Teh Impact of AI On The Labour Market

Organisation for Economic Co-operation and Development
DELSA/ELSA/WD/SEM(2021)3
For Official Use English - Or. English
21 January 2021
DIRECTORATE FOR EMPLOYMENT, LABOUR AND SOCIAL AFFAIRS
EMPLOYMENT, LABOUR AND SOCIAL AFFAIRS COMMITTEE
Cancels & replaces the same document of 12 January 2021
The impact of Artificial Intelligence on the labour market: What do we know so

far?
JEL Codes: J20, J81, J31, O14, O33.
Authorised for publication by Stefano Scarpetta, Director, Directorate for Employment, Labour and
Social Affairs.
Marguerita Lane, [email protected]
Anne Saint-Martin, [email protected]
JT03470411
OFDE
This document, as well as any data and map included herein, are without prejudice to the status of or sovereignty over any territory,
to the delimitation of international frontiers and boundaries and to the name of any territory, city or area.
2  DELSA/ELSA/WD/SEM(2021)3
OECD Social, Employment and Migration Working Papers
www.oecd.org/els/workingpapers
OECD Working Papers should not be reported as representing the official views of the
OECD or of its member countries. The opinions expressed and arguments employed are
those of the author(s).
Working Papers describe preliminary results or research in progress by the author(s) and
are published to stimulate discussion on a broad range of issues on which the OECD works.
Comments on Working Papers are welcomed, and may be sent to [email protected].
This series is designed to make available to a wider readership selected labour market,
social policy and migration studies prepared for use within the OECD. Authorship is
usually collective, but principal writers are named. The papers are generally available only
in their original language – English or French – with a summary in the other.
This document and any map included herein are without prejudice to the status of or
sovereignty over any territory, to the delimitation of international frontiers and boundaries
and to the name of any territory, city or area.
© OECD 2021
You can copy, download or print OECD content for your own use, and you can include
excerpts from OECD publications, databases and multimedia products in your own
documents, presentations, blogs, websites and teaching materials, provided that suitable
acknowledgment of OECD as source and copyright owner is given. All requests for
commercial use and translation rights should be submitted to [email protected].
For Official Use

DELSA/ELSA/WD/SEM(2021)3 3
Acknowledgements
This publication contributes to the OECD’s Artificial Intelligence in Work, Innovation,

Productivity and Skills (AI-WIPS) programme, which provides policymakers with new
evidence and analysis to keep abreast of the fast-evolving changes in AI capabilities and
diffusion and their implications for the world of work. The programme aims to help ensure
that adoption of AI in the world of work is effective, beneficial to all, people-centred and
accepted by the population at large. AI-WIPS is supported by the German Federal Ministry
of Labour and Social Affairs (BMAS) and will complement the work of the German AI
Observatory in the Ministry’s Policy Lab Digital, Work & Society. For more information,
visit https://oecd.ai/work-innovation-productivity-skills and https://denkfabrik-bmas.de/.
Special thanks must go to Stijn Broecke for his supervision of the project and to Mark
Keese for his guidance and support throughout the project. The report also benefitted from
helpful comments provided by colleagues from the Directorate for Employment, Labour
and Social Affairs (Glenda Quintini, Sandrine Cazes, Angelica Salvi Del Pero, Ann
Vourc’h, Anna Milanez, Alexandre Georgieff and Morgan Williams), from the Directorate
for Science, Technology and Innovation (Karine Perset and Luis Aranda) and from the
Centre for Skills (Elena Crivellaro). Thanks also to Diana Gehlhaus at the Centre for
Security and Emerging Technology at Georgetown University for her valuable comments.
Thanks to Katerina Kodlova for providing publication support.
For Official Use

Abstract
Recent developments in Artificial Intelligence (AI) have stoked new fears about large-scale
job loss, stemming from its ability to automate a rapidly expanding set of tasks (including
non-routine cognitive tasks), and its potential to affect every sector of the economy.
Furthermore, there are concerns about employee well-being and the broader work
environment, linked to the idea that AI may soon become pervasive in the workplace and
threaten and undermine humans’ place in it. However, AI also has the potential to
complement and augment human capabilities, leading to higher productivity, greater
demand for human labour and improved job quality.
From a theoretical perspective, the impact of AI on employment and wages is ambiguous,
and it may depend strongly on the type of AI being developed and deployed, how it is
developed and deployed, and on market conditions and policy. However, the empirical
evidence based on AI adopted in the last 10 years does not support the idea of an overall
decline in employment and wages in occupations exposed to AI. While AI is capable of
performing some non-routine cognitive tasks, some bottlenecks to adoption still remain,
and many tasks still require humans to carry them out. Thus, much of the impact of AI on
jobs is likely to be experienced through the reorganisation of tasks within an occupation.
Certain groups of workers may be more capable or better positioned to take advantage of
the benefits that AI brings, use AI in a way that is complementary to their work, and avoid
its negative impacts.
AI is likely to reshape the work environment of many people, by changing the content and
design of their jobs, the way workers interact with each other and with machines, and how
work effort and efficiency are monitored. AI can play an important role in facilitating
human-machine collaboration, helping workers in the execution of tedious or physically
demanding tasks while allowing them to leverage their own uniquely human abilities.
However, the same AI applications could also entail significant risks for the work
environment, especially if applied badly or with the singular motivation to cut costs.
For Official Use

Synthèse
Les récents progrès de l’intelligence artificielle (IA) ont ravivé les craintes de destructions
massives d’emplois, craintes dont l’origine tient au fait que cette technologie permet
l’automatisation d’un ensemble de tâches qui s’élargit rapidement, et qu’elle est
susceptible de trouver des applications dans tous les secteurs de l’économie. En outre, le
bien-être des salariés et l’environnement de travail dans son ensemble font aussi l’objet de
préoccupations, liées à l’idée que l’IA pourrait devenir bientôt omniprésente dans
l’entreprise et fragiliser la place de l’humain en son sein. Pourtant, l’IA est aussi à même
de compléter et d’augmenter les capacités humaines et, partant, de susciter des gains de
productivité, de soutenir la demande de main-d’œuvre et d’accroître la qualité des emplois.
Du point de vue théorique, les effets de l’IA sur l’emploi et les salaires sont ambivalents et
dépendent probablement dans une large mesure de la nature des technologies en question,
de la manière dont elles sont développées et déployées, des conditions de marché et des
politiques en place. Quoi qu’il en soit, les observations empiriques portant sur l’adoption
de telles technologies au cours de ces 10 dernières années n’accréditent pas la thèse d’un
recul global de l’emploi et des salaires dans les professions exposées à l’IA. Les systèmes
d’IA sont certes capables d’exécuter un certain nombre de tâches cognitives non
répétitives, cependant il demeure quelques freins à leur adoption, et de nombreuses tâches
exigent encore l’intervention d’opérateurs humains pour être menées à bien. Il s’ensuit que
l’essentiel des répercussions de l’IA sur l’emploi se matérialisera sans doute à travers la
réorganisation des tâches relevant d’une profession donnée. Certaines catégories de
travailleurs pourraient être mieux préparées ou mieux positionnées pour tirer le meilleur
parti de l’IA, en utilisant les nouveaux outils qu’elle apporte de façon complémentaire à
leur activité professionnelle, évitant ainsi les effets négatifs de cette technologie.
L’IA va vraisemblablement remodeler l’environnement de travail de beaucoup en
modifiant la teneur de leur emploi et la manière de le concevoir, leurs interactions avec
leurs semblables et avec les machines, et la façon dont l’effort et l’efficacité au travail
seront observés. L’IA peut jouer un rôle important en facilitant la collaboration entre
humains et machines, en aidant les travailleurs dans l’exécution de tâches fastidieuses ou
physiquement éprouvantes, tout en leur permettant d’exploiter au mieux leurs compétences
uniques, spécifiques à l’humain. Cela étant, ces mêmes applications de l’IA peuvent aussi
faire planer des risques non négligeables sur l’environnement de travail, surtout si elles
sont mises en œuvre de manière inappropriée ou à seule fin de réduire les coûts.
For Official Use

Kurzfassung
Die jüngsten Entwicklungen im Bereich der künstlichen Intelligenz (KI) haben die Furcht
vor weitreichenden Arbeitsplatzverlusten durch KI erneut angefacht. Grund dafür ist, dass
immer mehr Tätigkeiten durch KI automatisiert werden können und jeder Wirtschaftssektor
davon betroffen sein könnte. Darüber hinaus werden negative Auswirkungen auf das
Wohlergehen der Arbeitskräfte und das allgemeine Arbeitsumfeld befürchtet. Dahinter
steht der Gedanke, dass KI am Arbeitsplatz bald allgegenwärtig sein könnte und die
menschliche Arbeitskraft dadurch an Bedeutung verlieren und verdrängt werden könnte.
KI bietet aber auch das Potenzial, menschliche Fähigkeiten zu ergänzen und zu steigern.
Die Folge sind höhere Produktivität, größere Nachfrage nach menschlicher
Arbeitsleistung und bessere Arbeitsplatzqualität.
Aus theoretischer Sicht ist der Effekt der künstlichen Intelligenz auf die
Beschäftigungs- und Lohnentwicklung uneindeutig. Er kann stark davon abhängen, um
welche Art von KI es sich handelt, wie sie entwickelt und eingesetzt wird, und wie die
Marktbedingungen und das Politikumfeld aussehen. Die empirischen Befunde auf Basis
der KI-Nutzung der letzten zehn Jahre lassen in Berufen, in denen KI besonders gut
eingesetzt werden kann, jedoch keinen generellen Beschäftigungs- und Lohnrückgang
erkennen. Obwohl KI einige nichtroutinemäßige kognitive Aufgaben übernehmen kann, ist
dies nach wie vor nicht in allen Bereichen möglich und für viele Tätigkeiten sind weiterhin
Menschen erforderlich. Der Effekt der künstlichen Intelligenz auf die Arbeitswelt dürfte
daher vor allem in einer Neuorganisation der Aufgaben bestehen, aus denen sich die
jeweilige berufliche Tätigkeit zusammensetzt. Bestimmte Gruppen von Arbeitskräften sind
möglicherweise besser in der Lage, künstliche Intelligenz zu ihrem Vorteil zu nutzen, sie
zur Ergänzung ihrer eigenen Arbeitsleistung einzusetzen und ihre negativen Folgen zu
vermeiden.
KI dürfte das Arbeitsumfeld vieler Menschen erheblich verändern. Dies betrifft u. a. die
Arbeitsinhalte und die Arbeitsgestaltung, die Interaktion der Arbeitskräfte untereinander
und mit Maschinen und die Methoden, mit denen Arbeitsleistung und Effizienz gemessen
werden. KI kann einen wichtigen Beitrag zur Verbesserung der
Mensch-Maschine-Interaktion leisten, indem die Arbeitskräfte von eintönigen oder
körperlich anstrengenden Tätigkeiten entlastet werden und stattdessen ihre menschlichen
Fähigkeiten stärker einbringen können. Von denselben KI-Anwendungen könnten jedoch
auch erhebliche Risiken für das Arbeitsumfeld ausgehen, insbesondere wenn sie
unsach-gemäß genutzt oder ausschließlich zur Kostensenkung eingesetzt werden.
For Official Use

Table of Contents
OECD Social, Employment and Migration Working Papers ............................................................................ 2

Acknowledgements ................................................................................................................................................ 3
Abstract .................................................................................................................................................................. 4
Synthèse .................................................................................................................................................................. 5
Kurzfassung ........................................................................................................................................................... 6
Executive summary ............................................................................................................................................... 9
Résumé ................................................................................................................................................................. 11
Zusammenfassung ............................................................................................................................................... 14
1. Introduction ..................................................................................................................................................... 17
2. What can AI do? .............................................................................................................................................. 19
2.1. AI can be considered a general purpose technology ................................................................................... 19
2.2. AI can be considered an automation technology ........................................................................................ 20
2.3. Certain occupations are more exposed to AI .............................................................................................. 22
2.4. AI will not only lead to automation, but will also complement labour ...................................................... 25
3. Impact of AI on productivity, employment and wages ................................................................................ 27
3.1. AI will increase productivity but the size of the impact is debated ............................................................ 27
3.2. Theoretical models are ambiguous on the impact of AI on employment and wages .................................. 29
3.3. The limited empirical evidence does not support the idea that AI has reduced employment and wages ... 32
3.4. Firms are divided in their expectations about AI’s impact on overall labour demand ............................... 34
4. The AI transition ............................................................................................................................................. 35
4.1. Much of the impact of AI will be seen through the reorganisation of tasks ............................................... 35
4.2. The impact of AI will also result in the creation of new tasks ................................................................... 36
4.3. Workers may need to re-skill or up-skill in order to adapt to AI-induced changes in the labour market ... 38
4.4. Certain workers will be more capable of adapting to change ..................................................................... 40
4.5. Does AI favour owners of capital and superstar firms? .............................................................................. 41
5. AI and the work environment ........................................................................................................................ 43
5.1. The jury is out on whether the reorganisation of tasks as a result of AI adoption improves the work
environment ....................................................................................................................................................... 44
5.2. AI can promote close human-robot collaboration ...................................................................................... 45
5.3. AI can support human resource management and career development ...................................................... 46
5.4. AI may also entail risks for the work environment .................................................................................... 48
References ............................................................................................................................................................ 52
Tables
Table 2.1. Occupations most and least exposed to AI ........................................................................................... 23
Table 3.1. Estimated economic potential of AI ..................................................................................................... 29
Table 5.1. AI applications in Human Resource Management (HRM) and career development ............................ 47
For Official Use

Boxes
Box 2.1. Recent trends in AI and its adoption in the workplace ........................................................................... 20
Box 2.2. AI and the singularity ............................................................................................................................. 21
Box 3.1. Estimates of the economic potential of AI .............................................................................................. 29
For Official Use

Executive summary
Recent developments in Artificial Intelligence (AI) have stoked new fears about large-scale
job loss, stemming from its ability to automate a rapidly expanding set of tasks (including
non-routine cognitive tasks), and its potential to affect every sector of the economy.
Furthermore, there are concerns about employee well-being and the broader work
environment, linked to the idea that AI may soon become pervasive in the workplace and
threaten and undermine humans’ place in it. However, AI also has the potential to
complement and augment human capabilities, leading to higher productivity, greater
demand for human labour and improved job quality.
From a theoretical perspective, the impact of AI on employment and wages is ambiguous,
and it may depend strongly on the type of AI being developed and deployed, how it is
developed and deployed, and on market conditions and policy. If AI facilitates the
automation of tasks and delivers only modest increases in productivity, workers are
unlikely to share in the benefits of this new technology. To produce positive outcomes for
workers, AI must create new high-productivity tasks to replace those automated and boost
productivity sufficiently to raise consumer demand, hence increasing demand for human
labour.
The empirical evidence based on AI adopted in the last 10 years does not support the idea
of an overall decline in employment and wages in occupations exposed to AI. Some studies
suggest a positive impact of AI on wage growth.
The occupations judged to be most exposed to AI include high-skilled occupations
involving non-routine cognitive tasks, such as lab technicians, engineers and actuaries.
However, high exposure does not necessarily mean that jobs in these occupations will
disappear. While AI’s capabilities have expanded substantially, some bottlenecks to
adoption still remain, and many tasks still require humans to carry them out. Thus, much
of the impact of AI on jobs is likely to be experienced through the reorganisation of tasks
within an occupation, with some workers ultimately complemented in their work by AI,
rather than substituted by it.
Workers may need to re-skill or up-skill in order to adapt to the reorganisation of tasks and
the emergence of new tasks, and to weather potential job loss and navigate transitions to
new jobs. This will not only mean acquiring AI-related skills, but also acquiring skills in
areas that AI cannot perform so well, such as creative and social intelligence, reasoning
skills, and dealing with uncertainty. The smoothness of the AI transition and the extent of
the impact on workers will also depend on firm-level incentives to retain and retrain staff
and on institutional factors, such as the general infrastructure for training and job-search
available in the country, direct government funding, tax incentives and social benefit
systems.
Certain groups of workers may be more capable or better positioned to take advantage of
the benefits that AI brings, use AI in a way that is complementary to their work, and avoid
its negative impacts. While some high-skilled occupations are among those most exposed
to AI, there is evidence that individuals in higher wage occupations and/or with higher
educational attainment experience higher wage growth linked to AI, suggesting some
degree of complementarity. This suggests that AI adoption could increase income
inequality.
For Official Use

