Buffer or bottleneck? Employment exposure to generative AI and the digital divide in Latin America

Gmy ek, Paweł; Winkle , He nan; Ga gan a, San iago
Wo king Pape
Bu e o bo leneck? Employmen exposu e o gene a i e
AI and he digi al di ide in La in Ame ica
ILO Wo king Pape , No. 121
P o ided in Coope a ion wi h:
In e na ional Labou O ganiza ion (ILO), Gene a
Sugges ed Ci a ion: Gmy ek, Paweł; Winkle , He nan; Ga gan a, San iago (2024) : Bu e o
bo leneck? Employmen exposu e o gene a i e AI and he digi al di ide in La in Ame ica, ILO
Wo king Pape , No. 121, ISBN 978-92-2-041003-5, In e na ional Labou O ganiza ion (ILO), Gene a,
h ps://doi.o g/10.54394/TFZY7681
This Ve sion is a ailable a :
h ps://hdl.handle.ne /10419/302848
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/3.0/igo/
XBu e o Bo leneck? Employmen
Exposu e o Gene a i e AI and he
Digi al Di ide in La in Ame ica
Au ho s / Paweł Gmy ek, He nan Winkle , San iago Ga gan a
July / 2024
ILO Wo king Pape 121
Copy igh © In e na ional Labou O ganiza ion and he Wo ld Bank 2024
This is an open access wo k dis ibu ed unde he C ea i e Commons A ibu ion 3.0 IGO
License (h p://c ea i ecommons.o g/licenses/by/3.0/igo). Use s can euse, sha e, adap
and build upon he o iginal wo k, as de ailed in he License. The ILO and The Wo ld Bank mus
be clea ly c edi ed as he owne s o he o iginal wo k. The use o he emblem o he
ILO and The Wo ld Bank is no pe mi ed in connec ion wi h use s’ wo k.
A ibu ion – The wo k mus be ci ed as ollows: Gmy ek, P., Winkle , H., Ga gan a, S. Bu e
o
Bo leneck? Employmen Exposu e o Gene a i e AI and he Digi al Di ide in La in Ame ica. ILO
Wo king
Pape 121. Gene a: In e na ional Labou O ice and The Wo ld Bank, 2024.
T ansla ions – In case o a ansla ion o his wo k, he ollowing disclaime mus be added
along
wi h he a ibu ion: This ansla ion was no c ea ed by he In e na ional Labou
O ganiza ion (ILO) o The Wo ld Bank and should no be conside ed an o icial ILO o Wo ld Bank
ansla ion. The ILO and The Wo ld Bank a e no esponsible o he con en o accu acy o his
ansla ion.
Adap a ions – In case o an adap a ion o his wo k, he ollowing disclaime mus be added
along wi h he a ibu ion: This is an adap a ion o an o iginal wo k by he In e na ional Labou
O ganiza ion (ILO) and The Wo ld Bank. Responsibili y o he iews and opinions exp essed in he
adap a ion es s solely wi h he au ho o au ho s o he adap a ion and a e no endo sed by he
ILO o The Wo ld Bank.
This CC license does no apply o non-ILO o Wo ld Bank copy igh ma e ials included in his
publica ion. I he ma e ial is a ibu ed o a hi d pa y, he use o such ma e ial is solely
esponsible o clea ing he igh s wi h he igh holde .
Any dispu e a ising unde his license ha canno be se led amicably shall be e e ed o
a bi a
ion in acco dance wi h he A bi a ion Rules o he Uni ed Na ions Commission on
In e na ional T ade Law (UNCITRAL). The pa ies shall be bound by any a bi a ion awa d
ende ed as a esul o such a bi a ion as he inal adjudica ion o such a dispu e.
All que ies on igh s and licensing should be add essed o he ILO Publishing Uni (Righ s and
Licensing), 1211 Gene a 22, Swi ze land, o by email o [email p o ec ed].
ISBN 9789220410028 (p in ), ISBN 9789220410035 (web PDF), ISBN 9789220410042
(epub), ISBN
9789220410066 (mobi), ISBN 9789220410059 (h ml). ISSN 2708-3438
(p in ), ISSN 2708-3446 (digi al)
h ps://doi.o g/10.54394/TFZY7681
The designa ions employed in ILO and Wo ld Bank publica ions and he p esen a ion o
ma e ial he ein do no imply he exp ession o any opinion wha soe e on he pa o he
ILO and The Wo ld Bank conce ning he legal s a us o any coun y, a ea o e i o y o o i s
au ho i ies, o conce ning he delimi a ion o i s on ie s o bounda ies. De ails a
www.ilo.o g/disclaime
This a icle is a p oduc o he s a o he Wo ld Bank and he ILO. I has been eleased bo h in
The Wo ld Bank Policy Resea ch Wo king Pape Se ies and he ILO Wo king Pape Se ies. The
esponsibili y o opinions exp essed in signed a icles, s udies and o he con ibu ions
es s
solely wi h hei au ho s, and publica ion does no cons i u e an endo semen by he ILO
o The Wo ld Bank o he opinions exp essed in hem.
Re e ence o names o i ms and comme cial p oduc s and p ocesses does no imply hei en-
do semen by he ILO o The Wo ld Bank, and any ailu e o men ion a pa icula i m,
comme cial p oduc o p ocess is no a sign o disapp o al.
In o ma ion on ILO publica ions and digi al p oduc s can be ound a : h ps://www.ilo.o g/
esea ch-and-publica ions
ILO Wo king Pape s summa ize he esul s o ILO esea ch in p og ess, and seek o
s imula e
discussion o a ange o issues ela ed o he wo ld o wo k. Commen s on his ILO
Wo king Pape
a e welcome and can be sen o [email p o ec ed].
Au ho iza ion o publica ion: Richa d Samans, Di ec o , Resea ch Depa men
ILO Wo king Pape s can be ound a : h ps://www.ilo.o g/ esea ch-and-publica ions#wo king-
pape s
Sugges ed ci a ion:
Gmy ek, P., Winkle , H., Ga gan a, S. 2024. Bu e o Bo leneck? Employmen Exposu e
o Gene a i e AI and he Digi al Di ide in La in Ame ica, ILO Wo king Pape 121 (Gene a,
ILO and The Wo ld Bank, 2024). h ps://doi.o g/10.54394/TFZY7681
01 ILO Wo king Pape 121
Abs ac
Empi ical e idence on he po en ial impac s o gene a i e a i icial in elligence (GenAI) is mos ly
ocused on high-income coun ies. In con as , li le is known abou he ole o his echnology
on he u u e economic pa hways o de eloping economies. This pape con ibu es o ill his gap
by es ima ing he exposu e o he La in Ame ican labou ma ke o GenAI. I p o ides de ailed
s a is ics o GenAI exposu e be ween and wi hin coun ies by le e aging a ich se o ha monized
household and labou o ce su eys. To accoun o he slowe pace o echnology adop ion in
de eloping economies, i adjus s he measu es o exposu e o GenAI by using he likelihood o
accessing digi al echnologies a wo k. This is hen used o assess he ex en o which he digi al
di ide ac oss and wi hin coun ies will be a ba ie o maximize he p oduc i i y gains among
occupa ions ha could o he wise be augmen ed by GenAI ools. The indings show ha ce ain
cha ac e is ics a e consis en ly co ela ed wi h highe exposu e. Speci ically, u ban-based jobs
ha equi e highe educa ion, a e si ua ed in he o mal sec o , and a e held by indi iduals wi h
highe incomes a e mo e likely o come in o in e ac ion wi h his echnology. Mo eo e , he e is
a p onounced il owa d younge wo ke s acing g ea e exposu e, including he isk o job au-
oma ion, pa icula ly in he inance, insu ance, and public adminis a ion sec o s. When adjus -
ing o access o digi al echnologies, he indings show ha he digi al di ide is a majo ba ie
o ealizing he posi i e e ec s o GenAI on jobs in he egion. In pa icula , nea ly hal o he po-
si ions ha could po en ially bene i om augmen a ion a e hampe ed by lack o use o digi al
echnologies. This nega i e e ec o he digi al di ide is mo e p onounced in poo e coun ies.
Abou he au ho s
Paweł Gmy ek is a Senio Resea che a he Resea ch Depa men o he ILO.
He nan Winkle is a Senio Economis a he Wo ld Bank Po e y and Equi y Global P ac ice o
La in Ame ica and he Ca ibbean.
San iago Ga gan a is a Senio Resea che a he Cen e o Dis ibu i e, Labo and Social S udies
(CEDLAS) o he Na ional Uni e si y o La Pla a (UNLP).

