o hese echnologies while minimising
po en ial ha m. When conside ed in he con ex o WAITs, his limi ed ocus on e hical issues
6 T ans e 00(0)
becomes pa icula ly conce ning o US ech companies. As Neg ón (2021a) obse es, hese com-
panies a e leading in WAITs de elopmen , sugges ing ha many WAITs deployed in he EU a e
likely o be o US o igin. Acco ding o Me cal e al. (2019), many US-based AI companies a oid
po en ially shame ul public exposu e o hei p oduc s and p ac ices by c ea ing a açade o e hical
e o ins ead o making he necessa y s uc u al changes o add ess he unde lying e hical issues
inhe en in hei p oduc s. This app oach makes WAITs de eloped and ma ke ed by US-based ech
companies highly likely o pe pe ua e exploi a i e wo kplace p ac ices ha unde mine wo ke s’
igh s and eedoms, as hese sys ems a e o en designed o p io i ise e iciency and p o i abili y
o e ai ness, well-being and us in he wo kplace (Özkizil an and Hassel, 2021). Nex , we will
discuss he impac o AI on wo ke s’ igh s and eedoms, as ou lined by he Cha e .
AI as a h ea o wo ke s’ igh s in he Eu opean wo kplace
The undamen al igh s and eedoms ha p o ec wo ke s a e ensh ined in he Cha e , which
gua an ees, amongs o he hings, espec o p i a e li e (A icle 7), he p o ec ion o pe sonal da a
(A icle 8), eedom o assembly and associa ion (A icle 12), non-disc imina ion (A icle 21),
equali y be ween women and men (A icle 23), wo ke s’ igh s o in o ma ion and consul a ion
(A icle 27), he igh o collec i e ba gaining and ac ion (A icle 28), and ai and jus wo king
condi ions (A icle 31).
The deploymen o WAITs, as documen ed in schola ly esea ch, may pose signi ican chal-
lenges o a i ming some o hese undamen al igh s. Fo example, he ex ensi e use o WAITs in
moni o ing and su eillance h ea ens wo ke s’ igh s o p i acy and he p o ec ion o pe sonal da a
(Adams-P assl, 2019; Ajunwa e al., 2017; Rani e al., 2024). Using WAITs ained on biased da a
and lawed algo i hms in hi ing, i ing and p omo ion decisions may pe pe ua e disc imina ion and
unde mine he p inciples o non-disc imina ion and gende equali y (De S e ano, 2019; Özkizil an
and Hassel, 2021). Algo i hmic managemen ools may also in inge upon wo ke s’ eedom o
assembly and associa ion by exe ing con ol o e hei ac i i ies and limi ing hei abili y o o gan-
ise and ac collec i ely (Kellogg e al., 2020; Se oz, 2019).
Mo eo e , deploying WAITs wi hou adequa e pa icipa ion om wo ke s o hei ep esen a-
i es isks unde mining he igh o wo ke s o in o ma ion and consul a ion. To add o his, algo-
i hmic managemen ools ha se un easonable wo k hou s, dis ibu e shi schedules con a y o
wo ke s’ needs and apply a bi a y pe o mance c i e ia (Özkizil an and Hassel, 2021; Rani e al.,
2024) could iola e he igh o wo ke s o ai and jus wo king condi ions. Addi ionally, as WAITs
become mo e p e alen in labou ma ke s, wo ke s whose asks and skills a e highly suscep ible o
au oma ion may encoun e excessi ely long o sho wo king hou s, educed wages, insecu e em-
po a y wo k con ac s, and diminished ba gaining powe (Özkizil an and Hassel, 2021). These
condi ions u he in inge upon he igh o wo ke s o ai and jus wo king condi ions. Nex , we
examine how he EU AI Ac go e ns WAITs deploymen and de elopmen and i s sho comings.