Similarly, some firms may be better placed than others to develop and/or deploy AI.
Moreover, if the gains of AI accrue to a small number of superstar innovators or firms with
excessive market power, this could produce a divide between innovators and workers and
further reinforce the potentially negative impact of AI on inequality.
AI is likely to reshape the work environment of many people, by changing the content and
design of their jobs, the way workers interact with each other and with machines, and how
work effort and efficiency are monitored. AI can play an important role in facilitating
human-machine collaboration, helping workers in the execution of tedious or physically
demanding tasks while allowing them to leverage their own uniquely human abilities. AI
can offer cheaper, faster and more scalable solutions in the field of human resource
management, enabling workers to advance their own careers, helping managers to manage,
and enhancing training.
However, the same AI applications could also entail significant risks for the work
environment, especially if applied badly or with the singular motivation to cut costs. A lack
of transparency and explainability around algorithmic predictions and decisions can make
employees feel insecure, either psychologically or physically. By enabling extensive
monitoring of workers’ performance, AI can increase work pressure and generate stress
about productivity and about how managers may interpret data.
Many questions remain for future research. Surveys and qualitative research may be useful
for understanding how firms and workers view the AI transition, how decisions are made
and under what management models and what national policy and institutions, and what
measures lead to positive outcomes. Of particular interest is capturing evidence from
different environments, including for instance the use of AI in facilitating close
human-machine collaboration in manufacturing environments, the role of AI assisting the
highly skilled in prediction tasks, and the use of AI-enabled career development tools.
Further empirical analysis will help establish to what extent the impact of AI resembles the
impact of previous waves of automation, in terms of its potential to substitute and/or
complement human labour and to create new tasks, and the implications for labour demand
and income inequality. This will rely on data collection and the creation of indicators that
capture AI and its inherent attributes (in addition to considering automation technologies
more generally).
For Official Use

DELSA/ELSA/WD/SEM(2021)3  11
Résumé
Les récents progrès de l’intelligence artificielle (IA) ont ravivé les craintes de destructions
massives d’emplois, craintes dont l’origine tient au fait que cette technologie permet
l’automatisation d’un ensemble de tâches qui s’élargit rapidement (et s’étend notamment à
des tâches cognitives non répétitives), et qu’elle est susceptible de trouver des applications
dans tous les secteurs de l’économie. En outre, le bien-être des salariés et l’environnement
de travail dans son ensemble sont aussi l’objet de préoccupations à l’idée qu’elle pourrait
devenir bientôt omniprésente dans l’entreprise et y menacer et compromettre la place des
intervenants humains. Pourtant, l’IA est aussi à même de compléter et d’augmenter les
capacités humaines et, partant, de susciter des gains de productivité, de soutenir la demande
de main-d’œuvre et d’accroître la qualité des emplois.
Du point de vue théorique, ses effets sur l’emploi et les salaires sont ambivalents et
dépendent probablement dans une large mesure de la nature de ses applications, de la
manière dont elles sont mises au point et déployées, de la situation du marché et du cadre
réglementaire en place. Si cette nouvelle technologie facilite l’automatisation des tâches et
n’apporte que des gains modestes sur le plan de la productivité, alors les travailleurs ont
peu chances d’en recevoir eux aussi les bienfaits. Pour leur être bénéfique, l’IA doit faire
émerger de nouvelles tâches fortement productives, en remplacement de celles qui auront
été automatisées, et doper la productivité dans une mesure suffisante pour tirer la demande
des consommateurs, et par là même la demande de main-d’œuvre.
Les données concrètes recueillies au sujet des systèmes adoptés au cours de ces
10 dernières années n’accréditent pas la thèse d’un recul global de l’emploi et des salaires
dans les professions exposées à l’IA. Il semblerait, d’après certaines études, que celle-ci ait
une influence positive sur la croissance des salaires.
On trouve, parmi les professions considérées comme les plus exposées à l’IA, des
professions très qualifiées impliquant l’exécution de tâches cognitives non répétitives, ainsi
celles de technicien de laboratoire, d’ingénieur et d’actuaire. Cela étant, forte exposition ne
rime pas nécessairement avec destruction d’emplois. En dépit des progrès substantiels de
l’IA, il demeure quelques freins à son adoption, et de nombreuses tâches exigent encore
l’intervention d’opérateurs humains pour être menées à bien. Il s’ensuit que l’essentiel des
répercussions de l’IA sur l’emploi se matérialisera sans doute à travers la réorganisation
des tâches relevant d’une profession donnée, de sorte qu’in fine, cette technologie viendra
compléter l’activité de certains travailleurs et non pas se substituer à eux.
Les travailleurs auront sans doute besoin de recycler ou de développer leurs compétences
pour s’adapter à cette réorganisation et à l’apparition de tâches nouvelles, ainsi que pour
surmonter une éventuelle perte d’emploi et se reconvertir dans l’exercice d’une activité
nouvelle. Cela supposera d’acquérir des compétences non seulement dans le domaine de
l’IA, mais aussi dans des domaines où celle-ci n’est pas en mesure de rivaliser avec
l’humain, tels ceux de la créativité et de l’intelligence sociale, du raisonnement et de la
gestion de l’incertitude. La souplesse de la transition vers l’IA et l’ampleur de ses
conséquences pour les travailleurs dépendront aussi des incitations faites aux entreprises
pour qu’elles conservent et reconvertissent leur personnel et de différents facteurs
institutionnels, comme l’infrastructure générale de formation et de recherche d’emploi en
place dans le pays, les financements publics directs, les incitations fiscales et les systèmes
de prestations sociales.
For Official Use

Certaines catégories de travailleurs seront peut-être mieux à même, en raison de leurs

capacités ou de leur situation, de tirer parti des avantages apportés par l’IA, de se servir des
systèmes d’IA comme d’un auxiliaire dans leur activité professionnelle et de s’affranchir
des effets négatifs de cette technologie. En dépit du fait que quelques professions très
qualifiées comptent parmi les plus menacées, il apparaît que c’est à ceux qui occupent un
emploi relativement bien rémunéré et/ou ont un niveau d’études élevé que l’IA procure la
plus forte progression salariale, ce qui laisse supposer une certaine complémentarité. Il est
dès lors permis de penser que la diffusion des applications de l’IA pourrait contribuer au
creusement des inégalités de revenu.
De même, certaines entreprises seront probablement mieux à même que d’autres de mettre
au point et/ou déployer des systèmes d’IA. En outre, si l’IA profite exclusivement à un petit
nombre d’innovateurs « superstars » ou d’entreprises disposant d’une position de force
excessive sur le marché, un fossé pourrait se creuser entre ceux qui innovent et ceux qui
travaillent, et les possibles conséquences négatives sur les inégalités s’aggraver.
L’IA va vraisemblablement remodeler l’environnement de travail de beaucoup en
modifiant la teneur de leur emploi et la manière de le concevoir, leurs interactions avec
leurs semblables et avec les machines, et la mesure de l’effort de travail et de l’efficience
professionnelle. L’IA peut faciliter grandement la collaboration entre l’homme et la
machine, de même que l’exécution de tâches fastidieuses ou physiquement difficiles, et
permettre dans le même temps aux travailleurs d’exploiter les aptitudes spécifiquement
humaines dont ils sont dotés. Elle peut aussi apporter, dans le domaine des ressources
humaines, des solutions moins onéreuses, plus rapides et plus facilement reproductibles à
grande échelle, permettant aux travailleurs de gérer leur évolution professionnelle, aidant
les responsables à remplir leur rôle et favorisant la formation.
Cela étant, ces mêmes applications de l’IA peuvent aussi faire planer des risques non
négligeables sur l’environnement de travail, surtout si elles sont mises en œuvre de manière
inappropriée ou à seule fin de réduire les coûts. Le manque de transparence et
d’explicabilité des prédictions et décisions algorithmiques peut susciter chez les salariés un
sentiment d’insécurité, aussi bien psychologique que physique. Du fait qu’elle permet un
suivi complet de la performance des travailleurs, l’IA est susceptible d’accentuer les
pressions professionnelles et d’être cause d’anxiété au sujet de la productivité et de
l’interprétation que les dirigeants pourront faire des données à leur disposition.
De nombreuses questions demeurent, auxquelles la recherche devra apporter une réponse.
Des enquêtes et des études qualitatives pourraient aider à comprendre comment les
entreprises et les travailleurs envisagent la transition vers l’IA, comment les décisions sont
prises, en vertu de quels modèles de gestion, dans quel cadre réglementaire et sous l’égide
de quelles institutions nationales, et quelles mesures donnent des résultats positifs. Il serait
particulièrement intéressant de réunir des éléments factuels issus de différents
environnements, par exemple sur l’utilisation de l’IA au service d’une collaboration plus
étroite entre l’homme et la machine dans les activités manufacturières, sur l’aide apportée
aux travailleurs hautement qualifiés pour l’établissement de prévisions, et sur l’utilisation
d’outils de développement professionnel fondés sur cette technologie.
De nouvelles analyses empiriques aideront à voir dans quelle mesure le déploiement de
l’IA s’apparente, dans ses répercussions, à celui de vagues antérieures d’automatisation, du
point de vue de sa capacité de se substituer au travail humain et/ou de le compléter et de
susciter des tâches nouvelles et de ses conséquences sur la demande de main-d’œuvre et
les inégalités de revenu. Ces analyses prendront appui sur la collecte de données et la
For Official Use

construction d’indicateurs sur l’IA et ses caractéristiques propres (en plus de la prise en
compte des technologies qui, de manière plus générale, rendent possible l’automatisation).
For Official Use

Zusammenfassung
Die jüngsten Entwicklungen im Bereich der künstlichen Intelligenz (KI) haben die Furcht
vor weitreichenden Arbeitsplatzverlusten durch KI erneut angefacht. Grund dafür ist, dass
immer mehr Tätigkeiten (auch nichtroutinemäßige kognitive Aufgaben) durch KI
automatisiert werden können und jeder Wirtschaftssektor davon betroffen sein könnte.
Darüber hinaus werden negative Auswirkungen auf das Wohlergehen der Arbeitskräfte und
das allgemeine Arbeitsumfeld befürchtet. Dahinter steht der Gedanke, dass KI am
Arbeitsplatz bald allgegenwärtig sein könnte und die menschliche Arbeitskraft dadurch an
Bedeutung verlieren und verdrängt werden könnte. KI bietet aber auch das Potenzial,
menschliche Fähigkeiten zu ergänzen und zu steigern. Die Folge sind höhere Produktivität,
größere Nachfrage nach menschlicher Arbeitsleistung und bessere Arbeitsplatzqualität.
Aus theoretischer Sicht ist der Effekt der künstlichen Intelligenz auf die
Beschäftigungs- und Lohnentwicklung uneindeutig. Er kann stark davon abhängen, um
welche Art von KI es sich handelt, wie sie entwickelt und eingesetzt wird, und wie die
Marktbedingungen und das Politikumfeld aussehen. Wenn KI die Automatisierung von
Aufgaben erleichtert und nur geringfügige Produktivitätssteigerungen bewirkt, ist nicht
davon auszugehen, dass auch die Arbeitskräfte von dieser neuen Technologie profitieren.
Positive Effekte für die Arbeitskräfte ergeben sich dann, wenn durch KI neue
hochproduktive Tätigkeiten als Ersatz für die automatisierten Aufgaben entstehen und die
Produktivitätssteigerungen so hoch sind, dass die Verbrauchernachfrage wächst und
dadurch die Nachfrage nach menschlicher Arbeitsleistung angekurbelt wird.
Die empirischen Befunde auf Basis der KI-Nutzung der letzten zehn Jahre lassen keinen
generellen Beschäftigungs- und Lohnrückgang in Berufen erkennen, in denen KI besonders
gut eingesetzt werden kann. Einige Studien zeigen einen positiven Effekt der KI-Nutzung
auf das Lohnwachstum.
Zu den Berufen, in denen das KI-Potenzial als besonders groß eingeschätzt wird, zählen
hochqualifizierte Tätigkeiten mit nichtroutinemäßigen kognitiven Aufgaben, wie z. B.
Laborant*innen, Ingenieur*innen und Versicherungsmathematiker*innen. Ein hohes
KI-Potenzial bedeutet jedoch nicht zwangsläufig, dass in diesen Berufszweigen
Arbeitsplätze wegfallen. Obwohl die Fähigkeiten der künstlichen Intelligenz erheblich
zugenommen haben, kann KI nach wie vor nicht in allen Bereichen genutzt werden und für
viele Tätigkeiten sind weiterhin Menschen erforderlich. Der Effekt der künstlichen
Intelligenz auf die Arbeitswelt dürfte daher vor allem in einer Neuorganisation der
Aufgaben bestehen, aus denen sich die jeweilige berufliche Tätigkeit zusammensetzt. In
einigen Berufen wird KI die Tätigkeit der Arbeitskräfte eher ergänzen, anstatt sie zu
ersetzen.
Die Arbeitskräfte benötigen möglicherweise Umschulungen oder Höherqualifizierungen,
damit sie auf die umstrukturierten bzw. neu entstehenden Aufgaben vorbereitet sind und
einen potenziellen Arbeitsplatzverlust und Jobwechsel bewältigen können. Sie müssen
dazu nicht nur KI-bezogene Kompetenzen erwerben, sondern auch Kompetenzen in
Bereichen, die weniger KI-geeignet sind. Dazu zählen beispielsweise kreative und soziale
Intelligenz, logisches Denken und der Umgang mit Unsicherheit. Wie reibungslos die
KI-Einführung funktioniert und wie groß ihr Effekt auf die Arbeitskräfte ist, hängt auch
davon ab, wie stark sich die einzelnen Unternehmen bemühen, ihre Beschäftigten zu halten
und umzuschulen, und welche institutionellen Rahmenbedingungen in dem betreffenden
For Official Use

Land herrschen. Dies betrifft beispielsweise Aspekte wie die allgemeine Infrastruktur im
Bereich der Weiterbildung und Arbeitsvermittlung, direkte staatliche Förderung,
Steueranreize und die sozialen Sicherungssysteme.
Bestimmte Gruppen von Arbeitskräften sind möglicherweise besser in der Lage, künstliche
Intelligenz zu ihrem Vorteil zu nutzen, sie zur Ergänzung ihrer eigenen Arbeitsleistung
einzusetzen und ihre negativen Folgen zu vermeiden. So zählen zwar einige
hochqualifizierte Tätigkeiten zu den Berufen mit dem höchsten KI-Potenzial, es gibt aber
Anzeichen dafür, dass KI bei Arbeitskräften in besser bezahlten Berufen und/oder mit
höherem Bildungsabschluss zu einem höheren Lohnwachstum führt. Dies lässt auf eine
gewisse Komplementarität schließen. Es deutet jedoch auch darauf hin, dass KI-Nutzung
die Einkommensungleichheit verstärken könnte.
Einige Unternehmen dürften ebenfalls besser als andere in der Lage sein, KI zu entwickeln
und/oder einzuführen. Wenn die Vorteile von KI nur einigen führenden Innovatoren oder
Unternehmen mit übermäßiger Marktmacht zugutekommen, könnte dies einen Keil
zwischen Innovatoren und Arbeitskräfte treiben und den potenziell negativen Effekt von
KI auf die Ungleichheit weiter verstärken.
KI dürfte das Arbeitsumfeld vieler Menschen erheblich verändern. Dies betrifft u. a. die
Arbeitsinhalte und die Arbeitsgestaltung, die Interaktion der Arbeitskräfte untereinander
und mit Maschinen und die Methoden, mit denen Arbeitsleistung und Effizienz gemessen
werden. KI kann einen wichtigen Beitrag zur Verbesserung der
Mensch-Maschine-Interaktion leisten, indem die Arbeitskräfte von eintönigen oder
körperlich anstrengenden Tätigkeiten entlastet werden und stattdessen ihre menschlichen
Fähigkeiten stärker einbringen können. KI kann auch für kostengünstigere, schnellere und
besser skalierbare Lösungen im Personalmanagement eingesetzt werden. Dabei handelt es
sich beispielsweise um Anwendungen, die den Beschäftigten in ihrer beruflichen
Entwicklung helfen, Führungskräfte bei ihren Aufgaben unterstützen und die
Schulungsmöglichkeiten verbessern.
Von denselben KI-Anwendungen könnten jedoch auch erhebliche Risiken für das
Arbeitsumfeld ausgehen, insbesondere wenn sie unsachgemäß genutzt oder ausschließlich
zur Kostensenkung eingesetzt werden. Mangelnde Transparenz und Nachvollziehbarkeit
algorithmischer Vorhersagen und Entscheidungen kann bei den Beschäftigten zu einem
Gefühl psychischer oder physischer Unsicherheit führen. Die umfassende
Leistungskontrolle, die durch KI ermöglicht wird, kann den Arbeits- und
Produktivitäts-druck erhöhen und bei den Beschäftigten verstärkten Stress auslösen, weil
sie nicht wissen, wie ihre Vorgesetzten die Daten interpretieren.
Viele Fragen werden in künftigen Forschungsarbeiten noch zu klären sein. Erhebungen und
qualitative Untersuchungen könnten Aufschluss darüber geben, wie Unternehmen und
Arbeitskräfte die Einführung von KI beurteilen, wie und unter welchen
Managementmodellen und nationalen politischen und institutionellen
Rahmen-bedingungen Entscheidungen getroffen werden und welche Maßnahmen positive
Effekte bewirken. Von besonderem Interesse ist es dabei, Erkenntnisse aus
unter-schiedlichen Bereichen zu erlangen. Beispielsweise könnte untersucht werden, wie
KI die Mensch-Maschine-Interaktion in Fertigungsumgebungen verbessert, wie KI
hoch-qualifizierte Arbeitskräfte bei Prognoseaufgaben unterstützt oder wie KI in Tools für
die Karriereentwicklung genutzt wird.
Weitere empirische Analysen werden sich damit auseinandersetzen, inwiefern der Effekt
der künstlichen Intelligenz mit dem Effekt früherer Automatisierungswellen vergleichbar
For Official Use

ist. Dabei geht es sowohl um ihr Potenzial, die menschliche Arbeitsleistung zu ersetzen
und/oder zu ergänzen und neue Aufgaben zu schaffen, als auch um die Auswirkungen auf
die Arbeitsnachfrage und die Einkommensungleichheit. Für diese Analysen müssen Daten
erhoben und Indikatoren entwickelt werden, mit denen KI und ihre ganz spezifischen
Eigenschaften erfasst werden (neben einer allgemeineren Betrachtung von
Automatisierungstechnologien).
For Official Use