02 ILO Wo king Pape 121
Abs ac 01
Abou he au ho s 01
Ac onyms 05
XIn oduc ion 06
X1 LAC egion and he heo e ical e ec s o GenAI 08
X2 Me hods 15
Occupa ional exposu e o GenAI 15
Use o a compu e a wo k 19
X3 Findings 22
C oss-coun y compa isons o he le els o exposu e 22
Impac o digi al in as uc u e on he po en ial o ans o ma ion 26
Wi hin-coun y pa e ns 29
Which occupa ions d i e he e ec s? 30
Di e en ial exposu e ac oss ea nings le els 32
XFinal discussion 35
Appendix 38
Re e ences 45
Acknowledgemen s 50
Table o con en s
03 ILO Wo king Pape 121
Lis o Figu es
Figu e 1. GDP pe capi a, popula ion and income s a us o LAC coun ies in he sample 08
Figu e 2. Au oma ion and augmen a ion po en ial: LAC s o he egions 09
Figu e 3. In e ne co e age s pe capi a income: global and LAC 11
Figu e 4. Occupa ions in he LAC egion, by ISCO 1-digi and gende 13
Figu e 5. Co e age o ISCO-08 4-digi mic oda a in SEDLAC (WB) and ILO ha monized mic o-
da a collec ion 17
Figu e 6. Hie a chical clus e ing based on ISCO 2-digi sha es, GDP(PPP) and o al popula ion 18
Figu e 7. To al exposu e o GenAI by coun y 23
Figu e 8. Au oma ion po en ial - de ailed b eakdown o socio-economic cha ac e is ics 24
Figu e 9. Augmen a ion po en ial - de ailed b eakdown o socio-economic cha ac e is ics 25
Figu e 10. Jobs wi h augmen a ion po en ial and access o compu e a wo k, based on
PIAAC da a 27
Figu e 11. Exposu e by coun y, exposu e ype and access o digi al in as uc u e 28
Figu e 12. Exposu e by coun y, ype and de ailed coun y-le el cha ac e is ics 30
Figu e 13. ISCO 2-digi occupa ions by ype o exposu e and coun y (sha e o exposu e > 25%) 31
Figu e 14. Ea nings o occupa ions exposed o GenAI, by employmen s a us (exposu e
abo e 25%) 33
Figu e A 1. Compa ison o TechXposu e sco es s GBB sco es (mean by occupa ion, z-sco es) 38
Figu e A 2. Compa ison o Fel en e al. (2023) ML sco es s GBB sco es (z-sco es) 38
Figu e A 3. Labou ma ke dis ibu ion in LAC coun ies by ISCO-08 2-digi occupa ions and sex 39
Figu e A 4. Ranking o coun ies by he ype o GenAI exposu e 40
Figu e A 5. Compa ison o esul s on compu e use be ween PIAAC (a wo k) and SEDLAC (a
home) - augmen a ion ca ego y 40
Figu e A 6. Jobs in augmen a ion ca ego y ha do no use a compu ed a wo k: o als by
coun y 40
04 ILO Wo king Pape 121
Lis o Tables
Table 1. Dis ibu ion o AI Exposu e by Demog aphic and Socioeconomic Ca ego ies in
SEDLAC Da a 19
Table A 1. Indi idual SEDLAC obse a ions by coun y and yea 41
Table A 2. Es ima ed coe icien s o compu e use a wo k om PIAAC 41
Table A 3. Resul s o he pooled OLS wi h all indi idual obse a ions, wi h coun y-le el no -
malized popula ion weigh s 43
05 ILO Wo king Pape 121
Ac onyms
EM Eme ging Ma ke s
GDP G oss Domes ic P oduc
GBB Gmy ek, Be g and Bescond (as used in you s udy o ci a ion)
GenAI Gene a i e AI
GPT-4 Gene a i e P e- ained T ans o me 4
HIC High Income Coun ies
ILO In e na ional Labou O ganiza ion
IMF In e na ional Mone a y Fund
ISCO In e na ional s anda d Classi ica ion o Occupa ions
ISCO-08 In e na ional S anda d Classi ica ion o Occupa ions, 2008 e sion
LAC La in Ame ica and he Ca ibbean
LLM La ge Language Models
OECD O ganiza ion o Economic Coope a ion and De elopmen
PIAAC P og amme o he In e na ional Assessmen o Adul Compe encies
PPP Pu chasing Powe Pa i y
SEDLAC Socio-Economic Da abase o La in Ame ica and he Ca ibbean
TFP To al Fac o P oduc i i y
US Uni ed S a es
WB Wo ld Bank
WEF Wo ld Economic Fo um
12 ILO Wo king Pape 121
Fi h, he esul s o hese ecen expe imen s and mac oeconomic models do no conside gen-
e al equilib ium o second o de e ec s on employmen . Fo example, while inc eased p o-
duc i i y may b ing employmen and wage gains in sec o s acing a consume demand ha is
g owing apidly, ha may no be he case o sec o s acing a mo e s able consume demand
(Au o , 2024). The na u e o hese second o de e ec s is likely o be di e en ac oss coun ies.
In de eloping economies wi h a la ge ac ion o he wo k o ce in he in o mal sec o , and whe e
echnology adop ion and p i a e sec o in es men a e ypically concen a ed among a small
sha e o o mal i ms (Ci e a and C uz, 2022), wo ke s displaced om o mal sec o jobs may
ace mo e challenges inding high quali y jobs han hei coun e pa s in high-income coun ies.
While de ailed mac oeconomic modelling o such e ec s is beyond he scope o ou s udy, he
es ima es o jobs’ exposu e o GenAI p esen ed in his pape p o ide a p o ile o he socio-eco-
nomic g oups mo e likely o expe ience he i s -o de impac s.
His o ically, oge he wi h Sub-Saha an A ica, LAC is one o he mos unequal egions in he
wo ld (Wo ld Bank, 2016a), wi h le els o income inequali y s ongly in luenced by he changes
in he s uc u e o he labou ma ke (Aze edo e al., 2013). Conce ns abou he impac s o new
echnologies on inequali y in LAC a e consis en wi h b oade empi ical e idence abou he e -
ec s o ecen wa es o echnological change on labou demand, which ended o be skill-biased
and o widen he gap be ween low- and high-income wo ke s (Acemoglu and Res epo, 2022;
Au o e al., 2008). Acemoglu's (2024) mos ecen modelling o GenAI ou comes on wages and
inequali y also sugges s ha in nea ly all heo e ical scena ios, he deploymen o his echnol-
ogy a he wo kplace is likely o inc ease he inequali y be ween capi al and labou , and esul
in highe income inequali y be ween di e en demog aphic g oups, wi h pa icula ly nega i e
consequences o he incomes o low-educa ion women in he US.5 In he case o LAC coun ies,
Du z e al. (2018) conduc a comp ehensi e discussion o he se e al challenges o he egion
in e ms o digi al echnologies and how inclusi e hey migh be, looking a di e se case s udies
o echnology adop ion in La in Ame ica. In his ega d, he dis ibu i e impac o he AI adop-
ion depends s ongly on how he e ec s o inc eased p oduc i i y and ou pu could o e come
labou displacemen conduc ed by subs i u ion o echnology o wo ke s. Al hough he use o
AI echnologies in La in Ame ica emains s ill e y low, ecen empi ical e idence shows LAC la-
bou pa e ns mo e consis en wi h he skill-biased echnological change hypo hesis han he
job pola iza ion model6 (B ambilla e al., 2023; Messina e al., 2016; Messina and Sil a, 2018).
Simila conclusions a ise om he li e a u e on employmen and au oma ion in he es o he
de eloping wo ld (Das and Hilgens ock, 2022), al hough Maloney and Molina (2016) ind some
e idence o incipien pola iza ion in B azil and Mexico.
To u he heo ize he po en ial e ec s o GenAI di usion on inequali y in he egion, Figu e 4
p esen s he mos ecen b eakdown o LAC occupa ions by he highes , 1-digi le el o ISCO-08,7
e ealing isible di e ences in he employmen s uc u es ac oss gende s.
5Albei wi h smalle wage e ec s han he p e ious wa es o au oma ion (see Acemoglu and Res epo, 2022).
6The skill-biased echnological change (SBTC) hypo hesis sugges s ha echnology bene i s skilled wo ke s, inc easing demand o
high-skill jobs and widening wage inequali y. In con as , he job pola iza ion model posi s ha echnology c ea es mo e high-skill
and low-skill jobs, educing middle-skill job oppo uni ies and hollowing ou he middle class.
7Elemen a y occupa ions a e g ouped oge he wi h ag icul u al, ishe y and o es y wo k (96).

13 ILO Wo king Pape 121
XFigu e 4. Occupa ions in he LAC egion, by ISCO 1-digi and gende
No e: The b eakdowns a e p esen ed as a sha e o male and emale employmen sepa a ely and calcula ed as a mean sha e o
employmen ac oss he coun ies in each income b acke , based on ILO modelled es ima es (ILO, 2023a).
Fo men, he la ges sha e o employmen is in he elemen a y, ag icul u al, o es y and ishe y
wo k, ollowed by c a and ela ed ade wo ke s. Fo women, he la ges employmen ca ego-
ies conce n se ice and sales wo k, ollowed by elemen a y jobs. Among he “Se ice and sales
wo ke s”, he pa e n is e y simila ac oss coun y g oups, wi h male employmen dominan
only in p o ec i e se ices, and emale employmen ha ing much highe sha es in pe sonal ca e,
sales and pe sonal se ice wo k. A mo e de ailed analysis a ISCO-08 2-digi le el8 e eals ha –
excluding IT, science and enginee ing p o essions – women a e signi ican ly mo e ep esen ed
ac oss all p o essional ca ego ies, wi h pa icula p ominence in eaching, heal h, business ad-
minis a ion and legal, social and cul u al occupa ions. This end ex ends in o cle ical wo k and
ampli ies in line wi h coun ies’ income s a us. This wa an s a en ion, since ecen esea ch
has iden i ied cle ical and p o essional job ca ego ies as being mo e exposed o he isks o au-
oma ion wi h GenAI (Cazzaniga e al., 2024; Gmy ek e al., 2023; Ozdene on, Hakki, 2023; WEF,
2023), wi h p e-GenAI egional assessmen s also classi ying emale-held jobs in LAC as being a
a highe isk o au oma ion om digi al echnologies (Egana-delSol e al., 2022).
In acco dance wi h he echnical documen a ion o ISCO-08, such di e ences in he occupa ion-
al s uc u es also co espond o a ying le els o skills and educa ional a ainmen , as cle ical
suppo wo ke s, echnicians and p o essionals a e ypically classi ied in he mid- o high-skill
le el b acke s (ILO, 2023b). Gi en ha educa ional a ainmen and ea nings gaps ac oss skills
g oups ha e been impo an d i e s o income inequali y in LAC (Aze edo e al. 2013), he impac
o GenAI ha ollows he exis ing labou ma ke s uc u es would likely also ha e an e ec on
he o e all income inequali y. In he bes -case scena io, GenAI would boos he p oduc i i y o
lowe -skilled wo ke s in he exposed occupa ions, allowing hem o access highe incomes and
he e o e leading o a mo e b oad-based income dis ibu ion. In he wo s -case scena io, he
echnological ansi ion could esul in he au oma ion o la gely emale-held jobs in he cle ical,
echnical and p o essional occupa ions, while he oppo uni ies o new GenAI-augmen ed jobs
could be limi ed, gi en he high concen a ion o cu en employmen in elemen a y occupa ions
8Plo A1 in he appendix shows a ull b eakdown o occupa ions a a 2-digi le el o ISCO-08, by coun y and gende .
14 ILO Wo king Pape 121
and in he in o mal sec o , whe e echnology adop ion and p i a e sec o in es men a e low. To
be e unde s and how he i s -o de e ec s o GenAI may a ec inequali y, his s udy p o ides
a de ailed p o ile o he socio-economic g oups mos exposed o his echnology.
Finally, we acknowledge ha he inal ou comes o he echnological ansi ion p ocess will also
be la gely dependen on he exis ing and u u e policy amewo ks in he egion. While he anal-
ysis o coun y-le el polices and legal amewo ks is beyond he scope o his egional s udy, he
de ailed coun y-le el s a is ics ha we make publicly a ailable alongside his publica ion can
se e as use ul inpu s o he discussions unde pinning such policy esponses.9
9Access o de ailed da a a : h ps://pgmy ek.shinyapps.io/AI_Da a_Po al_Resea ch/.
15 ILO Wo king Pape 121
X2 Me hods
Occupa ional exposu e o GenAI
We combine mul iple da ase s o es ima e occupa ional exposu e o AI, le e aging he dis inc
ad an ages inhe en in each da ase o ensu e a comp ehensi e analysis.
We use he AI exposu e sco es a he 4-digi ISCO-08 le el om AI sco es om GBB (2023) as
he p incipal indica o o occupa ional exposu e o GenAI. We also conside al e na i e sco es
ha could be used o his pu pose, in pa icula he abili y-based sco es de eloped by Fel en e
al. (2021, 2023a, 2023b, 2018), based on he US O’NET classi ica ions, ecen ly linked o ISCO-08
4-digi occupa ions by (Cazzaniga e al., 2024), as well as he pa en -based sco es o exposu e
o digi al echnologies by P y ko a e al. (2024), om which a se o echnological g oups ele-
an o GenAI could be isola ed. Ha ing compa ed hese al e na i es, we ind ha Fel en e al.
sco es a e qui e aligned wi h GBB in b oad e ms, excep o ca ching a much wide g oup o
Manage s and P o essionals as highly exposed o AI echnologies.10 Since such b oad co e age
seems somewha un ealis ic in he con ex o many de eloping coun ies, we op o he sco es
o GBB, which p o ide a di ec link o ISCO-08 documen a ion and ocus exclusi ely on GenAI.