The EU AI ACT
The EU AI Ac , which came in o o ce in Augus 2024 wi h i s equi emen s se o be implemen ed
g adually o e ime (Fu u e o Li e Ins i u e, 2024), ma ks a signi ican miles one as he wo ld’s
i s comp ehensi e legal amewo k o AI go e nance (Council o he EU, 2024; Madinie ,
2021). I aims o achie e a wo-p onged e ec : o egula e AI ha poses a isk o undamen al
igh s, while bols e ing Eu ope’s global compe i i eness in AI de elopmen . The legisla ion
adhe es o a isk-based app oach, whe eby he AI sys ems wi h a highe isk o ha m ace s ic e
egula ions (Council o he EU, 2024). The Ac s ipula es ha p o ide s o high- isk AI sys ems
Özkizil an and Landini 7
mus ensu e ha hei p oduc s comply ully wi h he ele an EU legisla ion (A icle 8). The Ac
also equi es ha high- isk AI sys ems be designed and de eloped o allow o e ec i e human
o e sigh du ing deploymen (A icle 14). The AI Ac iden i ies AI sys ems used in employmen ,
wo ke s’ managemen and access o sel -employmen as high- isk, equi ing compliance wi h
Eu opean s anda ds wi hin 24 mon hs o he Ac ’s en y in o o ce (Fu u e o Li e Ins i u e, 2024).
The Ac p omo es human-cen ic and us wo hy AI, posi ioning i as a ool o people’s well-
being. I os e s an EU-wide commi men o hese p inciples by upholding he Cha e ’s p o ec-
ions o heal h, sa e y and undamen al igh s. Wi h ega d o he wo ld o wo k, he Ac s ongly
emphasises i s applica ion o os e employmen (Reci al 2) and unde sco es i s commi men o
ensu ing wo ke s’ p o ec ion (Reci al 9). Al hough he Ac ’s e o s o align AI de elopmen wi h
EU alues a e commendable, a c i ical analysis sugges s ha se e al sho comings signi ican ly
unde mine i s e ec i eness in p o ec ing wo ke s’ igh s and eedoms, as se ou in he Cha e .
Fi s , despi e championing he p inciples o a human-cen ic and us wo hy AI, he Ac ’s p i-
ma y goal is o es ablish a cohesi e legal amewo k o acili a e AI in eg a ion in o he in e nal
ma ke (Reci al 1). This is because legally he Ac is based on A icle 114 (The Ac , Reci al 3) o
he TFEU. This legal ounda ion con ines he p o ec ion o undamen al igh s o a p o ision ha
g an s he EU gene al au ho i y o ha monising egula ions ela ed o he in e nal ma ke (Almada
and Radu, 2024; Ebe s, 2024). This ocus on ma ke in eg a ion may incen i ise p io i ising eco-
nomic conside a ions o e he obus sa egua ding o undamen al igh s and eedoms o wo ke s
as ou lined by he Cha e , pa icula ly conce ning he p inciples o espec o p i a e li e (A icle
7), eedom o assembly and associa ion (A icle 12), non-disc imina ion (A icle 21), equali y
be ween women and men (A icle 23), he igh o wo ke s o in o ma ion and consul a ion wi hin
he unde aking (A icle 27), he igh o collec i e ba gaining and ac ion (A icle 28), and ai and
jus wo king condi ions (A icle 31).
Second, he EU AI Ac allows p o ide s o high- isk AI sys ems o sel -assess and sel -ce i y
he compliance o hei p oduc s wi h he Ac (A icle 43 and Annex VI). This app oach p io i ises
indus y ease o compliance, sides epping he in ol emen o independen ex e nal bodies, wo k-
e s and ade unions. This is pa icula ly isky o high- isk WAITs, in ela ion o which p o ide s
migh p io i ise speed o ma ke o e ho ough compliance checks du ing hei sel -assessmen s.
Such sys ems isk allowing ha m ul echnologies in o wo kplaces, po en ially iola ing wo ke s’
igh s and eedoms p o ec ed by he Cha e , including espec o p i a e li e (A icle 7), eedom
o assembly and associa ion (A icle 12), non-disc imina ion (A icle 21), equali y be ween women
and men (A icle 23), he igh o collec i e ba gaining and ac ion (A icle 28), and ai and jus
wo king condi ions (A icle 31).