1. Introduction
1. Artificial intelligence (AI) is reshaping economies and societies, offering new

products and services, and promising to generate productivity gains through greater
efficiency and lower costs. At the same time, AI also raises questions and fuels anxieties
about its impact on the labour market and society. Therefore, the purpose of this literature
review is to take stock of what is already known about the impact of AI on the labour
market, identify gaps in the evidence base and inform research under the OECD’s
three-year programme on AI in Work, Innovation, Productivity and Skills (AI-WIPS),
financed by the German Federal Ministry of Labour and Social Affairs (BMAS).
2. AI-WIPS, which started in January 2020, will provide valuable resources and
knowledge, including new in-depth analyses, measurement, international dialogue and
concrete policy assessments on the impact of AI on labour markets and society.
The AI-WIPS activity builds on previous OECD work on AI, including the OECD AI
Principles, which promote an AI that is innovative and trustworthy and that respects human
rights and democratic values. The OECD AI principles call on governments to build human
capacity and prepare for labour market transformation by:
 Empowering people to effectively use and interact with AI systems, including
equipping them with the necessary skills;
 Ensuring a fair transition for workers as AI is deployed, including via social
dialogue, training programmes, support for those affected by displacement, and
access to new opportunities in the labour market; and
 Promoting the responsible use of AI at work, to enhance the safety of workers and
the quality of jobs, to foster entrepreneurship and productivity, and aim to ensure
that the benefits from AI are broadly and fairly shared.
3. This literature review presents what is known about the impact of AI on the labour
market, including the impact on employment and wages, how AI will transform jobs and
skill needs, and the impact on the work environment. The important ethical issues raised
around the use of AI at work are not dealt with in this literature review, and are instead
examined in detail in the forthcoming issues note, “Ethical issues arising from AI
implementation at the workplace and associated policy challenges” (OECD, 2021[1]).
4. Two challenges had to be faced in establishing the scope of this literature review.
The first is that there is no widely accepted definition of AI. While this review tries to cast
a broad net, it is centred on the definition of an AI system established by the OECD’s AI
Experts Group (AIGO) (OECD, 2019[2]):
An AI system is a machine-based system that can, for a given set of human-defined
objectives, make predictions, recommendations or decisions influencing real or
virtual environments. It uses machine and/or human-based inputs to perceive real
and/or virtual environments; abstract such perceptions into models (in an
automated manner e.g. with machine learning (ML) or manually); and use model
inference to formulate options for information or action. AI systems are designed
to operate with varying levels of autonomy.
5. The second challenge is that the development and deployment of AI has not
happened in a vacuum, and as such, other technological advances (such as factory
For Official Use

automation and robotics) are frequently amalgamated with AI in the literature. Here, where
possible, an attempt is made to focus on AI and its attributes, while treating automation as
a potential consequence of AI and robotics as a potentially complementary technology.
6. In line with this, Chapter 2 begins by examining the capabilities of AI and what
relevance they might have for the labour market. It questions what sets AI apart from
previous technological changes, giving particular attention to the specific attributes of AI
and the associated implications for the labour market.
7. Chapter 3 then summarises the literature on the impact of AI on productivity,
employment and wages. These potential impacts are a source of concern for many, who
fear that AI will drive down demand for human labour and wages or even make human
labour obsolete.
8. Chapter 4 looks deeper into the mechanisms driving the AI transition and how they
may transform the way we work, reorganise tasks within any given occupation, and lead to
the emergence of new tasks and occupations. It also examines the abilities of different
groups to adapt to AI adoption and other factors that could drive inequalities.
9. Chapter 5 discusses how AI can reshape the work environment, by changing the
content and design of jobs, the way workers interact with each other and with machines,
and how work effort and efficiency are monitored.
For Official Use

2. What can AI do?
10. This chapter focuses on the capabilities of AI and what relevance they might have
for the labour market. It begins by laying out the characteristics of AI that are frequently
cited by researchers, in particular, those characteristics that have convinced researchers that
the impact of AI on the labour market (as distinct from the impact of technology or
automation more generally) is worthy of special attention. These include: AI’s potential to
affect multiple sectors and occupations across the economy, its ability to self-improve and
to expand the set of tasks that can be automated (including highly skilled ones) –
characteristics that could magnify the labour market impact, whether positive or negative.
11. AI has surpassed some of the limitations of previous technologies. In particular,
AI’s problem solving, logical reasoning and perception capabilities mean that the
automation of some non-routine cognitive tasks is now possible. This explains why some
high-skilled occupations such as radiologists, lab technicians, engineers, lawyers and
actuaries are judged to be highly exposed to AI, i.e. there is overlap between the tasks that
these occupations comprise and the tasks that AI can perform. However, high exposure
does not necessarily mean that jobs in these occupations will disappear.
12. Some bottlenecks in the development of AI remain: humans outperform AI in
creative and social intelligence, reasoning skills and dealing with uncertainty. Even when
AI facilitates automation of certain tasks, this still leaves other tasks that only humans can
perform. Indeed, in these cases, AI can complement workers and enable them to increase
their productivity. In this way, workers in occupations most exposed to AI could see parts
of their work complemented or substituted by AI, and could experience substantial change
in the tasks they perform.
2.1. AI can be considered a general purpose technology
13. The OECD (2019[3]) describes AI as a general purpose technology (or GPT), a
concept developed by Bresnahan and Trajtenberg (1992[4]) to label technologies with
potential application across a broad variety of sectors and occupations, and the ability to
improve over time and to generate complementary innovation. Other examples are
computing, electrification and the steam engine. Agrawal et al. (2019[5]) say that AI
qualifies as a GPT due to its ability to produce predictions, which can be inputs into
decision-making across occupations as diverse as teaching, radiology and translation.
Brynjolfsson et al. (2017[6]) point out that machine learning systems are specifically
designed to self-improve and give examples of how machine learning enables machines to
perceive the outside world, spurring a multitude of innovations. Cockburn et al. (2018[7])
call machine learning an “invention of a method of invention”, a concept introduced by
Griliches (1957[8]). They describe how machine learning and neural networks not only offer
productivity gains across a wide variety of sectors, but also offer transformation of the
innovation processes within those sectors. They highlight the potential of AI to contribute
to scientific discovery especially where research hinges on classification and prediction.
Box 2.1 provides a brief summary of recent trends in AI and its adoption in the workplace.
14. The economic significance of the GPT label is that it adds depth and scale to the
challenges and the opportunities presented by AI to the labour market. The potential for
application across multiple sectors and occupations means that AI is associated with much
larger potential for output and welfare gains (Brynjolfsson, Rock and Syverson, 2017[6]). It
For Official Use

also magnifies the labour market impact, whether positive or negative. One concern is that
since automation has already led to job loss in certain industries, AI could lead to job loss
in a much larger number of industries. On the other hand, the ability of AI to produce
further innovations, by changing “the process by which we create new ideas and
technologies, helping to solve complex problems and scaling creative effort” (Aghion et al.,
2017[9]), could generate entirely new industries and create a myriad of new jobs (as
discussed in section 4.2).
Box 2.1. Recent trends in AI and its adoption in the workplace
Coined as a term in 1956, AI has evolved from symbolic AI where humans built logic-based
systems, through the AI “winter” of the 1970s to the chess-playing computer Deep Blue in
the 1990s. Over the past few years, the availability of big data, cloud computing and the
associated computational and storage capacity and breakthroughs in an AI technology
called “machine learning” (ML), have dramatically increased the power, availability,
growth and impact of AI. Continuing technological progress is also leading to better and
cheaper sensors, which capture more-reliable data for use by AI systems. The OECD report
“Artificial Intelligence in Society” describes these developments in more detail (OECD,
2019[3])
ML is a set of techniques to allow machines to learn in an automated manner through
patterns and inferences from data rather than through explicit instructions from a human.
ML approaches often teach machines to reach an outcome by showing them many
examples of correct outcomes. However, they can also define a set of rules and let the
machine learn by trial and error. The technology driving the current wave of ML
applications is a sophisticated statistical modelling technique called “neural networks”,
which involve repeatedly interconnecting thousands or millions of simple transformations
into a larger statistical machine that can learn sophisticated relationships between inputs
and outputs.
Beyond large and established players in the technology sector, industry adoption of AI is
at an early stage. A US-based nationally representative firm survey (Beede et al., 2020[10])
shows low adoption rates for AI-related technologies such as machine learning, machine
vision, natural language processing and automated guided vehicles. Industries leading in
AI adoption include high tech, automotive and assembly, telecoms, transport and logistics,
financial services and consumer packaged goods, retail and healthcare (based on surveys
by Bessen et al. (2018[11]) and McKinsey (2019[12])). The same studies show that AI tends
to be embedded in technologies such as natural language understanding and text analysis,
natural language classification and decision management, visual recognition (including
image, face and video) and virtual agents or conversational interfaces (“chatbots”) and
robotic process automation.
2.2. AI can be considered an automation technology
15. Many economists also consider AI an automation technology, i.e. one designed to
facilitate the automation of tasks that would otherwise be performed by humans, thereby
reducing labour demand and wages for certain groups of workers (Acemoglu and Restrepo,
2018[13]; Aghion et al., 2017[9]). What may distinguish AI from other automation
technologies, such as industrial robots and other automated machinery, is its greater
potential to expand the range of tasks that can be automated. This may be particularly the
For Official Use

case with a technology such as machine learning, which is specifically designed to

self-improve. Some believe this ability to self-improve could lead to the singularity
(discussed in Box 2.2) and ultimately challenge humans’ place in the labour market and
society, though lagging productivity statistics throw doubt on this prediction.
16. There is already some evidence to suggest that AI can facilitate the automation of
tasks where automation was previously impossible. Until recently, automation has affected
mostly routine tasks (Autor, Levy and Murnane, 2003[14]) and low-skilled tasks
(Nedelkoska and Quintini, 2018[15])1. Some identify AI as a force to enable the automation
of non-routine and/or high-skilled tasks, because of the specific capabilities of AI (Aghion
et al., 2017[9]). The next section shows that the occupations judged to be most exposed to
AI include high-skilled occupations – including some traditionally ‘white-collar
professions’ requiring non-routine cognitive tasks, such as lab technicians, engineers and
actuaries.
17. As highlighted by the OECD 2019 Employment Outlook (2019[16]), the potential
threat to people who have been historically more sheltered from economic changes,
including white-collar workers with relatively high levels of education and secure jobs,
may be another factor driving public concerns around AI. At the same time, if AI did
displace higher paid workers, this could potentially mitigate the trends of increasing income
inequality and polarisation associated with automation technologies up to now2.
Box 2.2. AI and the singularity
In the eyes of some, AI’s ability to self-improve could lead to a singularity, which describes a
point in time at which machine intelligence exceeds human intelligence (Bostrom, 2006[17];
Good, 1966[18]) and economic growth accelerates “as an ever-accelerating pace of
improvements cascade through the economy” (Nordhaus, 2015[19]), challenging humans’ place
in the labour market. Others are more sceptical. For instance, Luc Julia, one of the creators of
voice assistant Siri argues that advancement in AI will always be dependent on human
knowledge and decision-making (Julia, 2019[20]).
Nordhaus (2015[19]) notes that the proponents of the singularity theory are most often computer
scientists although he identifies some economists (Brynjolfsson and McAfee, 2014[21]) who, in
his view, propose a “soft version” of the theory. Nordhaus tests a variety of hypotheses that
would indicate that technology-driven growth is accelerating but ultimately he does not find
sufficient support for this theorised acceleration. For instance, the capital-output ratio is not
rising rapidly, the decline in the cost of capital is not accelerating, and productivity growth is
not rising (as discussed in section 3.1). He concludes that the singularity is at least 100 years
away. This conclusion is roughly in line with the results of a survey of machine learning
researchers, which assigns a 50% chance to the outcome in which AI outperforms humans in
all tasks in 45 years and leads to the automation of all human jobs in 122 years (Grace et al.,
2017[22]).
1
These trends are referred to as routine-biased and skill-biased technological changes, respectively.
2
However, the size of this impact could be limited if higher skilled individuals began competing for
jobs typically performed by lower skilled individuals (Acemoglu and Restrepo, 2018[122]).
For Official Use

2.3. Certain occupations are more exposed to AI
18. One way to understand the capabilities of AI (and the likely labour market impact)
is to ask which occupations comprise the tasks that AI can perform.3 Researchers generally
refer to these occupations as being “exposed” to AI. It should be noted that “most exposed”
does not necessarily mean most likely to be replaced by AI, as the studies are based on
assessments of the technical feasibility of AI, and are limited in their consideration of other
factors. Additionally, “least exposed” to AI does not necessarily mean that that occupation
escapes the risk of automation. Some of these occupations are exposed to other
technologies, which have already led or could lead to their automation. Workers in
occupations most exposed to AI could see substantial change in the tasks they perform, but
could also see their work complemented (rather than substituted) by AI.
19. Whether AI has a positive or negative impact on jobs, one would expect the impact
to be strongest in occupations or sectors that rely most on the tasks that AI can carry out.
Researchers measure exposure using various methods:
 Webb (2020[23]) identifies AI patents (i.e. those with keywords such as “supervised
learning” and “reinforcement learning” together with “neural network” and “deep
learning” in their titles or abstracts) and then assesses the overlap (in verb-noun
pairs4) between the text of the patents and the text of job task descriptions (from
the O*NET database of occupations and tasks) in order to see which occupations
are most exposed to AI.
 Felten et al. (2019[24]) map (with the help of some computer science PhD students)
different AI categories (such as abstract strategy games, translation and image
recognition) to skills (from the O*NET database), in order to assess which
occupations rely on abilities where most AI progress has been seen.
 Brynjolfsson et al. (2018[25]) identify tasks (from the O*NET database) and
occupations (using Burning Glass data) suitable for AI by applying a rubric which
includes parameters such as: whether the task is describable with rules; whether it
requires complex, abstract reasoning; and whether it is highly routine and repeated
frequently.
2.3.1. High-skilled occupations are among those most exposed to AI