This choice is u he ein o ced by he a gumen s ecen ly ad anced by (Nu ski and Rue , 2024),
who ind ha ask- (GBB) and abili y-based (Fel en, 2023) sco es ende simila gene al esul s in
he Eu opean con ex . The au ho s sugges he ask-based analysis using GBB sco es is pa icula ly
ad an ageous o e alua ing employmen impac s, since ask bundles a e be e a ep esen ing
he daily eali y o occupa ions (see Au o , 2015), and conside ing he sha e o a ec ed asks o -
e s mo e scope o sepa a e he po en ial o job ans o ma ion o displacemen due o echno-
logical ad ancemen s. Indeed, ask-based app oaches ha e been widely used in he li e a u e o
his ype o analysis (Res epo, 2023, o conc e e examples see Acemoglu and Res epo, 2022,
2018; F ey and Osbo ne, 2017;), including he ecen ly inc easing use o AI-gene a ed ask-le el
sco es (Acemoglu, 2024; Eloundou e al., 2023), used as a bluep in o he GBB sco ing me hod.
In s ep 1, we ag occupa ions a 4-digi le el in ISCO-08 in o h ee ca ego ies es ablished by GBB:
“au oma ion po en ial”, “augmen a ion po en ial” and “ he big unknown”. We hen ely on he ILO
ha monized mic oda a collec ion11 o ob ain he sha es o employmen a 4-digi le el occupa-
ions o 18 coun ies o each o hese h ee AI exposu e ca ego ies (Figu e 7). We also calcula e
he sha es ha such exposed occupa ions make up in he highe , 2-digi le el o occupa ional
classi ica ion. F om his s ep, we swi ch o he ha monized household su eys om he Socio-
Economic Da abase o La in Ame ica and he Ca ibbean (SEDLAC) o calcula e AI exposu e ac oss
10 See Figu es A2 and A3 in he Appendix o a quick isual compa ison o GBB sco es o Fel en e al. (2023) and P y ko a e al. (2024).
Fel en e al.’s sco es co e a wide ange o AI han GenAI co e ed by GBB. P y ko a e al. (2024) ocus on ech abili ies in pa en s,
which can be a om eadiness o ma ke -le el deploymen . A de ailed analysis o hese sco es a he 4-digi occupa ional le el is
a ailable upon eques .
11 Calcula ions o 2-digi employmen om ILO Mic o da a eposi o y by Da id Bescond, ILO STATISTICS.
16 ILO Wo king Pape 121
and wi hin 16 La in Ame ican coun ies.12 Ou sample o SEDLAC da a consis s o abou 900,000
indi idual su ey obse a ions, wi h de ails by coun y p o ided in he Appendix (Table A1).
XBox 1. GBB Sco es o occupa ional exposu e o GenAI (Gmy ek e al., 2023)
GBB sco es we e de eloped based on he echnical documen a ion o ISCO-08, which con-
ains a lis o ypical asks o each o he 436 de ailed occupa ional g oups a he mos
de ailed, 4-digi le el, and which o ms he basis on which na ional s a is ical labou su -
ey epo s a e linked o he in e na ionally compa ably ISCO-08 s anda d a he ILO. GBB
build on he indings o Eloundou e al. (2023), who demons a e a close alignmen o GPT-
4 p edic ions wi h a su ey o 70 AI expe s on he po en ial o au oma ing occupa ional
asks wi h LLMs, and mo e b oadly on Bubeck e al. (2023), who p o ide ex ensi e es s
o he model’s capabili ies and demons a e i s capaci y o elabo a ing logical links be-
ween i ems, esol ing complex asks and p o iding jus i ica ions o i s decisions. Using
he Applica ion P og amming In e ace (API) o GPT-4, he au ho s designed a sequen ial
call ha loops o e each o 3,123 asks in ha documen a ion and eques ed he model
o assess he echnical easibili y o pe o ming a gi en ask wi h GPT-4 o LLM echnolo-
gy o simila capabili ies. The model is asked o a e asks on a scale o 0 o 1, wi h 1 ep-
esen ing he possibili y o pe o ming a gi en ask by he LLM in ull au onomy o m a
human ope a o , and o elabo a e a w i en jus i ica ion o each sco e (no ask ecei ed
a sco e o 1). The sco es and jus i ica ions a e hen e iewed o consis ency and s abili y
o p edic ions o e ime, wi h he w i en jus i ica ions e iewed by humans. Tasks wi h
sco es abo e 0.8 (high possibili y o au oma ion) a e ans o med in o embeddings, wi h
a seman ic clus e ing algo i hm applied o iden i y he majo g oups o such asks, which
a e subsequen ly e iewed by humans.
Task-le el sco es o each occupa ion a e used o calcula e he mean sco e and he s anda d
de ia ion (SD) o each occupa ion. These wo momen s o dis ibu ion a e subsequen ly
used o elabo a e a heo e ical amewo k o u he classi ica ion o sco es. Occupa ions
(i) wi h a high mean (µi > 0.6) and a high di e ence be ween he mean and SD (µi - σi > 0.5)
a e classi ied as jobs wi h a high au oma ion po en ial. Occupa ions wi h a low mean sco e
(µi < 0.4) and a high sum o he mean and SD (µi + σi > 0.6) a e conside ed o ha e a high
po en ial o augmen a ion, meaning ha while some o hei asks could be au oma ed,
he human ole emains c ucial o he majo i y o hei asks. Occupa ions be ween hese
wo ca ego ies a e classi ied as “ he big unknown”, since, depending on he p og ess o
echnology and he use o adjacen echnological applica ions (e.g. LLM-based agen s), hey
could all close o au oma ion o augmen a ion. Remaining occupa ions a e classi ied as
no a ec ed, wi h he unde s anding ha GenAI in i s cu en o m would ha e minimal
o no impac on hei asks. The sco es and indi idual ask dis ibu ions a e isualized by
he au ho s h ough a publicly a ailable in e ac i e app:
h ps://pgmy ek.shinyapps.io/AI_Da a_Po al_Resea ch/.
While GBB calcula e sepa a e sco es o high- and low-income coun ies, he esul s a e
e y simila and he eby only he high-income ones a e used o all coun ies ega dless
12 SEDLAC is p oduced by he Uni e si y o La Pla a’s cen e o Dis ibu ional, Labo and Social S udies (CEDLAS) and The Wo ld Bank’s
Equi able G ow h, Finance and Ins i u ions LCR-POV-Po e y and Equi y G oup (ELCPV).
This p ojec aims o imp o e he compa abili y o social and economic s a is ics ac oss 25 coun ies in he La in Ame ican and he
Ca ibbean (LAC) egion. This in ol es he ha moniza ion o household su ey a iables in eigh ca ego ies: income, demog aphics,
educa ion, employmen , in as uc u e, du able goods, se ices, and agg ega e wel a e.
17 ILO Wo king Pape 121
o hei income le els. See Appendix o a compa ison o GBB sco es o Fel en e al. (2023)
and P y ko a e al. (2024). See Gmy ek e al. (2023) o a de ailed desc ip ion o he sco e
gene a ing p ocess.
The ad an age o SEDLAC da abase is ha i con ains a hos o ha monized a iables a he indi-
idual le el, including he income agg ega es used o measu e po e y, as well as demog aphic
cha ac e is ics and labou ma ke ou comes. In s ep 2, we impu e he AI exposu e sco es om
GBB o indi idual esponden da a in SEDLAC, using he ISCO-08 occupa ion epo ed in he
household su ey. Such impu a ion is s aigh o wa d o he 8 coun ies wi h 4-digi ISCO-08
occupa ions in SEDLAC, and o which we can di ec ly compa e he calcula ions o he es ima es
om he ILO as an addi ional alida ion measu e (Figu e 5). In con as , he e a e 8 coun ies in
SEDLAC wi h 2-digi ISCO-08 sco es whe e he impu a ion is less ob ious and depends on o he
ci cums ances. We ha e wo ypes o such cases.
XFigu e 5. Co e age o ISCO-08 4-digi mic oda a in SEDLAC (WB) and ILO ha monized mic oda a collec ion
When we ha e he 4-digi ISCO-08 employmen s uc u e om ILO o he same coun y, we use
he es ima es o he sha es o exposu e a he 2-digi le el calcula ed in s ep 2 abo e.13 When we
do no ha e he 4-digi employmen s uc u e om a di e en sou ce o he same coun y, we
use ha o a “simila ” coun y.14 These coun y simila i ies a e de ined in s ep 3, by applying a
hie a chical clus e ing algo i hm o se e al coun y-le el cha ac e is ics including he ull b eak-
down o 2-digi ISCO-08 employmen sha es, GDP pe capi a (PPP) and o al popula ion (Figu e 6).
13 This conce ns B azil, Colombia, Cos a Rica and Mexico.
14 This conce ns A gen ina, Boli ia, Gua emala and Nica agua.

18 ILO Wo king Pape 121
XFigu e 6. Hie a chical clus e ing based on ISCO 2-digi sha es, GDP(PPP) and o al popula ion
In s ep 4, we impu e he es ima ed sha es o au oma ion, augmen a ion and he big unknown
o indi idual esponses a he 2-digi ISCO-08 occupa ion le el in SEDLAC. Ha ing a da a ame
wi h 2-digi le el sha es enables agg ega ion o indi idual esponses by main ca ego ies o in-
e es cap u ed in SEDLAC mic oda a. We ocus on gende (male, emale), a ea ( u al, u ban),
age (15-14, 25-34, 35-44, 45-54, 55-64), educa ion (low, medium, high), po e y s a us (non-poo ,
poo ), income quin iles (Q1 h ough Q5), o mali y (legal, p oduc i e), labou ela ionships (em-
ploye , sala ied employee, sel -employed, amily wo ke wi hou sala y) and sec o o econom-
ic ac i i y. The sha es o exposu e a e calcula ed in such a way ha au oma ion, augmen a ion,
big unknown and o he occupa ions add o 100 pe cen wi hin each ca ego y. This means ha
we can in e p e such esul s as a sha e o employmen in each ype o AI exposu e wi hin each
g ouping ca ego y ( o example, sha es o au oma ion, augmen a ion, big unknown and o he
occupa ions among people wi h low educa ion o among hose belonging o he age b acke o
35-44). Table 1 desc ibes hese a iables in mo e de ail.
19 ILO Wo king Pape 121
XTable 1. Dis ibu ion o AI Exposu e by Demog aphic and Socioeconomic Ca ego ies in SEDLAC Da a15
Va iable name Desc ip ion
Educa ion Low: ewe han 9 yea s o educa ion
Middle: 9 o 13 yea s o educa ion
High: 14 o mo e yea s o educa ion
Po e y An indi idual is conside ed poo (non-poo ) i hey li e in a household whose income
pe capi a is below (abo e) he po e y line o uppe middle-income coun ies (US$
6.85-a-day in pu chasing powe pa i y e ms)
Income quin iles Q1 h ough Q5, by whe he he indi idual’s household income pe capi a is in said quin-
ile.