The hi d sho coming is he EU AI Ac ’s exemp ion o p i a e companies deploying high- isk
WAITs om conduc ing undamen al igh s impac assessmen s (FRIAs). FRIAs play a c ucial
ole in iden i ying po en ial isks o ha m o undamen al igh s. The aim o such assessmen s –
whe e app op ia e, wi h he in ol emen o ele an s akeholde s – is o pinpoin pa icula ways
in which AI applica ions could nega i ely impac undamen al igh s wi hin he con ex o hei
use. Fundamen al igh s impac assessmen s go beyond simply iden i ying isks; hey also in ol e
de e mining mi iga ion s a egies speci ic o he use con ex (A icle 27 and Reci al 96). The
absence o manda o y FRIAs o p i a e employe s aises he isk ha ha m may be in lic ed by
deploying high- isk WAITs. Indeed, o one hing, i places he esponsibili y o con i ming ha
AI sys ems do no pose po en ial ha m o wo ke s’ undamen al igh s solely on he p o ide s’ sel -
assessmen . Fo ano he , his exemp ion e ec i ely bypasses he in ol emen o wo ke s and ade
unions in assessing WAITs be o e and du ing hei use. The Ac ’s exemp ion o p i a e deploye s
o high- isk AI om conduc ing undamen al igh s impac assessmen s hus heigh ens he isk o
o e looking o unde es ima ing he po en ial ad e se impac s o WAITs on undamen al igh s
8 T ans e 00(0)
gua an eed by he Cha e , pa icula ly conce ning he p inciples o espec o p i a e li e (A icle
7), eedom o assembly and associa ion (A icle 12), non-disc imina ion (A icle 21), equali y
be ween women and men (A icle 23), he igh o wo ke s o in o ma ion and consul a ion wi hin
he unde aking (A icle 27), he igh o collec i e ba gaining and ac ion (A icle 28), and ai and
jus wo king condi ions (A icle 31).
The ou h sho coming is he lack o manda o y egis a ion equi emen s o high- isk AI sys-
ems deployed by p i a e companies (A icle 49). Indeed, while p i a e sec o deploye s a e
exemp om his equi emen , p o ide s o high- isk AI sys ems and public au ho i ies using hese
sys ems mus egis e hem (Annex VIII). This omission makes i mo e di icul o wo ke s inde-
penden ly o iden i y which WAITs a e being used a hei wo kplace, weakening hei igh s o
in o ma ion and consul a ion wi hin he unde aking, as ou lined in he Cha e (A icle 27).
Closely ela ed o his, he i h sho coming conce ns he igh o wo ke s o an explana ion o
decisions made by high- isk AI sys ems impac ing hei heal h, sa e y o undamen al igh s. E en
hough in p inciple his igh is ensh ined in he Ac (A icle 86), i ails o p o ide meaning ul
sa egua ds in p ac ice. This is because his equi emen does no include consul a ion wi h wo ke s
(Indus iAll, 2024). Nei he does i gua an ee a de ailed desc ip ion o how he AI sys em makes
decisions (Hacke , 2024) be o e he ha m has occu ed. This means ha he Ac en i les wo ke s o
exe cise his igh only when a decision p oduces legal e ec s o simila ly signi ican impac s ha
hey pe cei e as ad e sely a ec ing hei heal h, sa e y, o undamen al igh s. Consequen ly,
wo ke s would no be able o an icipa e o unde s and wha migh lead o a ha m ul decision un il
a e he ha m has occu ed. Addi ionally, deploye s may only know as much as he AI p o ide s
disclose, and esea ch sugges s ha e en he de elope s may no ully unde s and complex sys ems
(Özkizil an and Hassel, 2021). This sho coming impedes wo ke s’ abili y o e ec i ely challenge
biased and un ai AI-d i en decisions in he wo kplace, in inging upon wo ke s’ igh s o in o ma-
ion and consul a ion wi hin he unde aking, as ou lined in he Cha e (A icle 27).
Six h, he unde - egula ion o gene al-pu pose AI (GPAI) is a signi ican conce n. Acco ding o
he Ac , he p o ide s o ounda ional gene al-pu pose AI models play a c i ical ole in he AI alue
chain, as hei models o en se e as he building blocks o downs eam applica ions. Consequen ly,
i is c ucial ha downs eam use s possess a comp ehensi e knowledge o hese models o ensu e
e ec i e in eg a ion and adhe ence o egula o y equi emen s (Reci al 101). Despi e his impo -
ance, he Ac ocuses p ima ily on ad anced GPAI models ha p esen sys emic isks (A icles 51
and 55). I manda es ha hei p o ide s e alua e he p oduc s using s anda d p o ocols, iden i y
and mi iga e sys emic isks, and epo se ious inciden s o he AI O ice and ele an na ional
au ho i ies. In con as , p o ide s o GPAI models classi ied as no posing sys emic isks a e sub-
jec only o speci ic anspa ency s anda ds. They mus main ain de ailed eco ds o hei p oduc
de elopmen and e alua ion p ocesses and sha e his in o ma ion wi h businesses in e es ed in
using hei GPAI models while sa egua ding hei in ellec ual p ope y (A icle 53).