20. Table 2.1 provides a summary of the findings regarding which occupations are most
and least exposed to AI (or, more specifically: machine learning, which all three papers
interpret as representing AI’s capabilities).
3
Most researchers adopt a task-based approach (based on Autor et al.’s influential paper (2003[14])),
such as this, to examine changes in the labour market. This is because the impact of AI is unlikely
to act on an entire job all at once. As such, AI might automate a part of a job rather than a job in its
entirety, and certain tasks will be more or less suitable for replacement by AI than others.
4
Popular verb-noun pairs in AI patents include, for example: classify image, predict quality,
generate rating.
For Official Use

Table 2.1. Occupations most and least exposed to AI
(Felten, Raj and Seamans, (Brynjolfsson, Mitchell and

(Webb, 2020[23])
2019[24]) Rock, 2018[25])
Most ‒ High-skilled occupations, including ‒ White-collar occupations, such ‒ Concierges, mechanical
exposed clinical lab technicians, optometrists and as chemical/civil/nuclear drafters, credit authorisers,
chemical engineers. engineers, epidemiologists, brokerage clerks, and
‒ Production jobs involving inspection and actuaries, statisticians, credit morticians, undertakers,
quality control, which Webb describes as analysts, accountants, computer and funeral directors.
representing a small proportion of the programmers, operations
low-skilled workforce research analysts.
Least ‒ High-skilled occupations requiring ‒ Physical occupations, including ‒ Massage therapists,
exposed reasoning about novel situations (e.g. maids and cleaners, cafeteria animal scientists,
researchers). attendants, dishwashers, hotel archaeologists, public
‒ Occupations requiring interpersonal skill porters, slaughterers and meat address system and other
(e.g. teachers and managers), including packers, roofers and painters, announcers, and
manual work such as baristas, food massage therapists, fitness plasterers and stucco
preparation workers or massage instructors. masons.
therapists.
Note: This is a selection of results. See papers for full lists of impacted occupations.
21. Looking across the results of these three studies, one of the most striking things is
that some high-skilled occupations are found to be among the most exposed to AI. The
results of Felten et al. show that almost all of the most exposed occupations are
“white-collar” jobs requiring an advanced degree. Brynjolfsson et al. and Webb appear to
find a greater skill mix among the most exposed occupations, but Webb notes that the
highly exposed low-skilled occupations represent only a small proportion of the
workforce.5
22. The finding that high-skilled occupations are exposed to AI can be contrasted with
other research which finds that low-skilled occupations are highly exposed to automation
technologies (in a broader sense) and therefore at highest risk of automation (while
high-skilled jobs are at lowest risk) – as in Nedelkoska and Quintini (2018[15]) and Frey and
Osborne (2017[26]). For instance, cleaners were considered at high risk of being automated
within a 20-year period according to Nedelkoska and Quintini, but not particularly exposed
to AI according to Felten et al. and Webb. One reason for this contrast may be the
technological innovations that have emerged in the last decade, as the studies by
Nedelkoska and Quintini and Frey and Osborne are based on an exercise performed in
2013.
23. Another reason may be differences in definition, as the analyses are likely to be
quite sensitive in this regard.6 The three more recent studies focus specifically on the
technical capabilities of machine learning, while the studies by Nedelkoska and Quintini
and Frey and Osborne consider a broader set of technological advances, including not only
5
Examining wage differentials, Webb finds that higher wage occupations tend to be more exposed
to AI while Brynjolfsson et al. find a low correlation between wages and exposure. Felten et al. do
not report such findings.
6
Indeed, Webb shows that by adapting his analysis to focus on technological changes other than
machine learning (specifically software and robots), a different set of occupation is designated as
highly exposed. Robots tend to automate repetitive manual tasks (e.g. materials movers in factories)
while software performs non-manual tasks that can be hard-coded in advance (e.g. broadcast
equipment operators).
For Official Use

AI but also mobile robotics (some of which use AI and some of which do not) (Frey and
Osborne, 2018[27])7. It is possible that the narrow focus of the three more recent studies on
the technical capabilities of machine learning could overlook less novel applications of AI.
This could be the case if AI replaces or is embedded in older automation technologies,
making them easier or cheaper to adopt. To the extent that AI could make older automation
technologies more attractive, the impact of AI along skill lines could be more similar to
previous waves of automation.
24. The researchers also attempt to identify the features that typify the occupations
most exposed to AI. While other automation technologies have typically been capable of
performing mostly routine tasks (Autor, Levy and Murnane, 2003[14]), Webb points to AI’s
capabilities to perform non-routine tasks, and particularly non-routine cognitive tasks.
Felten et al. draw attention to AI’s cognitive abilities, defined as abilities that “influence
the acquisition and application of knowledge in problem solving”. Improvements in AI in
relation to problem solving, logical reasoning, and perception may explain why highly
skilled technicians and engineers appear to be highly exposed to AI in the studies by Felten
et al. and Webb.
25. Turning to demographic characteristics, Webb finds that occupations requiring
higher skills, judgement and accumulated experience tend to be more exposed to AI, with
the result that more educated workers and workers older than 30 are more exposed. Webb
also finds that male workers are more likely to be exposed to AI (also to robotisation and
software), which he attributes to female-dominated occupations tending to require more
interpersonal skills and male-dominated occupations tending to require more technical
skills.
26. These studies identify the occupations most exposed to AI based on technical
feasibility, but are more limited in what they can say about whether workers in these
occupations will see their work substituted or complemented and about the overall impact
on demand for human labour.
2.3.2. Some bottlenecks to the development of AI remain

27. Despite advances in the technical capabilities of AI, some bottlenecks to
development remain. Webb identifies social interaction as an important feature of
occupations with low exposure to AI, regardless of skill level. He also finds that
occupations which combine manual work with interpersonal skills are among those least
exposed to AI, including for instance massage therapy, which is identified as having low
exposure to AI by all three papers. Physical tasks are identified by Felten et al. as having
low exposure to AI (although this does not exclude the possibility that they are exposed to
non-AI technologies). High-skilled research occupations which require either reasoning
about novel situations (Webb, 2020[23]) or some manual activities (e.g. animal scientists
and archaeologists (Brynjolfsson, Mitchell and Rock, 2018[25])) are also judged to be only
lightly exposed to AI.
7
Both studies are based on an exercise in which a group of machine learning researchers assessed a
list of occupations asking the question: “Can the tasks of this job be sufficiently specified,
conditional on the availability of big data, to be performed by state of the art computer-controlled
equipment?” The question does not reference AI specifically but it seems quite likely that machine
learning researchers would have considered the capabilities of AI in addition to other automation
technologies in their assessment.
For Official Use

28. It may not be obvious why lower skilled occupations might be only lightly exposed
to AI. Gries and Naudé (2018[28]) refer to Moravec’s Paradox, which observes that tasks
that require high-level reasoning demand relatively light computational resources while
tasks requiring sensor-motor skills (generally associated with lower skill occupations)
demand enormous computational resources. They point to the challenge in using
technology to replace humans in occupations such as security staff, cleaners, gardeners,
receptionists and chefs. Related to Moravec’s Paradox is Polanyi’s Paradox, which refers
to the challenge that computers face in performing tasks relying on tacit knowledge
(i.e. tradition, intuition, inherited practices, implied values, and prejudgments), such as
organising a closet. Michael Polanyi’s observation is that “We can know more than we can
tell” (Polanyi, 2009[29]). More precisely, there are many tasks which people understand
intuitively how to perform, but cannot elicit the rules or procedures they follow.
29. As AI continues to advance, some existing bottlenecks may be overcome,
potentially exposing certain occupations to AI that were previously unexposed. In
particular, there is a debate among computer scientists about whether AI (and machine
learning in particular) can in the future provide a solution to Polanyi’s paradox, by applying
statistics and inductive reasoning to supply best-guess answers in cases where formal
procedural rules are unknown (Autor, 2014[30]). But even in this case, machine learning
may only ever “get it right” on average while missing many of the most important and
informative exceptions. The solution may require a multifaceted set of inputs: brains and
brawn; technical mastery and intuitive judgment; adherence to rules and judicious
application of discretion; creativity and rote repetition. Typically, these inputs each play
essential roles: improvements in one do not obviate the need for the other. Perhaps, the
most important limitation of AI systems is that they are devices for answering questions,
not for posing them (Brynjolfsson and McAfee, 2017[31]). That means entrepreneurs,
innovators, scientists, creators, and other kinds of people who figure out what problem or
opportunity to tackle next, or what new territory to explore, will continue to be essential.
2.4. AI will not only lead to automation, but will also complement labour
30. The studies discussed in the previous section are based on the idea that certain tasks
are more suited to being performed by AI. In fact, the study by Brynjolfsson et al. (2018[25])
finds that many occupations comprise both tasks with high and low suitability for machine
learning. For instance, an economist may have to forecast economic trends using a dataset
(high suitability for machine learning due to the use of digital data inputs) and also write
reports and provide guidance based on their research (low suitability for machine learning
due to the reliance on complex, abstract reasoning and the importance that these outputs
are perceived to come from a human).8
31. As not all tasks within an occupation can be performed by AI, this suggests that the
impact of AI will be to replace workers in certain tasks and lead to the transformation of
occupations (discussed further in section 4.1) rather than their disappearance. Workers in
the transformed occupations will thus be predominantly complemented by AI (Fossen and
Sorgner, 2019[32]). Thus, AI has not only the potential to destroy jobs, but also to
complement them.
8
Overall, Brynjolfsson et al. assess the occupation of “economist” as close to average in terms of
suitability for machine learning.
For Official Use

32. AI is expected to increase productivity not only by enabling firms to replace labour
with cheaper capital, but also by complementing workers (Agrawal, Gans and Goldfarb,
2019[5]). AI is already enabling some workers to increase their productivity by leveraging
their social interaction skills and ability to reason about novel situations (areas in which
humans outperform AI, as shown in section 4.3.2). For instance, healthcare researchers
expect AI to complement human clinicians as their jobs transform to draw more on
uniquely human skills like empathy, persuasion and big-picture integration (Davenport and
Kalakota, 2019[33]). Another example highlighted in the OECD’s Artificial Intelligence in
Society (OECD, 2019[3]) is Alibaba’s chatbot, which handled more than 95% of customer
inquiries during a 2017 sale, thereby allowing human customer representatives to handle
more complicated or personal issues (Zeng, 2018[34]). Many other examples are provided
throughout this literature review from AI assisting teachers to deliver individualised
learning to AI enabling close human-machine collaboration in manufacturing
environments.
33. Agrawal et al. (2019[5]) note from their interactions with AI start-ups that, while
many talk about the impact of AI on labour in terms of potential substitution,
complementarity, and demand expansion, very few companies say that they are building
unambiguously labour-replacing technologies. In surveys, businesses also suggest that
decisions to adopt AI are motivated more by the aim of complementing human capabilities
than by the aim of substituting workers (Accenture, 2018[35]; Bessen et al., 2018[11]).
For Official Use

3. Impact of AI on productivity, employment and wages
34. Discussions about the predicted impact of AI on productivity, employment and

wages are filled with uncertainty. AI is expected to increase productivity but there is debate
about the size of the impact, particularly when predictions rely on advances that have yet
to be seen.
35. Even if AI boosts productivity substantially, it is not clear that workers will
necessarily share in the benefit in the form of higher employment and/or wages. This is
because AI can facilitate automation, contributing to downward pressure on the demand
for labour and a decoupling of productivity from labour market outcomes such as
employment and wages. These forces may counteract the productivity effect, which might
otherwise be expected to increase labour demand, employment and wages.
36. While the theory is ambiguous, the empirical evidence based on AI adopted in the
last 10 years does not support the idea of an overall decline in employment and wages in
occupations exposed to AI. Some studies suggest a positive impact of AI on wage growth,
with larger increases experienced by individuals in higher wage occupations and/or with
higher educational attainment. This suggests that these workers are more capable or better
positioned to use AI to complement their own labour, boost their productivity and to share
in the benefits.
3.1. AI will increase productivity but the size of the impact is debated
37. Much of the available economics literature on AI centres on its potential to increase
productivity, by reducing costs (including by enabling firms to replace labour with cheaper
capital), complementing labour and spurring complementary innovations (Agrawal, Gans
and Goldfarb, 2019[5]; Brynjolfsson, Rock and Syverson, 2017[6]; Cockburn, Henderson
and Stern, 2018[7]). However, the productivity paradox9 is the term used to refer to the fact
that productivity growth has been lagging over the past decade or so (Andrews, Criscuolo
and Gal, 2016[36]), despite substantial progress in AI (in particular, breakthroughs in
machine learning) and other technologies. Researchers attempt to understand the causes of
this productivity paradox, in order to be able to predict how AI will affect future labour
productivity growth.
38. One explanation is that the potential of AI (and more generally, other recent
technological advances) has been overestimated and that the predicted productivity
improvements will never come to pass. For instance, for Gordon (2018[37]), the modest
contribution of AI and robotics to productivity is part of the reason why productivity growth
was slower between 2006 and 2016 than in the preceding decade.10 In his view, much of
the impact of AI has been seen already (e.g. customer service bots, searching through legal
texts, assisting radiology diagnostics) and any further innovations (e.g. in medical research,
big data, and driverless vehicles) are more likely to be marginal improvements of past
9
Similar to the Solow Paradox in which “you can see the computer age everywhere but in the
productivity statistics” (Brynjolfsson, Rock and Syverson, 2017[6]).
10
Gordon posits that the IT revolution which boosted productivity between 1996 and 2006 has
reached “maturity”, as evidenced by the declining productivity of researchers across many fields
making true technological breakthroughs less likely in future.
For Official Use

technologies rather than true technological breakthroughs resulting in large productivity

boosts. He expects AI to replace workers in some jobs, but that this will happen at a steady
pace rather than as a sudden upheaval. He also expects that AI will complement workers in
other jobs, but in ways that produce only modest increases in productivity.
39. On the other hand, researchers who believe that AI has the potential to boost
productivity significantly (along the lines of a GPT, as explained in section 2.2) attribute
the productivity paradox primarily to lags in AI implementation and restructuring, which
can result in it taking years or even decades before substantial economic gains of GPTs are
seen (Brynjolfsson, Rock and Syverson, 2017[6]; Brynjolfsson, Mitchell and Rock,
2018[25]).11 OECD analysis confirms that uneven uptake and diffusion of digital
technologies throughout the economy is an important source of the productivity slowdown
(Andrews, Criscuolo and Gal, 2016[36]) and suggests that digitalisation may have
contributed to the widening performance gap between more and less productive firms (Gal
et al., 2019[38]), as less productive firms can find it harder to attract workers with the right
skills to help them adopt digital technologies efficiently.
40. Other factors that may have contributed to the productivity paradox include:
 Mismeasurement in the productivity statistics, due to difficulties in capturing
improvements in the quality of high tech products (Byrne and Sichel, 2017[39]);
 The gains of AI accruing mostly to a small number of superstar firms in a
“winner-takes-most” dynamic, whose market power enables them to engage in
wasteful and potentially anti-competitive efforts to block others from accessing the
technology (see (Brynjolfsson, Rock and Syverson, 2017[6]) and as discussed
further in section 4.5). However, Schwellnus et al. (2018[40]) attribute the
“winner-takes-most” dynamic to technological dynamism, rather than
anti-competitive forces.
 Automation being introduced at an excessive rate, resulting in deployment of the
wrong types of AI12 and mismatches between skills and new technologies, thereby
slowing productivity growth (Acemoglu and Restrepo, 2018[13]).
41. One challenge in weighing the relative merits of the arguments regarding AI’s
potential to increase productivity is that, at any point in time, it is difficult to forecast
productivity growth a couple of years into the future – especially when a predicted
productivity boost relies on the invention of an entirely new technology or an entirely new
application of a technology. To highlight the challenge of relying on “our own limited
imaginations” in this regard, Cowen (2016[41]) gives examples of past technological
breakthroughs that have come as a surprise, among them: x-rays, radio and transistors.
Cappelli (2020[42]) makes the point that claims reliant on future advances are easy to make
but impossible to refute. Despite these challenges, some consultancies have attempted to
put a value on the potential contribution of AI to economic growth (as detailed in Box 3.1),
estimating that AI could deliver additional global economic output of up to $15.7 trillion
by 2030 (PWC, 2018[43]).
11
Indeed, a survey by Ransbotham et al. (2019[71]) shows that 40% of organisations that make
significant investments in AI do not report business gains from it.
12
As discussed in (Acemoglu and Restrepo, 2020[49]), ones which are only good enough to replace
workers but not good enough to create new tasks and boost productivity sufficiently to raise demand
for human labour.
For Official Use