Fo mali y (legal) A sala ied wo ke is in o mal i hey do no ha e he igh o a pension linked o employ-
men when e i ed
Fo mali y (p oduc i e) An indi idual is conside ed an in o mal wo ke i hey belong o any o he ollowing
ca ego ies: (i) unskilled sel -employed, (ii) sala ied wo ke in a small p i a e i m, (iii) ze-
o-income wo ke . Unskilled wo ke s a e all indi iduals wi hou a e ia y o supe io ed-
uca ion deg ee. Small i ms a e hose wi h 5 o ewe employees. These c i e ia and de i-
ni ions e e o indi iduals’ main job.
Sec o o economic ac i i y P ima y sec o
Low- ech manu ac u ing ( ood, be e ages, obacco, ex iles and clo hing)
O he manu ac u ing
Cons uc ion
Re ail, es au an s, ho els and epai s
U ili ies, anspo and communica ions
Banking, inance, insu ance, p o essional se ices
Public adminis a ion
Educa ion, heal h and pe sonals se ices
Domes ic se ice
Use o a compu e a wo k
The me hod applied so a enables de ailed insigh s in o coun y-le el da a on exposu e o occu-
pa ions o GenAI, wi h u he b eakdowns by demog aphic and socioeconomic cha ac e is ics
o he a ec ed g oups (Table 1). A he same ime, he a ia ion in AI exposu e ac oss and wi hin
coun ies is only d i en by he a ia ion in he occupa ional s uc u es, because he same occu-
pa ion in di e en coun ies uses he same sco e o GenAI exposu e. To add ess his limi a ion,
in he nex s ep we in oduce c oss-coun y a ia ion o occupa ion-le el sco es, by accoun ing
o he a iabili y in he use o compu e equipmen in he same occupa ion loca ed in di e en
na ional con ex s.
15 Fo mo e de ails see h ps://www.cedlas.econo.unlp.edu.a /wp/wp-con en /uploads/Me hodological_Guide_ 201404.pd .
20 ILO Wo king Pape 121
We i s p oceed by impu ing he GenAI exposu e measu es a he 4-digi ISCO08 o he mic o-
da a om he P og amme o he In e na ional Assessmen o Adul Compe encies (PIAAC) col-
lec ed by he OECD. These su eys include ich in o ma ion on de ailed asks ca ied ou by peo-
ple a wo k, such as whe he wo ke s use a compu e (and in e ne )16 a wo k. Using his bina y
indica o , we spli each g oup o GenAI exposu e in o hose who use a compu e a wo k, and
hose who do no . No using a compu e a wo k means ha e en i he wo ke is in an occu-
pa ion ha is exposed o GenAI augmen a ion, such po en ial p oduc i i y gains a e unlikely o
ealize gi en he lack o access o digi al in as uc u e. We i s implemen his exe cise o he
ou coun ies in he LAC egion (Chile, Ecuado , Mexico and Pe u) and wo de eloped economies
(Slo enia and New Zealand) included in he PIAAC da ase .17
Since he e a e only ou La in Ame ican coun ies in PIAAC, we ex apola e he measu es o
compu e use a wo k om PIAAC o he ull se o coun ies in he SEDLAC da abase. In pa icu-
la , we es ima e a p edic i e model o he p obabili y o compu e use a he indi idual le el us-
ing he ull se o coun ies in PIAAC18 and independen a iables ha a e a ailable bo h in he
PIAAC and SEDLAC da abases.19 We hen use he es ima ed model and he se o independen
a iables o p edic he p obabili y o compu e use in he SEDLAC da abase. Mo e speci ically,
we i s es ima e he ollowing Logi model in PIAAC:
()()
compu e ISCOage emale eduin e ne b oadband
P =1 =,,,,GDP ,,
ci
ci
o
ci
a
ci ci cc c
,
,
,
,, (1)
Whe e
compu e ci
, is a bina y a iable equal o 1 i indi idual
i
in coun y
c
uses a compu e a
wo k;
IS
CO
ci
o
, is a ec o o 39 dummy a iables o each 2-digi ISCO08 occupa ion20;
ageci
a
, is a
ec o o 4 dummy a iables indica ing age g oups;
educi
, is a dummy a iable equal o one o
High School g adua es;
GDP
c
is he log o GDP pe capi a in 2017 US$ PPP;
in e ne c
is he a e o
in e ne use s pe 100 people, and;
b oadbandc
is he numbe o ixed b oadband subsc ip ions
pe 100 people. Since he e e ence yea o he PIAAC su eys a ies by coun y, we use he co -
esponding yea o he coun y-le el a iables (i.e. GDP, in e ne and b oadband). These coun-
y-le el a iables a e help ul o cap u e he link be ween he economy-wide le el o digi al and
economic de elopmen wi h he le el o compu e use a wo k.
In he nex s ep, we use he es ima ed equa ion (1) o p edic he p obabili y o using a compu e
a wo k a he indi idual le el in he SEDLAC da abase.21 The p obabili y o no using a compu -
e a wo k is simply 1-P (compu e =1). When choosing he e e ence yea s o he coun y-le el
a iables o he model, we use he e e ence yea o he SEDLAC su eys. Then, he p obabili y
16 Fo he main esul s p esen ed in he pape , we use a iable “g_q04”, which con ains he esponse o he ollowing ques ion: “Do
you use a compu e in you job?/Did you use a compu e in you las job?”. We also do obus ness checks by c ea ing a bina y a ia-
ble equal o 1 when he wo ke uses bo h a compu e and in e ne a wo k, by using he a iables “g_q05a”, “g_q05c”, “g_q05d” and
“g_q05h”, which con ain in o ma ion abou he equency (i.e. “ne e ”, “less han once a mon h”, “less han once a week bu a leas
once a mon h”, “a leas once a week bu no e e y day”, o “e e y day”) o in e ne use o mail, wo k ela ed in o ma ion, conduc
ansac ions and pa icipa e in ideo calls, espec i ely. Mo e speci ically, when he wo ke esponds ha he o she “ne e ” uses he
in e ne o any o hose ou easons a wo k, we conside ha he wo ke does no use in e ne a wo k. I he wo ke esponds ha
he o she uses he in e ne a wo k o any o hose pu poses wi h any equency o he han “ne e ”, hen we conside ha he wo k-
e uses in e ne a wo k. Simul aneous use o compu e and in e ne p esen s a mo e es ic i e condi ion, which esul s in g ea e
digi al gaps. These calcula ions a he coun y le el a e shown in Figu e A5. Mo e de ailed s a is ics a e a ailable upon eques .
17 While PIAAC includes se e al de eloped economies, we did no choose hem o his ini ial s ep o he analysis because hei su eys
we e ei he collec ed se e al yea s ea lie (2011-2012) and he adop ion o digi al echnologies inc eased d ama ically since hen. In
addi ion, some de eloped coun ies wi h mo e ecen su eys do no ha e ISCO 08 in o ma ion a he 4-digi le el (e.g. Uni ed S a es).
In con as , New Zealand and Slo enia’s su eys we e collec ed du ing he second ound (2014-2015), which is he same ime ame
o Chile’s su ey, and close o he ime ame o Ecuado , Mexico and Pe u (2017).
18 The e a e 38 coun ies wi h publicly a ailable PIAAC mic oda a, see h ps://www.oecd.o g/skills/piaac/da a/.
19 A simila exe cise was implemen ed by Ga o e e al., (2021) o adjus wo king- om-home measu es by in e ne access a es.
20
While he e a e 43 ISCO08 2-digi le el ca ego ies, we d op om he sample he ca ego ies 01 (Commissioned A med Fo ces O ice s),
02 (Non-Commissioned A med Fo ces O ice s) y 03 (A med Fo ces Occupa ions, O he Ranks) since hey a e no a ailable in PIAAC.
21 Table A2 in he appendix con ains he es ima ed coe icien s.
21 ILO Wo king Pape 121
o an indi idual being exposed o AI and o using a compu e a wo k will be he mul iplica ion
o bo h indi idual p obabili ies. Fo example, ake he case o a g oup o wo ke s (e.g. wo ke s
wi h high educa ion) ha , on a e age, has a 0.23 GenAI augmen a ion exposu e p obabili y (i.e.
he a e age o he bina y exposu e measu e a he 4-digi le el). Then le ’s assume ha , based on
he indi idual p edic ions om he Logi model, on a e age, indi iduals in such g oup ha e a 0.7
(0.3) likelihood o using (no using) a compu e a wo k. As a esul , we conclude ha wo ke s in
such g oup ha e a 0.161 (=0.7 x 0.23) p obabili y o being exposed o GenAI augmen a ion and
o using a compu e a wo k, while hey ha e a 0.069 (0.3 x 0.23) p obabili y o being exposed o
GenAI augmen a ion and o no using a compu e a wo k.22
In he inal s ep, we calcula e measu es o AI exposu e ac oss he same socio-demog aphic cha
-
ac e is ics as summa ized in Table 1, his ime wi h a simul aneous b eakdown by compu e use
a he wo kplace o each o hese cha ac e is ics.
22 The logic is he same when he measu es o GenAI exposu e a e impu ed a he 2-digi ISCO08 le el, which a e no bina y bu con-
inuous measu es be ween 0 and 1. Tha is, he exposu e o GenAI augmen a ion wi hin a g oup o wo ke s would be he a e age
o he con inuous exposu e measu e.
28 ILO Wo king Pape 121
XFigu e 11. Exposu e by coun y, exposu e ype and access o digi al in as uc u e
We can obse e ha , in he case o jobs exposed o po en ial au oma ion, he sha es o such jobs
ha do no use a compu e a e gene ally e y low ac oss coun ies. In o he wo ds, mos o such
occupa ions a e al eady digi ized. A no able excep ion applies o coun ies wi h lowe income,
such as Nica agua, Gua emala and Hondu as, whe e a ound hal o he jobs exposed o po en-
ial au oma ion a e p edic ed o no use a compu e . One way o hink abou his pa e n is ha ,
in poo e coun ies, he lack o digi al in as uc u e migh o e a empo a y bu e om he
isk o imminen au oma ion o some occupa ions in his ca ego y – a end ha likely ex ends
o coun ies wi h ela i ely lowe incomes ou side La in Ame ica. The plo also e-con i ms ha
such au oma ion-exposed jobs a e disp opo iona ely held by women.
The si ua ion is isibly di e en in he case o occupa ions wi h augmen a ion po en ial. Fi s , he
dis ibu ion o such jobs is mo e equal among women and men. In addi ion, he sha es o jobs
ha do no use a compu e a wo k a e also mo e e enly dis ibu ed ac oss men and women.