In e p e ed in he con ex o WAITs, his compliance amewo k implies ha de elope s elying
on GPAI models ha a e no conside ed o ca y sys emic isks may no ecei e su icien in o ma-
ion om he p o ide s o hese models abou he po en ial downs eam impac s on he sys ems
hey design. While he Ac p o ides a amewo k o assessing he compliance o AI p oduc s, Kak
e al. (2023: 2, b acke s added) highligh a conce ning aspec : ‘GPAI models ca y inhe en isks
. . . [ ha ] can be ca ied o e o a wide ange o downs eam ac o s and applica ions, [and] hey
canno be e ec i ely mi iga ed a he applica ion laye .’ This si ua ion is pa icula ly conce ning
o he design o WAITs, which use GPAI models as hei ounda ion, gi en ha hei po en ial o
bias and disc imina ion migh be ampli ied when hese powe ul models a e embedded wi hin
hese ools. Indeed, as Kak e al. (2023: 4) ap ly pu i : ‘ he ac ha such GPAI models can be ine-
uned o speci ic uses and asks only heigh ens he isk ha he esul s would be un ai , inaccu a e,
Özkizil an and Landini 15
Moss E and Me cal J (2020) E hics owne s: A new model o o ganiza ional esponsibili y in da a-d i en
echnology companies. A ailable a : h ps://da asocie y.ne /wp-con en /uploads/2020/09/E hics-
Owne s_20200923-Da aSocie y.pd (accessed 15 Feb ua y 2023).
Mu o M, Maxim R and Whi on J (2019) Au oma ion and A i icial In elligence: How Machines a e A ec ing
People and Places. Washing on, DC: Me opoli an Policy P og am a B ookings.
Nedelkoska L and Quin ini G (2018) Au oma ion, Skills Use and T aining. Pa is: OECD.
Neg ón W (2021a) Bosswa e and employmen ech da abase. A ailable a : h ps://home.cowo ke .o g/wo k-
ech/ (accessed 26 June 2024).
Neg ón W (2021b) Li le ech is coming o wo ke s: A amewo k o eclaiming and building wo ke powe .
A ailable a : h ps://home.cowo ke .o g/wo k ech/ (accessed 23 Ma ch 2023).
OECD (2019) A i icial In elligence in Socie y. Pa is: OECD Publishing.
Özkizil an D and Hassel A (2021) A i icial in elligence a wo k: An o e iew o he li e a u e. DOI: 10.2139/
ss n.3796746.
Ponce Del Cas illo A (2023) The AI Ac : De egula ion in disguise. A ailable a : h ps://www.socialeu ope.
eu/ he-ai-ac -de egula ion-in-disguise (accessed 24 July 2024).
Rani U, Pesole A and Gonzalez Vazquez I (2024) Algo i hmic Managemen P ac ices in Regula Wo kplaces:
Case S udies in Logis ics and Heal hca e. Luxembou g: Publica ions O ice o he Eu opean Union.
Se oz M (2019) AI: The u u e o wo k? Wo k o he u u e! Eu opean Commission. A ailable a : h ps://
ec.eu opa.eu/digi al-single-ma ke /en/news/ u u e-wo k-wo k- u u e (accessed 18 May 2020).
Webb M (2020) The impac o a i icial in elligence on he labo ma ke . A ailable a : SSRN: h ps://ss n.
com/abs ac =3482150 (accessed 12 June 2020).
Williams BA, B ooks CF and Shma gad Y (2018) How algo i hms disc imina e based on da a hey lack:
Challenges, solu ions, and policy implica ions. Jou nal o In o ma ion Policy 8: 78–115.
Yuan S, Shao Z, Wei X e al. (2020) Science behind AI: The e olu ion o end, mobili y, and collabo a ion.
Scien ome ics 124(2): 993–1013.