42. What are the implications for future productivity growth? If Gordon’s view holds
true, then the impact of future AI developments will be modest. If, on the other hand, slow
growth is due to lags in AI implementation and restructuring, this could still be consistent
with projections of substantial economic growth. In other words, growth would follow an
S-curve pattern with a slower start given the need to learn, invest and deploy the new
technology, followed by an acceleration driven by competition and improvements in
complementary technologies, and then a final period of slower growth again once the
technology is widespread and market competition lowers the returns earned by early
adopters13.
Box 3.1. Estimates of the economic potential of AI
Studies by consultancies suggest that AI has enormous potential to contribute to global

economic output, as shown in Table 3.1. Consultancies have generally tended to view AI as a
revolutionary and transformative force, and one that can boost economic output by increasing
productivity (by substituting workers and by complementing workers and capital), increasing
consumption, enhancing the diffusion of innovation and creating a revenue stream for
AI-producing firms. McKinsey’s estimate of $13 trillion by 2030 corresponds to an additional
1.2% annual contribution to GDP, which is greater than the additional 0.4% contributed by
robotisation during the 1990s and the additional 0.6% contributed by the spread of IT during
the 2000s (2018[44]). Accenture’s estimate of $14 trillion by 2035 is based on a projected
increase in labour productivity of up to 38% in some countries (2017[45]). The studies that
consider job loss generally assume it will be cancelled out by job creation in the long run.
Table 3.1. Estimated economic potential of AI
Impact Timeframe
(Accenture, 2017[45]) USD 14 trillion* By 2035
(Analysis Group, 2016[46]) in a study funded by Facebook USD 1.49-2.95 trillion By 2026
(McKinsey, 2018[44]) USD 13 trillion By 2035
(PWC, 2018[43]) USD 15.7 trillion By 2030
Note: * based on 16 industries in 12 economies that make up 50% of global economic output. Estimation approaches
differ, considering factors such as: productivity increases due to AI replacing workers and augmenting workers and
capital, additional consumption, diffusion of innovation and returns to AI-producing firms.
These studies pay relatively little attention to the transition period, with the exception of
McKinsey, who acknowledge potential negative externalities linked to displacement and wage
polarisation, and net out the costs of transition to get their final estimate. PWC and McKinsey
assume that adoption of AI follows an S-curve pattern, with a slower start followed by an
acceleration.
3.2. Theoretical models are ambiguous on the impact of AI on employment and

wages
43. This section explains why the impact of AI on employment and wages is
ambiguous. While AI-facilitated automation is expected to reduce labour demand, it can
increase it under specific circumstances.
13
The assumption that new technologies follow an s-shape pattern is almost a stylised fact,
according to Geroski (2000[124]).
For Official Use

3.2.1. AI-facilitated automation is expected to reduce labour demand and

decouple wages from productivity gains
44. Economists have proposed a theoretical framework centred on AI being an
automation technology to explain how AI has the potential to enhance productivity while
simultaneously reducing labour demand, wages and the labour share (Acemoglu and
Restrepo, 2018[13]). One limitation of this framework is that, by treating automation
technologies together, they assume that AI is similar to other automation technologies, such
as industrial robots and other automated machinery – an assumption which has yet to be
proven (Naudé, 2019[47]) – and do not pay attention to the inherent capabilities of AI. This
limitation is discussed in section 3.2.3.
45. The key feature of an automation technology is that it expands the set of tasks
within the production process that can be performed by capital, so that the share of tasks
performed by capital increases and the share of tasks performed by labour decreases
(Acemoglu and Restrepo, 2019[48]). This may give firms the possibility to replace labour
with cheaper capital, resulting in productivity gains. The displacement effect describes
capital taking over tasks previously performed by labour. This reduces labour demand, and
puts downward pressure on employment and wages. Additionally, because displacement
increases output at the same time, it tends to have the impact of reducing the share of labour
in national income14 and decoupling wages from any productivity gains. This is one reason
why productivity may increase but workers may not see their wages increasing at the same
rate. This is also why treating AI as a purely factor-augmenting technological change,
i.e. a force which increases the productivity of labour or capital while ignoring the
displacement effect, can lead to misleading conclusions (Acemoglu and Restrepo, 2019[48]).
3.2.2. However, the creation of new labour-intensive tasks can increase the
demand for labour in the long run
46. Even though the displacement effect may put downward pressure on employment
and wages, there are other countervailing forces at play (Acemoglu and Restrepo, 2018[13]):
 The productivity effect, whereby cost savings generated by automation increase
consumer demand (in the same sectors experiencing automation15 and/or in other
sectors16), which increases the demand for labour in non-automated tasks;
14
Schwellnus et al. (2018[40]) show that technological progress and (to a lesser extent) globalisation
can explain most of the contraction in the labour share over the last two decades. Capital-augmenting
technological progress or technology-driven declines in relative investment prices reduce the labour
share by fostering labour-capital substitution and increasing overall capital intensity.
15
Similar to how the introduction of ATMs in the banking industry in the 1970s produced cost
savings and additional consumer demand, which led banks to open additional branches, thereby
offsetting the original displacement of bank tellers (Bessen, 2015[120]). Bessen (2018[117]) shows that
a number of industries, including textiles, steel and automotive, experienced strong employment
growth during periods of rapid technological progress and productivity growth, which could have
been feared to cause a net job loss.
16
By increasing productivity and reducing prices, certain technologies have a positive impact on
employment in industries other than the ones where they are deployed (Autor and Salomons,
2018[119]). An example is a large supermarket chain introducing a new business model that generates
For Official Use

 The capital accumulation effect, whereby automation increases the capital intensity
of production, triggering accumulation of capital, which also raises the demand for
labour (in tasks where AI and automation are complementary to human labour)17;
 The deepening of automation, whereby technological improvements increase the
productivity of existing machines (i.e. with no additional displacement of labour),
boosting the productivity effect and further increasing the demand for labour; and
 The creation of new high-productivity, labour-intensive tasks, which increases the
labour share (potentially in the longer term), counteracting the impact of
automation18.
47. Acemoglu and Restrepo highlight an important dynamic within their model. While
the productivity effect, capital accumulation effect and deepening of automation are
important forces in counteracting the downward pressure on labour demand, employment
and wages, these forces are unlikely to be large enough to counter the displacement effect
in the short term. This is because, as automation continues, a falling share of the
productivity gains accrues to labour.19,20 However, in the longer term, through the creation
of new high-productivity, labour-intensive tasks (i.e. ones that “reinstate labour as a central
input into the production process” (Acemoglu and Restrepo, 2020[49])), a reinstatement
effect can increase the labour share and directly counteract the displacement effect. This
can ultimately set the growth process on a more balanced path, in which AI increases
productivity, employment and wages overall.
48. The creation of new labour-intensive tasks is thus a critical mechanism for
adjustment, but one that can be slow and one that depends crucially on the nature of AI
being deployed. Not all AI applications will create new labour-intensive tasks. For instance,
Acemoglu and Restrepo (2019[48]) decompose the change in the task content of production
in the US and find stronger displacement effects and considerably weaker reinstatement
effects during the last three decades than the decades before. They draw attention to the
need to develop the right types of AI, i.e. AI that creates new tasks and boosts productivity
sufficiently to raise demand for human labour (as opposed to the wrong types, which are
only just good enough to replace workers) (Acemoglu and Restrepo, 2020[49]).
3.2.3. How does this framework apply to AI specifically?

49. In their paper “Artificial Intelligence: The Ambiguous Labor Market Impact of
Automating Prediction” (2019[5]), Agrawal et al. adapt Acemoglu and Restrepo’s
considerable economies of scale and leads to lower prices, allowing consumers to increase their
spending in other industries.
17
Acemoglu and Restrepo (2018[13]) suggest that this effect may have mitigated the impact of the
rapid accumulation of tractors in the American economy in the first half of the 20th century, citing
Olmstead and Rhode (2002[121]).
18
Along the lines of how the Industrial Revolution spurred new jobs such as engineers, machinists,
repairmen, conductors, back-office workers and managers to support emerging technologies
(Acemoglu and Restrepo, 2018[13]).
19
It is this aspect in particular that encourages Acemoglu and Restrepo to reject the idea that
automation always leads to greater employment and wages.
20
When labour income does not reflect productivity gains, this may also suppress consumer demand
and dampen the productivity effect (Gries and Naudé, 2018[28]).
For Official Use

framework to hone in specifically on the impact of machine learning using a task-based

approach. They view machine learning as an advance in the field of prediction, which can
therefore substitute for human prediction tasks (i.e. displacement effect) as an input to
decision-making tasks (e.g. screening CVs for potential hires). By reducing uncertainty in
prediction, machine learning can increase the relative returns to labour (e.g. a surgeon
complemented in their work by an automated scan for cancer cells) or the returns to capital,
if it encourages the automation of the decision task also (e.g. once the prediction of
obstacles is automated, there is more reason to automating vehicle control). Machine
learning may also create new decision tasks, which can be performed by capital or labour
(in which case reinstatement can occur), although Agrawal et al. say that there are few
tangible examples of this at this stage.
50. The work by Agrawal et al. suggests that the framework developed by Acemoglu
and Restrepo can be reconciled with the concept of AI as a tool to enhance prediction,
beyond simply treating AI as an automation technology. In either case, the theory is
ambiguous on whether AI increases or decreases employment and wages. The ultimate
impact will depend on the type of AI being developed and deployed, how it is developed
and deployed, in addition to market conditions (Caselli and Manning, 2017[50])21 and policy
and institutions (OECD, 2019[16]; Aghion, Antonin and Bunel, 2020[51]; Acemoglu and
Restrepo, 2018[13]). Additionally, the impacts on inequality should not be underestimated
as there is no reason to believe that the displacement effects, productivity effects and
emergence of new labour-intensive jobs will be distributed evenly across industries,
regions, and socio-demographic groups.
3.3. The limited empirical evidence does not support the idea that AI has reduced
employment and wages
51. Given the ambiguity surrounding the impact of AI on employment and wages in
theoretical models, what does the empirical literature say? The body of literature on this
topic is not large (see (Seamans and Raj, 2018[52]) and (McElheran, 2018[53]) for discussion
of the related challenges) but this review has identified a few studies that examine recent
historical data in the United States22 for the impact of AI (or more specifically, machine
learning, which the measures tend to capture) as opposed to broader automation or
technological progress.
52. These studies do not support the idea of an overall decline in employment and
wages. Some studies suggest a positive impact of AI on wage growth, with larger increases
experienced by individuals in higher wage occupations and/or with higher educational
attainment.
21
Caselli and Manning (2017[50]) show that average wages will rise with the introduction of a new
technology given certain conditions: that the price of investment goods falls compared to more
labour-intensive consumer goods and that market competition is not reduced. As always, there are
winners and losers but one implication of average wages rising is that if workers can switch jobs
easily and/or redistribution is happening, all workers can benefit.
22
The fact that the empirical studies identified in this review analyse only employment trends in the
United States is likely due to the reliance of the measures of exposure to AI on the US-based O*Net
database of occupations and tasks.
For Official Use

53. Felten et al. (2019[24]) examine AI advances and US labour market trends at
occupation-state level between 2010 and 2015. They show that an occupation’s exposure
to AI (specifically, the areas of AI that have seen most advances in recent years) has a small
positive link with wages but no link with employment. The positive relation is mostly
driven by occupations that require a high level of familiarity with software and high-income
occupations23.
54. Fossen and Sorgner (2019[32]) apply the same measure to US individual-level panel
data from 2011 to 2018 and find that exposure to advances in AI is associated with greater
job stability (as indicated by an individual becoming non-employed or switching to a new
occupations) and wage growth overall. They interpret this result as indicating that AI is
predominantly complementary to human labour. The effects are even stronger for those
with higher levels of formal education and more experience (as indicated by age),
suggesting that these workers are more able to use AI to complement their own labour and
boost their productivity. In contrast, the authors find that measures representing exposure
to computerisation (as developed by Frey and Osborne (2017[26])) are associated with lower
job stability and job growth.
55. Acemoglu et al. (2020[54]) examine changes in US job postings between 2010 and
2018 across establishments and occupations according to their exposure to AI. The study
uses the exposure measures developed by Felten et al, Brynjolfsson et al, and Webb. Within
establishments most exposed to AI (but not producing and supplying AI themselves), they
observe a shift in the composition of job postings away from occupations most exposed to
AI to occupations least exposed, although effect sizes are modest and results are not robust
across specifications. They interpret this as suggesting any productivity and/or
complementary effects of AI in these establishments are small and lower than substitution
effects, although they acknowledge that it may simply be too early to observe an impact of
AI in overall employment patterns (i.e. outside the market for AI talent). Overall, they find
no significant impact on job quantity and no significant impact on the sets of skills required
in exposed occupations (i.e. representing either a need for new skills or obsolescence of
previously common skills).
56. None of these studies finds evidence to support the idea of an overall decline in
employment due to recent advances in AI. The studies by Felten et al. and Fossen and
Sorgner find a positive impact of AI on wage growth, with larger increases experienced by
individuals in higher wage occupations and/or with higher educational attainment. Felten
et al. note the possible implication that AI adoption could exacerbate income inequality.
57. While these backward-looking studies shine a light on the impact on employment
and wages of developments in AI over the last decade, it is unclear to what extent future
developments in AI will produce similar results. This question poses a substantial challenge
for researchers, who must attempt to predict the impact of entirely new technologies and
entirely new applications of existing technology, and for policymakers who may wish to
devise policy around this impact.
23
Exposure to AI was not related to employment growth or wage growth for low- or middle-income
occupations.
For Official Use

3.4. Firms are divided in their expectations about AI’s impact on overall labour
demand
58. Surveys of employers show that opinions are divided about whether they think that
AI will increase or decrease employment in future. In a survey of executives whose
companies have adopted AI (McKinsey, 2019[12]) the results pointed to an expected
decrease in the number of employees over the subsequent three years due to AI, despite
showing that AI had led to job growth in the preceding year. The reasons for this outlook
were not probed further, for instance whether it derives from expectations that upcoming
developments in AI would be more suited to substituting labour. In a survey of both tech
executives and the general population (Edelman, 2019[55]), two thirds of tech executives
surveyed believed that AI could increase employment. A minority within the general
population surveyed agreed, although a majority did agree that AI could produce an
increase in employment in the long term. Business surveys by Bessen et al. (2018[11]) and
McKinsey (2019[12]) suggest that businesses support the view that the impact of AI on jobs
is much more about the shifting of work from some occupations to others than about
eliminating labour overall.
59. Furthermore, there appears to be consensus that the impact will differ by occupation
and industry. The surveys by Bessen et al. (2018[11]) and McKinsey (2019[12]) suggest job
creation is likely to be experienced in sales and marketing (occupations which, Bessen et
al. note, involve the complementary use of AI) and that job loss is likely to be experienced
in manufacturing and some clerical occupations. The survey by Bessen et al. shows that
startups that sell AI products to customers in the agriculture, manufacturing, utilities and
transportation industries are much more likely to say that their products reduce labour costs
and/or automate routine tasks.
For Official Use