An in ui i e way o hinking abou his pa o Figu e 11 is ha he ligh blue zones (no compu -
e ) ep esen he ans o ma ion po en ial ha canno be a ained due o digi al in as uc u e
limi a ions. I one was o assume ha such a ans o ma ion could ansla e in o p oduc i i y

29 ILO Wo king Pape 121
gains in hose occupa ions, he zones wi hou a compu e can be seen as an una ainable po-
en ial p oduc i i y gains.34
Taking his analysis one s ep u he , we can quan i y hese e ec s. Fo example, i we calcula e
he una ained augmen a ion po en ial as an a e age alue ac oss he LAC coun ies (weigh -
ed by o al employmen o each coun y), i co esponds o 6.24 pe cen o emale employmen
and 6.22 pe cen o male employmen . I we hink o his gap as an a e age sha e o jobs wi h-
in all jobs wi h augmen a ion po en ial, i co esponds o a non-negligible 44 pe cen o such
jobs held by women and 50 pe cen o such jobs held by men. Applying hese calcula ions o he
2023 ILO modelled es ima es o o al employmen in each coun y, we es ima e ha he e a e
some 17 million jobs among he 16 LAC coun ies in ou SEDLAC sample ha could, in heo y,
expe ience addi ional p oduc i i y om he echnological ans o ma ion wi h GenAI, bu which
will no be in posi ion o do so due o he lack o digi al in as uc u e. Some 7 million such jobs
a e held by women and nea ly 10 million a e held by men (see Figu e A6 in Appendix o coun-
y-le el b eakdowns).
Wi hin-coun y pa e ns
In his sec ion, we p esen a sample o wi hin-coun y analysis, using he example o Colombia
and Cos a Rica (Figu e 12). De ailed b eakdowns by coun y a e p o ided in he Appendix and
made a ailable in a dynamic way on ou online po al.35
Wi hin each coun y, he plo can be ead as a de ailed b eakdown o a gi en ca ego y on he
e ical axis in o he sha es o employmen ep esen ed in ho izonal sepa a ions. Fo example,
in he case o Colombia, among women, 5.5 pe cen o employmen is exposed o he isk o au-
oma ion, 11.3 pe cen o employmen is in he augmen a ion ca ego y, and 22.5 pe cen in he
big unknown. The emainde , no p esen ed in he g aph, con ains all o he emale-held occupa-
ions ha do no all in o any o hose h ee exposu e ca ego ies. The da k shaded a eas ep e-
sen jobs wi h a compu e a wo k, while he ligh shading ep esen he jobs in each exposu e
ype ha do no use a compu e . We can obse e ha among he emale-held jobs exposed o
po en ial au oma ion, e y ew jobs do no use a compu e , whe eas among hose wi h a ans-
o ma ion po en ial, his digi al cons ain applies o abou hal o such jobs.
34 We also implemen a obus ness check whe e we use he a iable o ”compu e owne ship a home”, as sel - epo ed by households
in SEDLAC, o adjus o digi aliza ion. In pa icula , we eplace he impu ed ”compu e use a wo k” a iable o his measu e, o check
i he pa e ns p e ail. As seen in igu e A.5 in he appendix, while he absolu e alues o exposu e adjus ed by digi aliza ion a e di -
e en ac oss me hods, he ela i e igu es a e e y simila . In pa icula , among jobs exposed o GenAI augmen a ion, he sha e o
wo ke s wi h a compu e a home ends o be highe in iche coun ies.
35 h ps://pgmy ek.shinyapps.io/AI_Da a_Po al_Resea ch/
30 ILO Wo king Pape 121
XFigu e 12. Exposu e by coun y, ype and de ailed coun y-le el cha ac e is ics
Figu e 12 can also be ead e ically, allowing o wi hin-coun y analysis ac oss cha ac e is ics
o bo h he wo ke and he job. Con inuing wi h he example o Colombia, we can obse e ha
he sha e o jobs wi h high au oma ion po en ial is highe among women (5.5 pe cen ) han
men (1.6 pe cen ), and among u ban (3.9 pe cen ) han u al jobs (0.8 pe cen ). The sha e o
such occupa ions is also highes among young people and dec eases wi h age b acke s, while
i inc eases wi h educa ion le els and household income b acke s. While hese gene al ends
ha e al eady been discussed in he p eceding sec ion, he de ailed b eakdowns enable coun-
y-speci ic insigh s, necessa y o a b oade e lec ion on adequa e policy esponses. In o de
o suppo such p ocesses, in addi ion o ou pape , we make his de ailed coun y-le el da a
publicly a ailable online.
Which occupa ions d i e he e ec s?
To unde s and wha d i es hese e ec s, in Figu e 13 we ocus only on occupa ions whe e a leas
a qua e o all indi idual obse a ions unde a gi en ISCO 2-digi ca ego y alls in o one o he
exposu e g oups. Fo example, aking “augmen a ion & compu e ”, we can obse e ha , in many
coun ies, o e 50 pe cen o eaching p o essionals can be ound in his ca ego y. The excep ions
conce n lowe income coun ies, such as Gua emala, Hondu as and Nica agua, whe e he majo -
i y o eaching p o essionals a e in he ca ego y o po en ial augmen a ion, bu wi hou access
31 ILO Wo king Pape 121
o a compu e a wo k. This sugges s ha he di e ences in digi al de elopmen migh impose
impo an limi a ions on he bene i s o GenAI o he educa ion sys ems in poo e coun ies.
XFigu e 13. ISCO 2-digi occupa ions by ype o exposu e and coun y (sha e o exposu e > 25%)
O he occupa ions ha could bene i om augmen a ion and al eady use a compu e a wo k
conce n legal, social and cul u al p o essionals, and nume ical and ma e ial eco ding cle ks in
mos coun ies, as well as heal h p o essionals, in o ma ion and communica ion echnicians and
some o he assemble s in coun ies wi h ela i ely highe incomes. In elabo a ing on hese ca e-
go ies, examples could include legal p o essionals like ax lawye s using GenAI o case analysis,
social wo ke s employing case managemen so wa e, and cul u al p o essionals such as digi al
a chi is s. Assemble s wi h a compu e can include skilled wo ke s like elec icians who may u i-
lize GenAI-based diagnos ic o ins uc ional ools in highe -income coun ies.
32 ILO Wo king Pape 121
Ne e heless, a la ge sha e o he assemble s’ jobs alls in o he ca ego y wi h augmen a ion
po en ial bu no compu e a wo k, alongside pe sonal se ice and e use wo ke s. While such
jobs e ain a cen al human componen , in se ings wi h a compu e and in e ne access, some
o hei asks could bene i om he AI ans o ma ion. Fo example, pe sonal se ice wo ke s,
such as home heal h aides, could use a scheduling and clien managemen so wa e o enhance
se ice coo dina ion, while companies hi ing e use wo ke s could implemen was e acking sys
-
ems and ou e op imiza ion so wa e o imp o e e iciency and educe en i onmen al impac .36
Howe e , due o he lack o digi al in as uc u e in LAC coun ies, such po en ial bene i s would
simply emain ou o each o hose job ca ego ies.
Figu e 13 also demons a es a a ie y o occupa ions alling in o he “big unknown” ca ego y,
wi h a as majo i y o such jobs al eady using a compu e a wo k. This conce ns business and
adminis a ion p o essionals and associa e p o essionals, in o ma ion and communica ion ech-
nicians, cus ome se ice cle ks, gene al and keyboa d cle ks as well as hospi ali y, e ail and
o he se ices manage s. Decomposing hese g oups in o mo e de ailed occupa ions, i is easy
o illus a e why his ca ego y ep esen s he zone be ween he po en ial o ull job au oma ion
and augmen a ion wi h gene a i e AI. Fo example, business and adminis a ion p o essionals
could see GenAI so wa e s eamline complex da a analysis, associa e p o essionals migh use i
o manage logis ics mo e e icien ly, while cus ome se ice cle ks migh ely on GenAI o sup-
po wi h que y esolu ion. As he echnology e ol es, he balance be ween augmen a ion and
au oma ion o asks migh shi , po en ially ede ining some o he jobs in cus ome que y oles
signi ican ly as e and exposing hem o a highe isk o ull au oma ion han o he occupa ions
in his ca ego y. I should also be no ed ha a la ge sha e o cus ome que y cle ks is al eady
ound in he ca ego y o high au oma ion po en ial, alongside o he g oups o cle ical occupa-
ions. The ac ha , in mos cases, such jobs al eady use a compu e a wo k u he sho ens he
dis ance o he po en ial ull au oma ion o hese occupa ions, making po en ial shi s be ween
he “big unknown” and ull au oma ion mo e luid.
Di e en ial exposu e ac oss ea nings le els
As he inal s ep o he analysis, we look deepe in he ela ionship be ween labou ea nings and
he deg ee o occupa ional exposu e o GenAI, ocusing exclusi ely on he occupa ions ha e-
po al eady using a compu e . To do ha , we use he la es a ailable mic o da a o each coun y
and ocus sepa a ely on income o employees and sel -employed. To ensu e compa abili y ac oss
coun ies, we show he median income o each ISCO-08 2-digi le el occupa ion as a pe cen -
age o he median income o wage employees ac oss all obse a ions in each su ey sample.37
36 This clea ly assumes ha such echnologies would be adop ed in a way ha suppo s wo ke s’ asks, a he han imposed o inc ease
he le el o algo i hmic con ol and limi wo ke agency, which can ha e nega i e e ec s on wo king condi ions (Adams-P assl e al.,
2023; ILO, 2023c).
37 The median is a o ed in wage analysis as i esis s skew om ou lie s. I ep esen s he cen al endency o a da a se and p o ides
a clea e indica ion o ypical income le els, especially in ins ances whe e income dis ibu ion is no symme ical.
33 ILO Wo king Pape 121
XFigu e 14. Ea nings o occupa ions exposed o GenAI, by employmen s a us (exposu e abo e 25%)
No e: Incomes we e calcula ed as he median income in local cu ency o each ISCO-08 occupa ion, based on he la es a aila-
ble su ey da a o each coun y. They we e hen ecalcula ed as a dis ance om each g oup’s (employees/sel -employed) me-
dian income and no malized as a sha e o he median income o employees in each coun y, which p o ides a common e e -
ence poin . G ey do s ep esen all occupa ions o which he sample size and exis ing da a allow o calcula ion o he median
income. Colou ed do s ep esen only hese ISCO-08 2-digi occupa ions whe e a leas 25 pe cen o occupa ions in a coun y
wi hin a gi en 2-digi ca ego y a e es ima ed as exposed o au oma ion, augmen a ion o he big unknown and ha e a compu -
e a wo k ( heo e ical eadiness o GenAI-d i en ans o ma ion).
In Figu e 14, we i s plo all coun y-le el da a poin s o LAC (g ey do s), wi h he size o each do
ep esen ing he employmen sha es. The do s below he ho izon al e e ence line a ze o ep-
esen occupa ions wi h an income below he median income o wage employees in each coun-
y. Con e sely, do s abo e ha line e e o jobs wi h income abo e he median income o wage
employees, wi h he lack do ed line ep esen ing he o e all end o highe incomes acc uing o
jobs wi h a lowe numbe in he ISCO-08 s uc u e, ha is, p o essional and manage ial posi ions.

34 ILO Wo king Pape 121
Subsequen ly, we ma k he jobs ha a e exposed o an immedia e in e ac ion wi h GenAI, ha
is, occupa ions wi hin he ca ego ies o au oma ion, augmen a ion and he big unknown, which
epo al eady using a compu e a wo k. We highligh only hese occupa ions in each coun y
whe e a leas a qua e o jobs in a 2-digi ca ego y alls in o one ype o exposu e, wi h expo-
su e ypes ma ked in colou s and he sha e o employmen e lec ed in he size o each ma ke .
We can obse e ha , in he case o wage employees, nea ly all exposed occupa ions a e ei he
a ound o abo e he median income, whe eas o sel -employed we see mo e occupa ions wi h
below-median incomes. The exposu e o au oma ion is gene ally g ouped a ound occupa ions
wi h an income be ween he median and an equi alen o wo median incomes, wi h only some
o he in o ma ion and communica ion jobs exceeding ha h eshold. Occupa ions wi h a high
augmen a ion po en ial a e gene ally g ouped a ound he highe income b acke o somewhe e
be ween 1.5 o 3 alues o he median income o wage employees. The ca ego y o big unknown
is mo e sp ead ac oss he income dis ibu ion, wi h some sel -employed sales wo ke s below he
median le el and he op ea ne s a ound h ee median incomes.
In he bigge pic u e, Figu e 14 shows ha mos o he exposed ca ego ies conce n jobs a ound
wha could be de ined as middle- and uppe -middle income jobs, wi h ha dly any occupa ions
showing signi ican exposu e among he low-income occupa ions. In o he wo ds, he main h us
o he i s o de e ec s o GenAI echnologies can be expec ed among he people who al eady
ha e high incomes and who a e in jobs equi ing ela i ely highe skill le els, while he jobs o
he poo a e qui e likely o emain ou side he immedia e e ec s o his echnological ansi ion.
35 ILO Wo king Pape 121
XFinal discussion
This s udy examines he exposu e o GenAI wi hin he labou ma ke s o he LAC egion, e eal-
ing bo h widesp ead po en ial impac s and signi ican a iabili y ac oss di e en demog aphics
and sec o s. Ou indings indica e ha a subs an ial p opo ion – be ween 30 and 40 pe cen
o employmen in LAC – is exposed in some way o GenAI. This exposu e is linked wi h he eco-
nomic s a us o coun ies, sugges ing ha income le els a e a s ong co ela e o GenAI’s impac
on labou ma ke s. Howe e , i is c ucial o no e ha such exposu e does no imply au oma ion,
and ha o he as majo i y o hese jobs, he po en ial lies in ans o ming he asks ha hese
occupa ions pe o m. Ou es ima es o he po en ial e ec s o au oma ion in LAC amoun o 2
o 5 pe cen o employmen depending on he coun y. These igu es, while seemingly modes ,
should no be i ialized as hey ep esen indi iduals’ li elihoods ha a e a s ake. In addi ion,
some o he jobs om he la ge ca ego y o “ he big unknown” migh mo e close o au oma ion
o e ime, as he echnology and i s applica ions o wo kplace asks de elop u he .
Compa isons o ou esul s o o he s udies a e complica ed, due o he signi ican di e ences
in he concep s applied, occasional lack o de ailed da a ha would enable a mo e p ecise as-
sessmen , di e ging me hods o p esen ing he indings, and he gene al sca ci y o s udies ha
co e non-HIC coun ies (Comunale and Mane a, 2024). Fo example, Eloundou e al. (2023) s a e
ha up o 80% o he US wo k o ce could ha e a leas 10% o hei asks eplaced, while 19% o
wo ke s could lose a leas 50% o hei asks o LLMs – a inding ha is ha d o di ec ly ela e o
ou amewo k, excep o he simila i y o a much s onge augmen a ion e ec o e au oma-
ion. McKinsey (2023) poin s o a simila g oup o “knowledge wo k” as being mos exposed bu
ocus he analy ical wo k on addi ional alue gene a ion h ough p oduc i i y inc eases, a he
han on di ec e ec s on employmen . WEF Fu u e o Jobs (2023), e en hough global in scope,
ocuses exclusi ely on la ge en e p ises, poin ing o cle ical and adminis a i e jobs among oc-
cupa ions wi h he as es expec ed declines. Goldman Sachs (2023), based on ex apola ion o
O*NET occupa ions o eme ging economies, sugges s ha “mos jobs and indus ies a e only
pa ially exposed o au oma ion and a e hus mo e likely o be complemen ed a he han subs i-
u ed by AI”.38 Acco ding o ou bes knowledge, he e a e no p io s udies wi h de ailed insigh s
o GenAI exposu e in he LAC egion and ou es ima es o o al po en ial exposu e a e gene ally
lowe han he 40 pe cen es ima ed by Cazzaniga e al. (2024) o eme ging economies. Howe e ,
weal hie LAC coun ies show exposu e le els close o his es ima e.
Wi hin his con ex , i espec i e o coun y-speci ic di e ences, ou es ima ions show ha ce ain
cha ac e is ics consis en ly co ela e wi h highe GenAI exposu e. Speci ically, u ban-based jobs
ha equi e highe educa ion, a e si ua ed in he o mal sec o , and a e held by indi iduals wi h
highe ela i e incomes a e mo e likely o come in o in e ac ion wi h his echnology. Mo eo e ,
he e is a p onounced il owa ds younge wo ke s acing g ea e exposu e, including he isk o
job au oma ion, in pa icula in he inance, insu ance, and public adminis a ion sec o s. While
hese g oups migh also be be e posi ioned o eap he bene i s o new echnologies (Ananian
e al., 2006; Aube e al., 200639; Cazzaniga e al., 2024), shedding well-paid, o mal, skilled and
emale-domina ed jobs can ha dly be a posi i e scena io o he al eady highly in o mal and un-
equal economies o he LAC egion. Ou indings sugges ha – a leas as he i s -o de e ec
38 The la ge au oma ion sco es in ha s udy a e ha d o compa e o ou indings, since he unde lying da a is no public.
39 S udies o ea lie wa es o echnological change, such as he in oduc ion o he in e ne and digi al inno a ions he wo kplace,
demons a e ha such changes end o pu younge popula ions a an ad an age, in compa ison o olde wo ke s.
36 ILO Wo king Pape 121
– he middle class is he g oup whose jobs and ea nings ha e he highes le els o o e all expo-
su e o GenAI, wi h many possible di ec ions ha his ans o ma ion can ake.
These indings sugges an impo an ole o go e nmen in e en ions, aimed a minimizing
dis up ions esul ing om sudden job losses h ough job p o ec ion measu es, and maximizing
he p oduc i e bene i s o he ansi ion, o example h ough equipping wo ke s wi h ounda-
ional skills ha can help hem keep up wi h he changing cha ac e o jobs and d i e he p o-
duc i e cha ac e o such changes, a he han see hei skills become obsole e. The ac ha
some ulne able g oups, such as women and you h, ace g ea e exposu e o au oma ion high-
ligh s he impo ance o li e-long lea ning so ha ha wo ke s ha e he skills o adap o chang-
es in he wo ld o wo k. In he sho - e m, as shown in nume ous ILO s udies, social p o ec ion
sys ems can play an impo an ole as mac oeconomic shock s abilize s, and educe he impac
o ansi ions o he a ec ed wo ke s and hei households a he mic oeconomic le el (ILO,
2023), especially when hei use is combined wi h skills de elopmen p og ams (ILO, 2023e). In
he medium and long- e m, educing gende gaps in he exposu e o au oma ion would equi e
add essing ac o s ha pe pe ua e occupa ional seg ega ion by gende , such as gende -based
social no ms (Ca anza e al., 2023).
A he same ime, ou indings show ha he sha es o jobs ha could bene i om a p oduc i e
ans o ma ion wi h GenAI a e consis en ly highe han hose wi h au oma ion isks ac oss all
LAC coun ies, anging be ween 8 and 12 pe cen o employmen ac oss coun ies. This is pa -
icula ly he case o he jobs in educa ion, heal h and pe sonal se ices. In addi ion, he sec o s
o ien ed owa ds cus ome se ice ( e ail, ade, ho els, es au an s, e c.) ace an ele a ed expo-
su e o " he big unknown", which means ha a p oduc i e augmen a ion could also be sough
in hese jobs wi h he igh policies and incen i es in place. The e o e, a emp ing na a i e ha
can be cons uc ed based on hese s a is ics is ha , in he big pic u e, mo e can be gained han
los as a ne job and economic e ec o he ans o ma ion.
This is whe e ou analysis p o ides new in o ma ion o assess whe he he lack o digi al in a-
s uc u e could be a bu e o a bo leneck o eap he economic bene i s o GenAI. On he one
hand, ou indings show ha mos wo ke s exposed o GenAI au oma ion a e using digi al ech-
nologies, which sugges s ha he po en ial nega i e e ec s may no ake long o ma e ialize. On
he o he hand, we ind ha inadequa e digi al in as uc u e is a majo bo leneck o ealizing
he posi i e e ec s o augmen a ion, he eby impac ing a signi ican segmen o he labou o ce
in he LAC egion. Nea ly hal o he occupa ions ha could po en ially bene i om augmen-
a ion a e hampe ed by digi al sho comings ha will p e en hem om ealizing ha po en-
ial. Speci ically, 6.24 pe cen o jobs held by women and 6.22 pe cen o hose held by men a e
a ec ed due o hese gaps. Simila limi a ions apply o he jobs in he “big unknown” ca ego y:
e en hough some o hem could po en ially pi o owa ds augmen a ion h ough inc easing
complemen a i y be ween GenAI and he wo ke in hese occupa ions, he digi al gaps will p e-
en la ge sha es o hese jobs om bene i ing om such a scena io.
The ex ension o ou inding is ha he in luence o digi al di ides would likely be e en mo e s a k
in egions o lowe economic de elopmen han LAC, which unde lines he need o equalizing
digi al access in de eloping coun ies. Policies o achie e such goal should include no only hose
ela ed o digi al in as uc u e, bu also hose ha aim o s eng hen he incen i es o adop
digi al echnologies o a p oduc i e use (Wo ld Bank, 2016b)40 o ensu e ha he ans o ma i e
40 Wo ld Bank (2016) p o ides se e al examples o policies no di ec ly ela ed o expanding he digi al in as uc u e ha p omo e he
adop ion o digi al echnologies by i ms and people, such as os e ing mo e compe i ion in p oduc ma ke s (bo h domes ic and
h ough in e na ional ade), imp o ing he quali y o he educa ional sys ems, e c.