4. The AI transition
60. The impact of AI on the labour market is likely to run much deeper than changes
in employment and wages. The mechanisms underlying these changes could transform the
way we work, reorganising tasks within any given occupation. The adoption of AI may
result in the emergence of new tasks and occupations, which only humans can perform.
61. Workers may need to re-skill or up-skill in order to adapt to the reorganisation of
tasks and the emergence of new tasks, and to weather potential job loss and navigate
transitions to new jobs. Certain workers may be more capable or better positioned to do so,
potentially exacerbating already existing inequality. For instance, workers in high-skilled
occupations may have greater ability to learn new information, tend to possess skills which
cannot be easily automated and have greater access to lifelong learning. Inequalities may
also arise if the gains of AI are captured by the owners of capital and superstar firms, rather
than workers.
4.1. Much of the impact of AI will be seen through the reorganisation of tasks
62. Many researchers share the view that the impact of AI must be understood not only
in terms of its potential to destroy jobs, but also in terms of its potential to substantially
transform the nature and content of jobs that remain (OECD, 2019[16]).24 According to
Autor (2015[56]), failing to understand this is what leads to exaggerated claims of mass
unemployment. Combining Acemoglu and Restrepo’s theoretical framework with Autor et
al.’s task-based approach, AI adoption may result in a worker being displaced from a
certain task. Rather than their entire job being eliminated entirely, there may then be a
reorganisation of tasks within the job profile (with some tasks added and others removed)
so that AI is ultimately complementary to the worker. These dynamics reflect the interplay
between the displacement, productivity and reinstatement effects, which will determine the
overall impact on AI on demand for labour (and the related impact on employment and
wages).
63. A study by Brynjolfsson et al. (2018[25]) supports the idea of the reorganisation of
tasks, as they find very few jobs that can be fully automated by machine learning. Instead,
they find that many occupations comprise both tasks with high and low suitability for
machine learning. They observe lower variance in the AI-suitability scores between
occupations than between tasks, suggesting that suitability to AI is highly sensitive to the
task being performed and that some of this sensitivity is dulled when tasks are bundled into
an occupation. In their view, this supports the idea that the unleashing of the potential of
machine learning is reliant on the reengineering of processes and the reorganisation
24
Such transformations are typical of technological change and automation: Acemoglu and Restrepo
(2018[123]) suggest that about half of employment growth between 1980 and 2015 was in occupations
in which job titles or tasks performed by workers changed. In the case of automation, OECD analysis
(2018[125]) found that while 82% of regions across Europe and North America had experienced
employment decline in occupations at high risk of automation, 60% of regions had experienced an
employment increase in occupations at low risk of automation and these increases actually
accounted for a higher share of total employment than the decline. This supports the idea that
automation shifts the mix of occupations, rather than driving down overall employment.
For Official Use

(unbundling and rebundling) of tasks.25 They argue that this reorganisation of tasks should
be the focus of the debate about the impact of AI on labour markets, instead of the focus
on full automation of occupations.
64. An empirical study of how stock analysts have reacted to the introduction of AI
(Grennan and Michaely, 2017[57]) provides a useful demonstration of how AI can produce
both substitution and complementary effects, and results in a reorganisation of tasks.
Firstly, AI produces a substitution effect as it is able to outperform high-skilled sell-side
equity analysts in prediction tasks where sufficient data are available (in part because
analysts’ predictions can be subject to bias, due to conflicts of interest). The researchers
find that analysts who cover stocks with the most data available (and therefore most suited
to AI-based prediction) are more likely to leave the profession, representing the substitution
effect. At the same time, the analysts who stay in the profession tend to shift their attention
towards the types of stocks that are less suited to AI-based prediction, indicating a further
substitution effect at the task level. The researchers find evidence that these analysts book
more meetings with management teams, suggesting that the use of AI frees up time for
them to use their interpersonal skills to gather soft information on the stocks. This
represents a reorganisation of tasks and, in the researchers’ view, the new focus of analysts
on tasks that AI cannot perform reflects the complementary of AI to high-skilled labour.26
65. Ernst et al. (2018[58]) make the point that the decision to reorganise tasks will
depend in part on workers’ ability to adapt to the redesigned jobs. Crucially, the decision
will primarily rest with the management of a company, and will depend on the profitability
of doing so27, the company’s interest in supporting workers through this transition
(e.g. through training), and institutional factors, such as the general infrastructure for
training and job-search help available in the country, tax incentives and social benefit
systems (Ernst, Merola and Samaan, 2018[58]) citing (Sengenberger, 1987[59])and (Albertini
et al., 2017[60])).
4.2. The impact of AI will also result in the creation of new tasks
66. As established in previous chapter, AI’s potential to create labour-intensive tasks

(i.e. tasks that only humans can perform) is a critical mechanism for adjustment,
counteracting the displacement effect and ensuring that the productivity benefits of AI are
shared with workers. Most directly, AI will create jobs in entirely new occupations and
fields related to its own development and deployment. However, this is not sufficient to
ensure positive labour market outcomes. AI must also create new high-productivity tasks
for human labour as noted in section 3.2.2. New jobs may also be created due to innovations
enabled by AI and to spillovers from the AI industry.
25
Brynjolfsson et al. (2017[6]) point out that firms implementing large enterprise planning systems
almost always spend several times more on redesigning business processes and on training than they
do on the hardware and software. Considerable changes to hiring and other HR practices may also
be necessary to match human capital to the new structure.
26
The researchers also find that the accuracy of earning forecasts produced by stock analysts
declines in relation to the stocks that are more suitable for AI, but this decline in quality is attributed
to compositional effects: it is the better performing stock analysts who leave the profession.
27
This in turn depends on the impact on demand for the products and services linked to these jobs
(citing (Bessen, 2018[117])).
For Official Use

4.2.1. New tasks for humans will be created in developing and deploying AI
67. As with other technologies, the deployment of AI will create jobs through its own
need for further development, maintenance, operation and regulation (PWC, 2018[43]). In a
study of companies already using or testing AI and machine-learning systems, (Wilson,
Daugherty and Morini-Bianzino, 2017[61]) identified three different types of jobs that will
emerge thanks to AI development and that will be performed by humans:
 Trainers, who will train AI systems, e.g. reducing the error rate of language
translators, tagging data in a training dataset, adapting a chatbot to mimic human
behaviour or match the culture of a company;
 Explainers, who will interpret the outputs generated by AI systems to improve
accountability, e.g. explaining how an AI system arrived at a decision, inform
decision-makers about appropriate uses of AI throughout an organisation;
 Sustainers, who will monitor the work of AI systems to ensure they are working as
intended, e.g. monitoring a CV-screening process for bias, installing content filters
in machine learning for a chatbot.
68. In certain AI systems, the need for human intervention in development and
deployment is clear. For instance, trainers and sustainers may be required to ensure that
AI-enabled sentiment analysis tools are producing accurate results. Knowing what people
mean when they are giving feedback, and not necessarily what they say in that feedback,
is a very human skill (Ultimate Software, 2018[62]). Sentiment analysis tools can classify
content as “good, neutral, or bad”, but human intervention may be needed to detect sarcasm.
69. The use of AI could also lead to the creation of brand new service activities. For
instance, Guszcza et al. (2018[63]) anticipate that emerging AI regulation could create an
industry around algorithmic auditing, wherein independent auditors ensure accountability
in the “black box” of AI. They argue that algorithm auditing should become a profession
in its own right, with proper credentialing, standards of practice, disciplinary procedures,
ties to academia, continuing education, and training in ethics, regulation, and
professionalism. Independent bodies could be formed to deliberate and issue standards of
design, reporting and conduct, regarding AI systems that companies and organisation are
developing and deploying.
4.2.2. The creation of new tasks beyond those that enable AI is crucial for
ensuring positive labour market outcomes
70. Acemoglu and Restrepo (2020[49]) point out that the jobs created in the AI industry
will not be enough to compensate for job loss if AI automates jobs in every other industry.
This situation would lead to gross inequalities across sectors and transitions would be
challenging. In their view, in order to lead to positive economic and social outcomes, the
new tasks created must extend beyond those that enable AI. They give some examples of
applications of AI that would create new high-productivity tasks for human labour:
 Individualised teaching: AI could use real-time data to determine students’
individual learning styles and problem areas and then generate recommended
teaching methods to enable teachers to teach in a way that is adapted to each student
or small subset of students. This could enhance the productivity within teaching,
benefiting society and potentially increasing demand for teachers with diverse skills
to perform the individualised teaching.
For Official Use

 High-precision production: Using AI-enabled virtual reality in production

processes could enable workers and robots to work together safely via interactive
interfaces that augment human precision and perception.
71. In addition to this, AI has some characteristics that may lead to even more job
creation than recent technological advances, specifically its potential to produce further
innovations, enable scientific breakthroughs and generate entirely new industries (OECD,
2018[64]; Gartner, 2017[65]).28 Predicting the scale of future job creation, particularly when
the jobs rely on further technological breakthroughs or entirely new applications of the
technology, is just as challenging (if not more) as predicting the scale of future job loss.
However, one study has estimated that AI could lead to a net job creation of 2 million
worldwide by 2025 (Gartner, 2017[65]).
72. The development of the AI industry may induce an even larger indirect job creation
effect. Empirical work by Moretti (2010[66]; 2012[67]) shows that the creation of jobs in the
ICT sector can have large multiplier effects in local labour markets. For each additional job
in a high tech company in a local community, five additional jobs outside high-tech are
created in the same community.
4.3. Workers may need to re-skill or up-skill in order to adapt to AI-induced

changes in the labour market
73. Workers may need to re-skill or up-skill in order to adapt to the reorganisation of
tasks and the emergence of new tasks, and to weather potential job loss and navigate
transitions to new jobs. Some may choose to acquire AI-related skills so that they can take
advantage of opportunities in AI development and deployment. However, not all jobs
where AI is complementary to human labour will require specialised AI skills. Some of
these jobs will require skills in areas that AI cannot do so well, such as creative and social
intelligence, reasoning skills, and critical thinking.
4.3.1. The demand for AI-related skills is increasing

74. A range of different skill profiles will be required in order to develop and
implement AI. Focusing at the top end of the AI talent market , the venture capital fund
MMC Ventures (2019[68]) identified that in addition to a doctoral degree in mathematics,
statistics or programming, AI professionals are increasingly expected to have
sector-specific, engineering and commercial competencies, which further limits the
potential AI talent pool. They estimate that demand for AI talent nearly doubled between
2016 and 2018 with two roles are open for every AI expert available, signalling a shortage
of skills. However, they say that supply is growing with universities and companies
increasingly providing the necessary training and with AI firms providing free educational
resources.
75. While developing novel AI concepts and techniques will generally be the domain
of those educated to doctoral level, a much broader landscape of talent will be needed to
integrate these concepts into technologies and systems (Toney and Flagg, 2020[69]). Toney
et al. suggest that demand is rising even more quickly for other types of AI talent that do
28
To demonstrate the scale of the potential impact, the OECD (2018[64]) gives the example of the
discovery of DNA structure in the 1950s leading to a revolution in industrial biotechnology and the
creation of vast economic value (the global market for recombinant DNA technology has been
estimated at around USD 500 billion).
For Official Use

not require a PhD (including for example, software engineers and database architects)
relative to AI talent that does require a PhD. The researchers show that 80% of job postings
considered AI-adjacent require a minimum education level of a bachelor’s degree, with
more than half of these stating no preference for a more advanced degree.
76. AI skills do appear to attract a pay premium. One analysis suggests that postings
demanding AI skills offer on average 11% higher salary compared to job postings with no
such demands (and compared to a 6% premium for software skills), even when controlling
for unobserved firm characteristics (Alekseeva et al., 2019[70])).
77. A firm’s need for AI professionals will likely depend on whether they are
developing AI internally or simply buying and using an AI product, and the decisions they
make around training, outsourcing and recruitment. One business survey (Ransbotham
et al., 2019[71]) suggests that companies invest in AI talent internally (rather than relying on
AI vendors) in addition to bringing in experienced AI talent from outside for technical
leadership roles, and upskill their existing workforce to be able to work with AI, in order
to capture value from AI. In another survey (Bessen et al., 2018[11]), AI startups suggest
that most of their products require only general familiarity with computers within their
client companies, claiming that the need for specialised computer skills or specific training
among the workers who will use (and be complemented by) AI is modest. One survey
(Accenture, 2018[35]), which surveyed both employers and workers, suggests that
employers tend to underestimate the willingness of employees to acquire the skills
necessary to work with intelligent technologies. In fact, a majority of the employees
surveyed are positive about the impact of AI on their work (high-skilled employees are
more positive than low-skilled employees) and consider it important to develop their own
skills.
4.3.2. As smart as AI is, there are uniquely human skills it cannot replicate
78. Not all jobs that emerge in the AI transition will require AI-related skills. There are
human skills that AI cannot replicate, which means that certain applications of AI will
require human assistance.
79. One analysis (Sage-Gavin, Vazirani and Hintermann, 2019[72]) based on the U.S.
Department of Labor’s O*Net database, which ties abilities, skills, tasks and working styles
to occupations, shows that skills such as creativity, complex reasoning, and social and
emotional intelligence are growing in importance in many jobs. Many of these skills are
the same ones that the OECD Future of Education and Skills 2030 project (OECD, 2019[73])
has encouraged individuals to develop, on the basis that AI cannot do them so well (at least
currently), for example: creativity and originality29, complex social interaction and dealing
with uncertainty. The authors point out that humans can handle uncertainty better than AI,
through their ability to develop their beliefs and understanding of what is happening in the
world, and to discard beliefs when they are inaccurate, irrelevant or damaging. AI can
complete specific tasks efficiently, and respond effectively to complexity and to some
characteristics of uncertainty, but if the goals and context of the task are ambiguous or
change, then a “breakdown” often occurs. The OECD report also highlights the importance
that lifelong learning is responsive to the changing demands of the labour market, although
29
Additionally, given the malleability of AI and its large range of applications, it is the creativity
and imagination of the human users and designers of AI that will unlock its full benefits (Berkowitz
and Miller, 2018[126]).
For Official Use

low skilled individuals and workers at high risk of automation are much less likely to be
engaged in training (OECD, 2019[16]).
80. Not only will these skills help workers to use new technologies such as AI, but
workers can also use AI to augment these skills so that human-machine collaboration can
add up to more than the sum of its parts (Sage-Gavin, Vazirani and Hintermann, 2019[72]).
Thus, enhancing skills in these areas that AI cannot emulate will allow AI to be
complementary to human labour (OECD, 2019[73]).
4.4. Certain workers will be more capable of adapting to change
81. The evidence presented in the previous chapter showed that some high-skilled
occupations involving non-routine cognitive tasks are among the most exposed to AI,
suggesting that workers in these occupations are more likely to see their work substituted,
complemented and/or transformed by AI. If these workers (including lab technicians,
engineers and actuaries) can adapt to these changes, they may be able to benefit from the
AI transition.
82. The empirical studies presented in section 3.3 (Felten, Raj and Seamans, 2019[24];
Fossen and Sorgner, 2019[32]) found larger positive wage effects of AI adoption among
individuals with higher educational attainment and in higher wage occupations, suggesting
that these individuals have been most able to use AI in a complementary fashion. At the
same time, Acemoglu et al. (2020[54]) find that individuals in lower wage occupations have
been more likely to be substituted by AI. Fossen and Sorgner also found that the positive
wage effects were stronger for older, more experienced workers, suggesting that these
workers are more able to use AI to complement their own labour and boost their
productivity. In this case, AI adoption could increase inequality.
83. Explaining why workers in high-skilled occupations might be more capable of
adapting to such changes, Fossen and Sorgner (2019[32]) point to their greater ability to
learn new information and adapt to new technologies, as well as their tendency to possess
skills which cannot be easily automated, such as creative and social intelligence, reasoning
skills, and critical thinking. They suggest that the potential for AI to substitute labour
depends on the extent to which an occupation consists of non-routine cognitive tasks, which
is more likely in high-skilled occupations in their view.30
84. It has also been suggested that adjustments to white-collar jobs due to AI might be
slower than adjustments to blue-collar jobs due to automation (Wright, 2019[74]), because
of the greater need to adjust processes around reporting or controls, the value attached to
relationships and to expert judgment in complex decision-making, and the unlikeliness of
a situation in which an entire department (e.g. accounting) is dismissed all at once. The
article provides the example of Zurich Insurance piloting the use of machine learning in
the assessment of insurance claims for car crashes or burglaries. Zurich Insurance
ultimately decided not to roll the pilot out in full due to the frequency with which humans
had to step in to override the computer’s decision.
30
Agarwal et al. (2019[5]) present a similar argument in relation to the impact of machine learning
but one that does not tie the impact to skill level. Instead, the impact of machine learning on an
occupation is related to the extent to which the core skill of the occupation is prediction. In their
view, the HR professional whose core skill is screening CVs for potential hires may find the value
of this skill diminished whilst a surgeon using AI imaging to operate more precisely on a tumour
may find the value of their operating skills enhanced.
For Official Use