37 ILO Wo king Pape 121
p omise o GenAI does no bypass hose who a e mos in need o i s ad an ages. Focus on in o -
mali y in he ansi ion will also be o key impo ance. We ind ha wo ke s wi h o mal jobs a e
mo e exposed o GenAI au oma ion han hei coun e pa s in he in o mal sec o . E en hough
o mal jobs ypically o e some co e age o he social p o ec ion sys ems, e en ual au oma ion o
hese occupa ions does no gua an ee ha he same wo ke s would easily ind new o mal em-
ploymen . Since wo ke s in he in o mal sec o om de eloping coun ies a ely mo e o be e
paying and o mal jobs (Dono an e al., 2023), allowing au oma ion o slowly e ode he exis ing
o mal sec o would simply expand in o mali y. Policies aimed a educing he segmen a ion o
he labou ma ke along he o mal-in o mal lines would help imp o e he chances o displaced
wo ke s’ ansi ioning back o he o mal sec o . In his con ex , i is impo an o ecognize ha
echnological ans o ma ion can ac ually o e oppo uni ies o job e- o maliza ion h ough in-
no a i e go e nmen se ices (Chacal ana Janampa e al., 2024).
Finally, while his s udy p o ides a de ailed o e iew o he LAC egion, i is no wi hou limi a ions,
some o which can be conside ed as open a enues o u u e inqui y. The i s o hose conce ns
da a on he use o compu e s and in e ne a wo k, o which he impu a ion om PIACC was
he bes a ailable s a egy in ou case. Ob aining new su ey da a o a leas he mos exposed
occupa ions would su ely o e mo e p ecision o u u e es ima es in ha ega d. In he absence
o such da a, impu a ion based on he second cycle o OECD’s PIAAC, co e ing 2022-23, migh
be a iable op ion, as soon as such da a become publicly a ailable. Second, mo e could be done
o unde s and how he ask con en o occupa ions a ies ac oss coun ies and in sys ema izing
such di e ences acco ding o income g oups and possibly egional cha ac e is ics beyond LAC.
Thi d, u u e s udies could y o ob ain mo e ine-g ained cha ac e is ics o indi iduals’ in e ne
access, since ne wo k la ency, eliabili y o ype o de ices used may a ec he adop ion and im-
pac s o GenAI by wo ke s.41 Acco dingly, we will con inue ou e o s o collec da a on i ms’ and
wo ke s’ adop ion o GenAI in de eloping coun ies, in o de o alida e he exposu e measu es
de eloped in his s udy and o adap ou policy esponses o mo e de ailed indings. In ha e-
ga d, we would welcome collabo a ion wi h ins i u ions, including na ional adminis a ions, ha
migh be able o assis us in he collec ion o such da a o u u e esea ch.
41 Figu e A5 shows a obus ness check whe e exposu e o GenAI is spli by whe he wo ke s ha e access o in e ne and a compu e
(ins ead o compu e only).
44 ILO Wo king Pape 121
Dependen Va iable: AI Exposu e
Augmen a ion Au oma ion Big Unknown
(0.00327) (0.00146) (0.00341)
Cons uc ion -0.00372** -0.00328*** 0.0222***
(0.00159) (0.000812) (0.00192)
Re ail, wholesale ade, es au an s, ho els, e c. 0.111*** -0.00104 0.329***
(0.00187) (0.000840) (0.00257)
Elec ., gas, wa e , ansp., communica ion 0.147*** 0.0227*** 0.0721***
(0.00288) (0.00153) (0.00274)
Banks, inance, insu ance, p o essional ss. 0.0693*** 0.0856*** 0.191***
(0.00260) (0.00250) (0.00353)
Public Adm., de ence 0.0571*** 0.0339*** 0.145***
(0.00301) (0.00257) (0.00424)
Educa ion, Heal h, pe sonal se ices 0.295*** -0.0161*** 0.0155***
(0.00298) (0.00127) (0.00259)
Domes ic ss. 0.0467*** -0.0422*** -0.0632***
(0.00217) (0.00116) (0.00230)
Cons an 0.0193*** -0.0118*** 0.0297***
(0.00304) (0.00150) (0.00431)
Coun y and yea ixed e ec s Yes Yes Yes
Mean o dependen a iable 0.1203 0.0323 0.1821
Obse a ions 888,685 888,685 888,685
R2 0.139 0.076 0.228

45 ILO Wo king Pape 121
Re e ences
Acemoglu, D., 2024. The Simple Mac oeconomics o AI. Mass. Ins . Technol.
Acemoglu, D., Ande son, G., Beede, D., Bu ing on, C., Child ess, E., Dinle soz, E., Fos e , L.,
Goldschlag, N., Hal iwange , J., K o , Z., Res epo, P., Zolas, N., 2023. Ad anced Technology
Adop ion: Selec ion o Causal E ec s? AEA Pap. P oc. 113, 210–214. h ps://doi.o g/10.1257/
pandp.20231037
Acemoglu, D., Ande son, G.W., Beede, D.N., Bu ing on, C., Child ess, E.E., Dinle soz, E., Fos e ,
L.S., Goldschlag, N., Hal iwange , J.C., K o , Z., Res epo, P., Zolas, N., 2024. Au oma ion and he
Wo k o ce: A Fi m-Le el View om he 2019 Annual Business Su ey. NBER Chap e s.
Acemoglu, D., Res epo, P., 2022. Tasks, Au oma ion, and he Rise in U.S. Wage Inequali y.
Econome ica 90, 1973–2016. h ps://doi.o g/10.3982/ECTA19815
Acemoglu, D., Res epo, P., 2018. A i icial In elligence, Au oma ion and Wo k (No. w24196). Na ional
Bu eau o Economic Resea ch. h ps://doi.o g/10.3386/w24196
Adams-P assl, J., Ab aha, H., Kelly-Ly h, A., Silbe man, M. ‘Six’, Rakshi a, S., 2023. Regula ing
algo i hmic managemen : A bluep in . Eu . Labou Law J. 14, 124–151. h ps://doi.
o g/10.1177/20319525231167299
Ananian, S., Aube , P., Behaghel, L., 2006. T a ailleu s âgés, nou elles echnologies e change-
men s o ganisa ionnels : un éexamen à pa i de l’enquê e « Reponse ». Sui i d’un commen ai e
de Luc Behaghel : emploi des senio s - Des e e s du changemen echnologique aux ecomman-
da ions. Econ. S a . 397, 21–49. h ps://doi.o g/10.3406/es a .2006.7125
Aube , P., Ca oli, E., Roge , M., 2006. New echnologies, o ganisa ion and age: i m-le el e i-
dence*. Econ. J. 116, F73–F93. h ps://doi.o g/10.1111/j.1468-0297.2006.01065.x
Au o , D., 2024. Applying AI o Rebuild Middle Class Jobs. Wo king Pape Se ies. h ps://doi.
o g/10.3386/w32140
Au o , D., Do n, D., Ka z, L., Pa e son, C., an Reenen, J., 2020. The Fall o he Labo Sha e and
he Rise o Supe s a Fi ms*. Q. J. Econ. 135, 645–709.
Au o , D.H., 2015. Why A e The e S ill So Many Jobs? The His o y and Fu u e o Wo kplace
Au oma ion. J. Econ. Pe spec . 29, 3–30. h ps://doi.o g/10.1257/jep.29.3.3
Au o , D.H., Ka z, L.F., Kea ney, M.S., 2008. T ends in U.S. Wage Inequali y: Re ising he Re isionis s.
Re . Econ. S a . 90, 300–323.
Aze edo, J.P.W.D., Inchaus e Comboni, M.G., San elice, V., 2013. Decomposing he ecen inequal-
i y decline in La in Ame ica. Wo ld Bank.
Babina, T., Fedyk, A., He, A., Hodson, J., 2024. A i icial in elligence, i m g ow h, and p oduc in-
no a ion. J. Financ. Econ. 151.
46 ILO Wo king Pape 121
Bakke , B.B., Ghazanchyan, M., Emme ling, J., Vibha, V., 2020. The Lack o Con e gence o La in-
Ame ica Compa ed wi h CESEE: Is Low In es men o Blame? [WWW Documen ]. IMF. URL h ps://
www.im .o g/en/Publica ions/WP/Issues/2020/06/19/The-Lack-o -Con e gence-o -La in-Ame ica-
Compa ed-wi h-CESEE-Is-Low-In es men - o-Blame-49519 (accessed 3.26.24).
B ambilla, I., Césa , A., Falcone, G., Gaspa ini, L., Lomba do, C., 2023. Rou iniza ion and Employmen :
E idence o La in Ame ica. Re . Desa o. Soc. 131–176. h ps://doi.o g/10.13043/DYS.95.4
B ynjol sson, E., Li, D., Raymond, L.R., 2023. Gene a i e AI a Wo k. Wo king Pape Se ies. h ps://
doi.o g/10.3386/w31161
B ynjol sson, E., Rock, D., Sy e son, C., 2021. The P oduc i i y J-Cu e: How In angibles Complemen
Gene al Pu pose Technologies. Am. Econ. J. Mac oecon. 13, 333–372. h ps://doi.o g/10.1257/
mac.20180386
Ca anza, E., Das, S., Ko ikula, A., 2023. Gende -Based Employmen Seg ega ion: Unde s anding
Causes And Policy In e en ions.
Caunedo, J., Kelle , E., Shin, Y., 2023. Technology and he Task Con en o Jobs ac oss he De elopmen
Spec um. Wo ld Bank Econ. Re . 37, 479–493. h ps://doi.o g/10.1093/wbe /lhad015
Cazzaniga, M., Jaumo e, F., Li, L., Melina, G., Augus us J., P., Ca lo, P., Emma J., R., Mendes, M., 2024.
Gen-AI: A i icial In elligence and he Fu u e o Wo k [WWW Documen ]. IMF. URL h ps://www.
im .o g/en/Publica ions/S a -Discussion-No es/Issues/2024/01/14/Gen-AI-A i icial-In elligence-
and- he-Fu u e-o -Wo k-542379 (accessed 2.19.24).
Chacal ana Janampa, J., de Ma os, F.B., Ga cía, J.M., 2024. New echnologies, e-go e nmen and
in o mali y | In e na ional Labou O ganiza ion. ILO Wo k. Pap. 112, ILO Wo king Pape s, Gene a.
Comunale, M., Mane a, A., 2024. The Economic Impac s and he Regula ion o AI: A Re iew o
he Academic Li e a u e and Policy Ac ions [WWW Documen ]. IMF. URL h ps://www.im .o g/en/
Publica ions/WP/Issues/2024/03/22/The-Economic-Impac s-and- he-Regula ion-o -AI-A-Re iew-
o - he-Academic-Li e a u e-and-546645 (accessed 3.26.24).
Das, M., Hilgens ock, B., 2022. The exposu e o ou iniza ion: Labo ma ke implica ions o de-
eloped and de eloping economies. S uc . Change Econ. Dyn. 60, 99–113.
Dieppe, A., 2021. Global P oduc i i y: T ends, D i e s, and Policies.
Doellgas , V., O’B ady, S., Kim, J., Wal e s, D., 2023. AI in Con ac Cen e s: A i icial In elligence
and Algo i hmic Managemen in F on line Se ice Wo kplaces.
Dono an, K., Lu, W.J., Schoellman, T., 2023. Labo Ma ke Dynamics and De elopmen *. Q. J. Econ.