85. Some expect AI advancements in problem solving, logical reasoning and

perception to open up specialised AI tools to non-experts, thus complementing low-skilled
workers while substituting high-skilled workers, e.g. agricultural systems that provide
guidance what and when to plant, and construction systems that plan maintenance activities
(Ernst, Merola and Samaan, 2018[58]). A similar example is the use of AI-enabled
on-the-spot training programmes enabled by glasses or tablets with virtual or augmented
reality functionality in manufacturing processes that require the worker to adapt to each
specific order (Moore, 2019[75]). While the use of these devices means that workers need
less pre-existing knowledge or training, Moore points out that this also means that workers
do not learn long-term skills and do not specialise.
86. There may be a mix of factors influencing older workers’ ability to adapt. On the
one hand, the idea of age-biased technological change, whereby the adoption of technology
disadvantages older workers, is well established ( (Behaghel, Caroli and Roger, 2011[76])
and elsewhere). Webb (2020[23]) points out that older workers are likely to find it more
difficult to adapt to changes in the labour market due to AI. They are generally less mobile
and may have less incentive to retrain (due to fewer remaining years of working life). On
the other hand, Webb considers that the impact of AI on employment may be felt via the
entry margin rather than the exit margin (i.e. fewer young workers joining the occupation,
rather than more older workers leaving it), which would result in a higher impact on
younger workers.
4.5. Does AI favour owners of capital and superstar firms?
87. Just as changes in relative demand for different groups of workers can result in
inequality, so can changes in relative demand for different factors of production. When AI
replaces labour in a production process and complements other factors of production (say,
land or capital), what workers lose in terms of wages, the owners of land and capital gain
as a kind of unearned windfall (Korinek and Stiglitz, 2017[77]).
88. Korinek and Stiglitz (2017[77]) identify a further channel through which AI (and
other technologies) may increase inequality. Specifically, the excludable nature of
technology (e.g. through intellectual property rights) may enable innovators to build up
market power in a “winner-takes-most” dynamic31 (potentially to the point where they
reach superstar status, as mentioned in previous section), which they can then use to earn
a surplus in excess of the costs of innovation. This would create not only inequality among
innovators32, but also between innovators and workers.
89. Since technological progress can increase productivity and drive economic growth,
there is reason to explore how AI can be implemented in an inclusive way, without
increasing inequalities and societal resistance to technological progress. Korinek and
Stiglitz (2017[77]) consider the scope for redistribution in a situation where AI is introduced
as a labour-replacing technology, producing winners and losers (via the channels described
31
Schwellnus et al. (2018[40]) demonstrate the existence of the “winner-takes-most” dynamic but
attribute it to technological dynamism, rather than anti-competitive forces.
32
Inequalities could also arise between workers in superstar firms and workers in firms struggling
to keep at pace with digitalisation (OECD, 2019[16]).
For Official Use

above). Korinek and Stiglitz find that redistribution33 (e.g. via a tax on capital inputs) can
still produce a Pareto improvement whereby all parties are better off as a result of AI (even
in the case of a singularity), as long as the costs of redistribution are sufficiently low.
Korinek and Stiglitz also point out that technology that is privately optimal may not be
socially optimal, and hint at potential intervention in the innovation process in order to
make AI less likely to substitute labour.
33
Korinek and Stiglitz suggest achieving redistribution via a tax on capital inputs that earn windfall
gains (particularly on those that are fixed in supply, such as land or energy, as they would be less
distortionary). If this type of redistribution is not feasible, Korinek and Stiglitz suggest shortening
the term of patent protection, resulting in the lower prices for consumers/workers after expiration of
the patent and thereby some redistribution of the innovation surplus.
For Official Use

5. AI and the work environment
90. Work not only rewards people financially, it also provides them with a chance to
fulfil their ambitions, feel useful in society and build self-esteem. Work can be either good
or bad for health, fulfilling or meaningless, depending on factors such as the tasks
completed, work organisation and management practices, amongst others.34 AI has the
potential to reshape the work environment of many people, by changing the content and
design of their jobs, the way they interact with each other and with machines, and how
work effort and productivity are monitored. Deployment of AI-enabled technologies in the
workplace is still at an early stage, and at this moment, it is an open question whether AI
will improve or worsen the work environment overall, and how this might differ across
different types of AI, different workers and different modes of implementation.
91. The quality of the work environment is an important element of overall job quality,
in addition to earnings and labour market security (OECD, 2018[78]). The concept refers to
the setting in which workers operate, reflecting the interplay between job demands such as
physical risk factors, emotional demands and work intensity, and job resources such as task
organisation, work autonomy, learning opportunities, workplace relationships and good
management practices (OECD, 2017[79]).
92. This chapter first examines how the adoption of AI and the reorganisation of tasks
can improve or worsen the work environment, to the extent that it steers workers towards
or away from safe and fulfilling tasks. The discussion then moves to explore how
AI-enabled technologies might change the interaction between worker and machine via the
introduction of collaborative robots (or cobots).
93. This chapter also explores the multiple potential applications of AI in the field of
human resource management and recruitment, enabling workers to advance their careers,
helping managers to manage, and supporting career guidance and training. These
applications demonstrate that AI can complement human capabilities and support
management practices and career development. However, some of the same features that
make AI algorithms so powerful in these respects may diminish the quality of the work
environment in other ways or for other workers, generating new concerns as well as
amplifying existing concerns. This chapter discusses how the collection and use of vast
amounts of data and the lack of transparency and low explainability due to complex inner
workings, may generate stress and increase psychosocial risks at work. Even AI
technologies with potential to improve the work environment may have the opposite effect
if applied badly or with the sole objective to increase productivity (even if driven by
genuine business need) at the expense of other factors.
34
This is well documented in various strands of literature, from occupational health studies that
identify major risks for both physical and mental health in the workplace, to occupational
psychology and behaviour research, to people management research which follows a more positive
perspective under the basic premise that people are more productive when they enjoy their work and
workplace (Saint-Martin, Inanc and Prinz, 2018[82]).
For Official Use

5.1. The jury is out on whether the reorganisation of tasks as a result of AI adoption
improves the work environment
94. One of the most direct ways AI can affect the quality of the work environment is
through the automation of tasks and the resulting reorganisation of tasks within an
occupation. To the extent that AI can facilitate the automation of hazardous, repetitive or
demeaning tasks and steer workers toward safer and more fulfilling ones, it can enhance
the work environment. On the other hand, if the reorganisation has the effect of removing
safe and fulfilling tasks from workers, the work environment will deteriorate.
95. One survey of workers in Japan (Yamamoto, 2019[80]) suggests that the
reorganisation of tasks in the wake of AI adoption contributes both to greater job
satisfaction and increased stress. The authors suggest that AI allows workers to concentrate
on more complex tasks that can only be performed by humans. These more complex tasks
may intensify work-related stress but may also provide a greater sense of satisfaction once
accomplished.
96. Occupational Safety and Health (OSH) agencies have already identified the
potential for robots to replace workers in strenuous work activities and dangerous work
environments (e.g. chemical or ergonomic hazards), thereby reducing OSH risks (EU-
OSHA, 2018[81]). A new generation of robots, powered by AI systems (instead of traditional
programming algorithms) opens up even more opportunities. Many people are still exposed
to risk factors for physical health at work, even though many tasks have been automated
that formerly required hard physical labour (Saint-Martin, Inanc and Prinz, 2018[82]). Of
particular concern is the high prevalence of risk factors for work-related musculoskeletal
injuries in manufacturing and construction, but also in a number of service activities such
as healthcare and retail.
97. In manufacturing and warehousing environments, AI may offer some solutions, but
may also introduce some new risks. First, AI-enabled smart robots can perform a much
wider range of tasks, including some physically onerous tasks that less advanced
technologies left to human workers. Second, even when full automation may not be
feasible, smart robots (including the cobots described in the next section) can work
alongside operators to reduce the health consequences of physical efforts, repetitive
movements or awkward postures, which are key risk factors for musculoskeletal injuries.
However, the same technologies may produce psychosocial risks if people are driven to
work at the robot’s pace, as well as physical risks due to potential collisions. In addition,
there may be questions about liability and responsibility in the case of injury or damage
(Moore, 2019[75]).35
98. As established in previous chapters, AI’s problem solving, logical reasoning and
perception capabilities could lead to high-skilled workers (including lab technicians,
engineers and actuaries) seeing substantial change in the tasks they perform and to their
work environment. However, there does not appear to be much discussion in the literature
35
A set of case studies (Jaehrling et al., 2018[107]) illustrates some additional risks, in this case
following the introduction of an automatic sorting system in warehouses. The resulting
reorganisation of the job of “manual picking” into two jobs ("feeding” and “palletising”) made tasks
less varied and work more repetitive as the tasks left to workers were those routine tasks that were
too costly to automate. Workers expressed a sense of alienation, as if they were a “mere appendage
of a machine”. Even though the automatic sorting systems covered by the case studies did not appear
to have involved AI, similar results might be expected with systems that do rely on AI.
For Official Use

on how the reorganisation of tasks due to AI for high-skilled workers could improve or
deteriorate the quality of their work environment. One rare example is the discussion by
Jha and Topol (2016[83]) of the potential for visual fatigue among radiologists as their
profession becomes more data-rich due to the requirements of AI technology, and as their
activities depend less on inference and more on detection. This gap in the literature may be
due to lower levels of concern about the quality of the work environment among those in
higher skill jobs (who are, on average, more likely to work in high quality working
environments (OECD, 2017[79])). Those highly skilled workers who are exposed to AI may
also have greater say in how AI is adopted, given that they are more likely to possess
specialist knowledge essential for the functioning of the organisation, implying that the
withdrawal of their co-operation could be costly.
5.2. AI can promote close human-robot collaboration
99. One factor that has traditionally limited human-robot collaboration in

manufacturing or warehousing environments is the physical danger associated with humans
and robots sharing the same space. Some point to AI-enabled technologies as a way to
allow humans and robots to work in close collaboration while safeguarding the health and
well-being of workers (Daugherty and Wilson, 2018[84]; [85]). One example36 is AI-enabled
cobots.
100. Cobots allow firms to combine a robot’s strength and endurance with a human’s
tacit knowledge and agile decision-making (Knudsen and Kaívo-Oja, 2020[86]), thereby
complementing and augmenting human capabilities (rather than replacing them)
(Daugherty and Wilson, 2018[84]; [85]), and enhancing performance compared to purely
robotic processes. Collaborative robotics has been described as one of the fastest-growing
sectors of the robotics market (Goldberg, 2019[87]). Villani et al. (2018[88]) identify the
industrial applications where cobots are most advantageous, according to the literature:
handling; welding; assembly; and applications in the automotive industry (where demand
is currently the greatest). In these applications, cobots generally assist the operator by
performing mundane and or physical tasks such as moving materials, holding heavy objects
or performing sample tests.
101. Despite these advantages, when AI-enabled cobots work in close proximity to a
worker, new physical and psychosocial risks are likely to arise. This is why robots have
typically been sectioned off from human workers in industrial environments, with little
physical interaction.37 Indeed, collaborative robotic assembly tasks have been shown to
produce mental strain as evidenced by the monitoring of psychological and physiological
(e.g. sweating) responses, which are more pronounced when the cobot is within 2 metres
of the worker and moves quickly and without warning (Arai, Kato and Fujita, 2010[89]).
36
Another example, mentioned in section 4.2.2, is the use of AI-enabled virtual reality in production
processes enabling workers and robots to work together safely via interactive interfaces that augment
human precision and perception (Acemoglu and Restrepo, 2020[49]).
37
For decades, robots have typically been large pieces of machinery, usually sectioned off from
human workers, that would perform a dedicated task – unloading a stamping press, for example
(Daugherty and Wilson, 2018[84]). That specific task was part of a rigid, fixed chain of work that
would generally include humans doing other predefined tasks – for instance, inspecting the stamped
metal parts in order to discard defected parts.
For Official Use

102. However, some believe that these risks can be mitigated through the use of
intelligent navigation systems. Cobots can be equipped with AI-powered intelligent
navigation systems, which allow them to navigate a shared workspace with the human
element in mind (e.g. human field of vision and human motion patterns) (Lasota and Shah,
2015[90]). The aim is not only to avoid collision but also to achieve fluid human–robot
interactions so that the individual feels safe and comfortable working with a robot
teammate. In one experiment, Lasota and Shah (2015[90]) showed that individuals felt more
satisfied, safer and more comfortable working in a shared workspace with an AI-enabled
robot than with a standard robot, in addition to performing the task more quickly38.
103. While close collaboration facilitated by cobots may enhance performance in certain
settings, there is also concern about increased work intensity if the worker is driven to work
at the robot’s pace rather than the reverse (Moore, 2019[75]). As more autonomy is given to
the cobot, there are questions about whether this diminishes the operator’s task discretion
and autonomy. These factors have been shown to be closely associated with workers’ job
satisfaction, physical and psychological well-being, and to act as a buffer against the
damaging effect of high work intensity ( (OECD, 2017[79]) citing ( (Karasek, 1979[91];
Karasek and Theorell, 1990[92])).
104. Some experiments have suggested that there may be particular circumstances in
which a cobot equipped with an AI-powered plan execution system is well suited to act as
a team leader and to allocate tasks among human team members. One experiment
(Gombolay et al., 2015[93]), in which 24 pairs of human participants collaborated with a
cobot equipped with a plan execution system to build Lego kits, found that participants
preferred when the cobot was given full authority over team coordination compared to
when one participant was given authority or authority was shared with the cobot. The
results surprised the researchers who expected that participants would prefer to keep some
control over team coordination. It could be that the participants found the process of
scheduling to be burdensome, especially given the tight deadlines, and preferred to be part
of an efficient team rather than play a role in the team coordination process.
105. The potential of AI to enhance performance by enabling close human-robot
collaboration through cobots has been an active area of research and development. A
substantial amount of this literature has examined the impact of human-robot collaboration
on operators’ feelings of trust, safety and stress (e.g. (Villani et al., 2018[88]), (TNO
(Netherlands Organisation for Applied Scientific Research), 2018[94]), (Hancock et al.,
2011[95])). For instance, the literature review performed by Villani et al. (2018[88]) identifies
safety issues as the primary challenges for developing cobots and calls for developers to
prioritise safety over performance. They suggest that certain advances in collaborative
robotics that would improve human-robot interaction (e.g. human-friendly interfaces) have
not yet transferred from the laboratory to industrial settings. As such, it is not clear that the
promise of cobots to safeguard the health and well-being of workers is yet being fully
realised.
5.3. AI can support human resource management and career development
106. AI has a few features that make it particularly attractive to firms that wish to employ
more modern, participative and engaging human resource management models and to
workers wishing to advance their careers. However, potential benefits for workers are very
38
The study did not make comparisons with a work environment consisting of only human workers.
For Official Use

much dependent on how employers will use these new technologies, which have the
potential to collect and produce vast amount of data on work performance, and therefore,
may increase work pressure (see section 5.4). With these caveats in mind, Table 5.1
presents some applications that demonstrate how AI can enable personalised coaching for
individuals, help managers to manage, match skills to jobs, and improve training tools and
programmes.
Table 5.1. AI applications in Human Resource Management (HRM) and career development
Application What does it offer? How does it use AI?