138, 2287–2325. h ps://doi.o g/10.1093/qje/qjad019
Egana-delSol, P., Bus elo, M., Ripani, L., Sole , N., Viollaz, M., 2022. Au oma ion in La in Ame ica:
A e Women a Highe Risk o Losing Thei Jobs? Technol. Fo ecas . Soc. Change 175, 121333.
h ps://doi.o g/10.1016/j. ech o e.2021.121333
Eloundou, T., Manning, S., Mishkin, P., Rock, D., 2023. GPTs a e GPTs: An Ea ly Look a he Labo
Ma ke Impac Po en ial o La ge Language Models.
47 ILO Wo king Pape 121
E umban, A., Samaan, D., Van A k, B., 2024. La in Ame ica’s P oduc i i y Puzzle: Insigh s in o di-
e si ica ion and s uc u al change. P esea a ion a he CEPAL con e ence: h ps://www.cepal.
o g/si es/de aul / iles/e en s/ iles/daniel_samaan_p esen a ion.pd .
Fa e io, M., Tyson, A., 2022. Wha he da a says abou Ame icans’ iews o a i icial in elligence.
Pew Res. Cen . URL h ps://www.pew esea ch.o g/sho - eads/2023/11/21/wha - he-da a-says-
abou -ame icans- iews-o -a i icial-in elligence/ (accessed 3.14.24).
Fel en, E., Raj, M., Seamans, R., 2021. Occupa ional, indus y, and geog aphic exposu e o a i i-
cial in elligence: A no el da ase and i s po en ial uses. S a eg. Manag. J. 42, 2195–2217. h ps://
doi.o g/10.1002/smj.3286
Fel en, E.W., Raj, M., Seamans, R., 2023a. Occupa ional He e ogenei y in Exposu e o Gene a i e
AI. h ps://doi.o g/10.2139/ss n.4414065
Fel en, E.W., Raj, M., Seamans, R., 2023b. How will Language Modele s like Cha GPT A ec
Occupa ions and Indus ies? SSRN Elec on. J. h ps://doi.o g/10.2139/ss n.4375268
Fel en, E.W., Raj, M., Seamans, R., 2018. A Me hod o Link Ad ances in A i icial In elligence o
Occupa ional Abili ies. AEA Pap. P oc. 108, 54–57. h ps://doi.o g/10.1257/pandp.20181021
F ey, C.B., Osbo ne, M.A., 2017. The u u e o employmen : How suscep ible a e jobs o compu e isa-
ion? Technol. Fo ecas . Soc. Change 114, 254–280. h ps://doi.o g/10.1016/j. ech o e.2016.08.019
Ga o e, S.D., Gomez, P.N., Ozden, C., Rijke s, B., Viollaz, M., Winkle , H., 2021. Who on Ea h Can
Wo k om Home? Wo ld Bank Res. Obs.
Gmy ek, P., Be g, J., Bescond, D., 2023. Gene a i e AI and Jobs: A global analysis o po en ial e -
ec s on job quan i y and quali y (Wo king pape ).
Goldman Sachs, 2023. Global Economics Analys The Po en ially La ge E ec s o A i icial
In elligence on Economic G ow h (B iggsKodnani), Global Economics Analys .
G ampp, M., B andes, D., Laude, D., 2023. Gene a i e AI’s as and u ious en y in o Swi ze land
(Delloi e).
Gu ié ez, G., Philippon, T., 2017. Declining Compe i ion and In es men in he U.S. NBER Wo k.
Pap., NBER Wo king Pape s.
Hui, X., Reshe , O., Zhou, L., 2023. The Sho -Te m E ec s o Gene a i e A i icial In elligence on
Employmen : E idence om an Online Labo Ma ke . h ps://doi.o g/10.2139/ss n.4527336
ILO, 2023a. ILO Modelled Es ima es (ILOEST da abase) [WWW Documen ]. ILOSTAT. URL h ps://
ilos a .ilo.o g/ esou ces/concep s-and-de ini ions/ilo-modelled-es ima es/ (accessed 6.20.23).
ILO, 2023b. ISCO documen a ion: pa III. De ini ions o majo g oups, sub-majo g oups, mino
g oups and uni g oups. [WWW Documen ]. URL h ps://www.ilo.o g/public/english/bu eau/s a /
isco/docs/g oupde n08.pd (accessed 6.20.23).
ILO, 2023c. Wo ld Employmen and Social Ou look 2023: The alue o essen ial wo k [WWW
Documen ]. URL h ps://www.ilo.o g/digi alguides/en-gb/s o y/weso2023-key-wo ke s (accessed
7.31.23).
48 ILO Wo king Pape 121
ILO, 2023d. Wo ld Social P o ec ion Repo 2020–22: Social p o ec ion a he c oss oads – in pu -
sui o a be e u u e. ILO Flaghip Repo .
ILO, 2023e. The ILO s a egy on skills and li elong lea ning 2030 | In e na ional Labou O ganiza ion.
IMF, 2022. P oduc i i y in La in Ame ica and he Ca ibbean: Recen T ends and he COVID-19 Shock.
Lane, M., Williams, M., B oecke, S., 2023. The impac o AI on he wo kplace: Main indings om
he OECD AI su eys o employe s and wo ke s. OECD, Pa is. h ps://doi.o g/10.1787/ea0a0 e1-en
Lewandowski, P., Pa k, A., Ha dy, W., Du, Y., 2019. Technology, Skills, and Globaliza ion: Explaining
In e na ional Di e ences in Rou ine and Non ou ine Wo k Using Su ey Da a.
Lo Bello, S., Sanchez Pue a, M.L., Winkle , H.J., 2019. F om Ghana o Ame ica : The Skill Con en
o Jobs and Economic De elopmen . Policy Res. Wo k. Pap. Se ., Policy Resea ch Wo king Pape
Se ies.
Maloney, W.F., Molina, C.A., 2016. A e au oma ion and ade pola izing de eloping coun y labo
ma ke s, oo ? Policy Res. Wo k. Pap. Se ., Policy Resea ch Wo king Pape Se ies.
McKinsey, 2023. Economic po en ial o gene a i e AI | McKinsey.
Messina, J., Pica, G., O iedo, A., 2016. Job Pola iza ion in La in Ame ica. In e -Ame ican De elopmen
Bank, Washing on, DC.
Messina, J., Sil a, J., 2018. Wage Inequali y in La in Ame ica. Wo ld Bank Publ. - Books.
Noy, S., Zhang, W., 2023. Expe imen al e idence on he p oduc i i y e ec s o gene a i e a i icial
in elligence. Science 381, 187–192. h ps://doi.o g/10.1126/science.adh2586
Nu ski, L., Rue , N., 2024. Exposu e o gene a i e AI in he Eu opean labou ma ke : dis ibu ion-
al implica ions.
OECD, 2023. The impac o AI on he wo kplace: Main indings om he OECD AI su eys o
employe s and wo ke s (No. 288), OECD A i icial In elligence Pape s. h ps://doi.o g/10.1787/
ea0a0 e1-en
OECD/PIAAC, 2019. Skills Ma e : Addi ional Resul s om he Su ey o Adul Skills. O ganisa ion
o Economic Co-ope a ion and De elopmen , Pa is.
Ozdene on, Hakki, 2023. AI-Exposed Jobs Employ Mo e Women. URL h ps://www. e eliolabs.
com/news/mac o/ai-exposed-jobs-employ-mo e-women/ (accessed 6.26.23).
Peng, S., Kalliam akou, E., Cihon, P., Demi e , M., 2023. The Impac o AI on De elope P oduc i i y:
E idence om Gi Hub Copilo . h ps://doi.o g/10.48550/a Xi .2302.06590
Pew Resea ch Cen e , 2023. Ame icans explain in hei own wo ds wha makes hem ei he mo e
conce ned o mo e exci ed abou he inc eased p esence o AI in daily li e. In e ne Sci. Tech.
URL h ps://www.pew esea ch.o g/in e ne /2022/03/17/ai-and-human-enhancemen -ame icans-
openness-is- empe ed-by-a- ange-o -conce ns/ps_2022-03-17_ai-he_01-02/ (accessed 3.14.24).
49 ILO Wo king Pape 121
P y ko a, E., Pe i , F., Deyu, L., Suga , C., Tommaso, C., 2024. The Employmen Impac o Eme ging
Digi al Technologies.
Res epo, P., 2023. Au oma ion: Theo y, E idence, and Ou look. Wo king Pape Se ies. h ps://
doi.o g/10.3386/w31910
Ru ge s, 2024. U.S. Wo ke s Assess he Impac s o A i icial In elligence on Jobs: Topline Su ey
Resul s, Held ich Cen e o Wo k o ce De elopmen .
Schady, N., Holla, A., Saba wal, S., Sil a, J., Yi, C., 2023. Collapse and Reco e y: How he COVID-19
Pandemic E oded Human Capi al and Wha o Do abou I .
S anbe g, M., Li, W., Fleming, M., Goeh ing, B., Thompson, N., 2024. Beyond AI Exposu e: Which
Tasks a e Cos -E ec i e o Au oma e wi h Compu e Vision? h ps://doi.o g/10.2139/ss n.4700751
Webb, M., 2020. The Impac o A ificial In elligence on he Labo Ma ke .
WEF, 2023. The Fu u e o Jobs Repo 2023.
Wo ld Bank, 2016a. Po e y and Sha ed P ospe i y 2016: Taking on Inequali y, Po e y and Sha ed
P ospe i y. The Wo ld Bank. h ps://doi.o g/10.1596/978-1-4648-0958-3
Wo ld Bank, 2016b. Wo ld De elopmen Repo 2016: Digi al Di idends (Tex /HTML).

50 ILO Wo king Pape 121
Acknowledgemen s
We would like o hank Richa d Samans, Se gei Sua ez Dillon Soa es, Janine Be g, Leona do
Iaco one, Paulo Bas os, Ca los Rod iguez Cas elan, William F. Maloney and s a om ILO Resea ch
and om he Wo ld Bank Po e y eam o aluable commen s. We also hank ILO STATISTICS
eam o da a con ibu ion, pa icula ly Da id Bescond, who p oduced se e al inpu s o ou es-
ima es based on ILO mic o da a eposi o y, as acknowledged in de ail in he pape .
Resea ch Depa men (RESEARCH)
In e na ional Labou O ganiza ion
Rou e des Mo illons 4
1211 Gene a 22
Swi ze land
T +41 22 799 6530
esea [email p o ec ed]
www.ilo.o g/ esea ch
XAd ancing social jus ice, p omo ing decen wo k
The In e na ional Labou O ganiza ion is he Uni ed Na ions agency o he wo ld o wo k. We b ing oge he go e nmen s, employe s and wo ke s
o imp o e he wo king li es o all people, d i ing a human-cen ed app oach o he u u e o wo k h ough employmen c ea ion, igh s a wo k,
social p o ec ion and social dialogue.
I S B N 9789220410028
9HSTCMA*ebaaci+
9 789220 410028
Con ac de ails