Humu's Nudge Ongoing and personalised coaching for Uses machine learning to customise hints and
Engine employees and managers using insight suggestions. e.g. what timing, messaging, and
from behavioural sciences and motivational techniques are effective for each
occupational psychology employee
Dristi's video-enabled Increases efficiency along assembly lines Image recognition software converts videos of
productivity by identifying weaknesses and workers along the assembly line to data points for use
monitoring inefficiencies on the factory floor in a user-friendly app
Humanyze’s Employees wear "sociometric” badges AI-enabled analytics identify weaknesses in
wearable devices that record communication patterns so communication patterns. e.g. based on tone of voice,
that feedback on teamwork can be given gesturing, frequency of interruption
Image recognition software analyses facial
HireVu's facial Analyses recorded interviews to make the expression, body language and word choice filter to
recognition software hiring process more efficient assist with selection
IBM's predictive Predicts whether an employee is likely to IBM's AI platform, Watson, makes predictions based
attrition program leave the company so that managers can on a wide variety of data points
take strategic action
Answers questions about the army's NLP-enabled chatbot speaks with potential recruits
U.S. Army's chatbot recruiting process and refers users to using the same language and style as an army
recruiter human recruiters when necessary recruiter
Jobiri's resume Builds CVs and cover letters for AI algorithm evaluates CVs and provides customised
builder jobseekers feedback
Instant Coach Flight Enables individuals to practice skills in A chatbot provides a private setting for individuals to
Simulator between training sessions practice skills, including difficult, dangerous or
embarrassing ones
IBM's Blue Matching Delivers personalised internal job Predictive analytics generate recommendations that
recommendations to IBM employees on a fit employees' qualifications, aspirations and AI-
voluntary basis inferred skills
Note: This table is intended to illustrate a variety of applications of AI in HRM and career development tools
and is not intended to be comprehensive. Information in the table was collected from developers’ websites.
107. AI’s ability to process large amounts of data and learn in real time could make it
particularly suitable for supporting continuous feedback and development (e.g. Ultimate
Software's UltiPro Perception). Rather than waiting for the results of an annual employee
engagement survey, regular feedback helps managers take timely actions to ensure that
employees’ needs and goals are being met (Nielsen et al., 2016[96]; Kark, Van Dijk and
Vashdi, 2018[97]).
108. AI’s capabilities in the collection and analysis of data may make it an attractive tool
to support management decision-making, from monitoring worker productivity
(e.g. Dristi’s video-enabled solution) to communications (e.g. Humanzye’s wearable
devices). Human resource management (HRM) professionals may be attracted to
AI-enabled technologies on the basis that they offer a more data-driven approach to
recruitment (e.g. HireVu’s facial recognition software) or retention (e.g. IBM's predictive
attrition program), or that they are time-saving (e.g. US Army chatbot recruiter).
For Official Use

109. AI may offer advantages for workers who wish to advance their careers, from
helping jobseekers to draft their CVs (Jobiri's resume builder) to enhancing training
programmes (Instant Coach Flight Simulator). AI-enabled data analytics (such as IBM's
Blue Matching) could improve matching between skills and jobs, which has been shown to
be a key driver of job satisfaction and performance (Saint-Martin, Inanc and Prinz,
2018[82]).
110. Some AI applications complement sophisticated data analytics with insights from
neuroscience, behavioural science and applied research in various fields, such as
communication and organisational psychology (such as Humu's Nudge Engine). These
kinds of knowledge can help develop a positive workplace culture, but all too often, they
remain outside the range of skills and competencies that employers are focusing on to run
their core business activities. In this regard, AI-enabled technologies may offer cost
efficiencies and scalability relative to a human coach (Bersin, 2018[98]).
111. For companies wanting to improve workplace culture and the quality of the work
environment, such technologies can be particularly appealing. However, they may have the
opposite effect if they enable excessive monitoring on behalf of the firm, undermine data
privacy or lack transparency and explainability (as discussed in section 5.4).
112. Evidence on the current prevalence of AI applications within HRM functions is
mixed but points to the potential for this to grow in the future. One report (PWC, 2018[99])
suggests that 40% of HRM functions in international companies (mostly US-based) are
currently using AI applications, mostly for the recruitment and hiring process. HR
practitioners participating in a roundtable (Mathis, 2018[100]) agreed that this was the main
application, while a few also mentioned using AI to customise e-learning content.
The researchers note that HRM applications represent a very small share of the total
investment in AI development, compared to other industries such as health, robotics and
marketing, sales and customer relationship management (citing (HRWins, 2018[101])) but
that the number of HR-related AI sales increased considerably from 2013 to 2018 (citing
(CB Insights, 2018[102])). Another recent survey across various industries in more than
30 countries showed that a vast majority of HR professionals surveyed believed that AI
could improve internal matching and visibility of opportunities (Zhang, Feinzig and
Hemmingham, 2018[103]). However, the same survey suggested that adoption levels are still
modest: almost two thirds of HR professionals surveyed had not yet adopted such tools
while only 6% of them report using AI solutions moderately or to a great extent.
5.4. AI may also entail risks for the work environment
113. Some of the same features that make AI algorithms so powerful may also entail
risks for the quality of the work environment. Excessive monitoring may generate
psychosocial risks, increasingly recognised as an important component of occupational
safety and health (Leka and Jain, 2010[104]; ILO, 2016[105]). Concerns about data privacy,
transparency and explainability may exacerbate these risks, in addition to raising questions
about the ethics39 of introducing AI to the workplace. AI can also amplify some of the
psychosocial risks associated with digitalisation more generally40 as AI is embedded in
39
A forthcoming OECD publication separately explores these ethical questions in more detail. This
literature review focuses on the impact on the work environment and the quality of jobs only.
40
e.g. stress, discrimination, heightened precariousness, musculoskeletal disorders, and the
possibilities of work intensification and job losses
For Official Use

pre-existing tools or new tools for workplace management and design (Moore, 2019[75]).
However, Moore notes a general lack of discussion in high-level governmental and
organisational reports about the implications for OSH of introducing AI into the workplace.
114. Even those who warn of the risks associated with the use of AI in the workplace,
point out that the use of such tools to supervise work activities is not necessarily harmful
(Moore, 2019[75]; De Stefano, 2018[106]). According to Moore (2019[75]), it is the
implementation rather than the technology itself that creates a negative or positive impact
on working conditions. Jaehrling et al. (2018[107]) note the importance of mediating factors
such as management attitudes and perceptions as well as employee bargaining power, in
determining the impact of technological innovations on the work environment.
5.4.1. Enable excessive monitoring on behalf of the firm

115. Excessive monitoring of employees, in the form of data collection and processing,
may cause stress and undermine well-being. Surveillance at work is not necessarily new41,
but AI tools can only exacerbate the situation, not least because it is the very way those
tools perform – every bit of data is potentially valuable (Van den Broek, 2017[108]).
116. Many of the HRM-related applications mentioned in Table 5.1 require additional
data to be collected in the workplace. Even cobots – which are not intended to monitor
employees’ behaviours but rather are geared towards helping them execute work tasks –
produce a myriad of granular data on work performance (e.g. workers’ and machines’ idle
times). In addition, some innovative approaches to put human psychosocial risks at the
heart of human-robot collaboration involve the operator wearing a smart-watch that
monitors stress levels (Landi et al., 2018[109])
117. Even if effectively anonymised and aggregated, data collection can be highly
invasive and may capture personal elements, including the level of interaction with
colleagues and the mood of workers (De Stefano, 2018[106]). Moore (2019[75]) describes the
potential for AI tools to increase the degree of monitoring in call centre work, already
considered repetitive, demanding and subject to high levels of monitoring (citing
(Woodcock, 2016[110])). While AI tools to analyse sentiment or facial expressions could be
employed to identify and combat overwork or stress, the requirement for invasive data
collection could equally exacerbate the issue of monitoring and create new sources of
stress.
118. How AI tools impact managerial practices is also important. If these tools are used
to implement micromanagement and other practices that increase pressure on workers, they
may cause stress and anxiety (Moore, 2019[75]), and may even cause efficiency and
productivity to decline ( (De Stefano, 2018[106]) citing (Moore, Akhtar and Upchurch,
2018[111])).
5.4.2. Data privacy and protection issues: is “Big Brother” watching you?
119. Data privacy and protection issues come to the fore where AI-enabled technologies
rely on data in the individual’s private sphere. For instance, sentiment analysis may rely on
capturing and analysing a worker’s written exchanges, from e-mails, to instant messages,
to blogs and communications on various social networks.
41
For instance, computer monitoring that measures employee keystroke speed and accuracy, or
video surveillance that detects safety issues but also employee misconduct, have been used for some
time now, well before the deployment of AI technologies in the work environment.
For Official Use

120. Data privacy and protection issues cut across all AI-enabled technologies, from
cobots to individual virtual coaches, to smart apps and platforms that support
decision-making in sensitive areas of HR management, such as hiring and performance
management processes. Collection, storage, processing and analysis of large amounts of
data are the very essence of these technologies. In a recent white paper (2020[112]), the
European Commission stated its intention to examine on a continuous basis any additional
risks posed by AI systems and the application of General Data Protection Regulation
(GDPR) (European Commission, 2018[113]) to these risks.
5.4.3. Lack of transparency, explainability and fairness

121. Many of the concerns discussed above can be exacerbated when there is no
transparency about what AI technologies are being used in the workplace and how they
feed into decision-making. This lack of transparency and explainability could result in the
misuse of AI and the use of inaccurate and/or biased AI; it could also reinforce ethical and
privacy concerns and pose a barrier for workers wishing to have a say on how AI is used
in the workplace. AI has the potential to produce results that are inaccurate and/or biased
and therefore lead to unfair and discriminatory decisions.
122. HR professionals may be attracted to AI-enabled technologies on the basis that they
will overcome individual biases of supervisors and replace them with more objective and
neutral metrics, but may find instead that one set of biases has been replaced by another
(De Stefano, 2018[106]). For instance, if a system learns which job applicants to accept for
an interview by using a dataset of decisions made by human recruiters in the past, it may
inadvertently learn to perpetuate their racial, gender, ethnic, or other biases (Brynjolfsson
and McAfee, 2017[31]). For instance, in 2018, Amazon had to scrap an experimental
AI-enabled recruitment tool, which was discovered to be biased against female candidates
(Dastin, 2018[114]). Due to the prevalence of men already in technical roles, the system
downgraded CVs that mentioned all-women’s colleges or female-indicating phrases like
“women’s chess club.” A recent study shows that, to find the best workers (and
simultaneously overcome bias, even if this may not be a specific goal), hiring algorithms
must find a balance between selecting from groups with proven track records and selecting
from under-represented groups to learn about quality (Li, Raymond and Bergman,
2020[115]).
123. Similarly, AI-enabled performance management systems shaped by subjective
assessments of intangibles, such as engagement and cultural fit, risk introducing racial
stereotypes into decision-making and harming diversity (Bodie et al., 2017[116]). Moreover,
these biases may not appear as an explicit rule but, rather, be embedded in subtle
interactions among the thousands of factors considered, so that diagnosing and correcting
the problem can be a challenging task (Brynjolfsson and McAfee, 2017[31]).42 De Stefano
points to the importance of workers being able to “negotiate the algorithm” (including
through collective bargaining) when AI is used in performance management systems, a task
which is more challenging when transparency is low.
42
Technical solutions do appear to be emerging, like the Local Interpretable Model-Agnostic
Explanations (LIME) (Daugherty and Wilson, 2018[84]). For instance, if an expert HR system has
identified the best candidate for a particular job, LIME can identify the variables that led to that
conclusion (such as education and deep expertise in a particular narrow field) as well as the evidence
against it (such as inexperience in working on collaborative teams).
For Official Use

124. Employers need to learn how to correctly use the data provided by monitoring
technologies, as information should be placed in context before proper sense-making can
be applied (Van den Broek, 2017[108]). Even though today’s technologies are getting
increasingly smarter, context remains very hard to grasp and interpret for a machine.
Therefore, when combined with algorithmic decision-making, data-driven performance
management may increase psychosocial risks and stress significantly, unless it is backed
up with ethical consideration and adequate human intervention (Moore, 2019[75]). Risks of
stress and anxiety arise when workers feel that decisions are being made through automated
processes based on numbers and data that they have neither access to nor control over. This
may generate great uncertainty as to the accuracy and fairness of the decision-making
process, especially if the latter determines promotion and remuneration, job description
changes, and hiring and firing.
125. For companies operating in the EU area, the GDPR already establishes the right for
people not to be subject to an automated individual decision-making (including profiling
with respect to the individual’s performance at work), therefore requiring some forms of
human intervention in decision-making processes. Guszcza et al. (2018[63]) anticipate that
emerging regulations might lead to the creation of brand new service activities such as
algorithmic auditing. According to Guszcza et al., it may require no less creativity, hard
work, and innovation for companies to improve transparency and explainability around
AI-enabled technologies (including those that affect their employees) than to develop the
AI technologies themselves.
For Official Use

References
Accenture (2018), Reworking the Revolution, https://www.accenture.com/_acnmedia/PDF-69/Accenture- [35]

Reworking-the-Revolution-Jan-2018-POV.pdf (accessed on 9 June 2020).
Accenture (2017), How AI boosts industry profits and innovation, Accenture, [45]
https://www.accenture.com/fr-fr/_acnmedia/36DC7F76EAB444CAB6A7F44017CC3997.pdf
(accessed on 27 March 2020).
Acemoglu, D. et al. (2020), AI and Jobs: Evidence from Online Vacancies, [54]
https://conference.nber.org/conf_papers/f143876.pdf (accessed on 30 September 2020).
Acemoglu, D. and P. Restrepo (2020), “The Wrong Kind of AI? Artificial Intelligence and the Future of [49]
Labor Demand”, Cambridge Journal of Regions, Economy and Society, Vol. 13/1, pp. 25-35,
http://dx.doi.org/10.1093/cjres/rsz022 (accessed on 19 February 2020).
Acemoglu, D. and P. Restrepo (2019), “Automation and New Tasks: How Technology Displaces and [48]
Reinstates Labor”, Journal of Economic Perspectives, Vol. 33/2, pp. 3-30,
http://dx.doi.org/10.1257/jep.33.2.3.
Acemoglu, D. and P. Restrepo (2018), Artificial Intelligence, Automation and Work, NBER, [13]
http://www.nber.org/papers/w24196 (accessed on 20 February 2020).
Acemoglu, D. and P. Restrepo (2018), “Low-Skill and High-Skill Automation”, Journal of Human [122]
Capital, Vol. 12/2, pp. 204–232, https://www.journals.uchicago.edu/doi/full/10.1086/697242 (accessed
on 30 March 2020).
Acemoglu, D. and P. Restrepo (2018), “The race between man and machine: Implications of technology [123]
for growth, factor shares, and employment”, American Economic Review, Vol. 108/6, pp. 1488-1542,
http://dx.doi.org/10.1257/aer.20160696.
Aghion, P., C. Antonin and S. Bunel (2020), “Artificial Intelligence, Growth and Employment: The Role [51]
of Policy”, Economie et Statistique / Economics and Statistics 510-511-512, pp. 149-164,
http://dx.doi.org/10.24187/ecostat.2019.510t.1994.
Aghion, P. et al. (2017), “Artificial Intelligence and Economic Growth”, NBER Working Paper Series, [9]
No. 23928, NBER, http://www.nber.org/papers/w23928 (accessed on 3 February 2020).
Agrawal, A., J. Gans and A. Goldfarb (2019), “Artificial Intelligence: The Ambiguous Labor Market [5]
Impact of Automating Prediction”, The Journal of Economic Perspectives, Vol. 33/2, pp. 31-50,
http://dx.doi.org/10.2307/26621238.
Albertini, J. et al. (2017), “A Tale of Two Countries: A Story of the French and US Polarization”, IZA [60]
Discussion Papers, https://www.iza.org/publications/dp/11013/a-tale-of-two-countries-a-story-of-the-
french-and-us-polarization (accessed on 19 August 2020).
Alekseeva, L. et al. (2019), The Demand for AI Skills in the Labor Market, [70]
http://dx.doi.org/10.2139/ssrn.3470610 (accessed on 2 November 2020).
For Official Use

American Economic Association & Committee on Economic Statistics (ed.) (2020), Measuring [10]
Technology Adoption in Enterprise-Level Surveys: The Annual Business Survey,
https://www.aeaweb.org/conference/2020/preliminary/powerpoint/7rdEN8G6 (accessed on
2 October 2020).
Analysis Group (2016), Global Economic Impacts Associated with Artificial Intelligence, [46]
https://www.analysisgroup.com/Insights/publishing/global-economic-impacts-associated-with-
artificial-intelligence/ (accessed on 27 March 2020).
Andrews, D., C. Criscuolo and P. Gal (2016), “The Best versus the Rest: The Global Productivity [36]
Slowdown, Divergence across Firms and the Role of Public Policy”, OECD Productivity Working
Papers, No. 5, OECD Publishing, Paris, https://dx.doi.org/10.1787/63629cc9-en.
Arai, T., R. Kato and M. Fujita (2010), “Assessment of operator stress induced by robot collaboration in [89]
assembly”, CIRP Annals - Manufacturing Technology, Vol. 59/1, pp. 5-8,
http://dx.doi.org/10.1016/j.cirp.2010.03.043.
Autor, D. (2015), “Why are there still so many jobs? the history and future of workplace automation”, [56]
Journal of Economic Perspectives, Vol. 29/3, pp. 3-30, http://dx.doi.org/10.1257/jep.29.3.3.
Autor, D. (2014), “Polanyi’s Paradox and the Shape of Employment Growth”, NBER Working Papers, [30]
No. 20485, NBER, http://dx.doi.org/10.3386/w20485.
Autor, D., F. Levy and R. Murnane (2003), “The Skill Content of Recent Technological Change: An [14]
Empirical Exploration”, The Quarterly Journal of Economics, Vol. 118/(4), pp. 1279-1333,
https://economics.mit.edu/files/11574 (accessed on 14 February 2020).
Autor, D. and A. Salomons (2018), Is automation labor-displacing? Productivity growth, employment, and [119]
the labor share, https://www.brookings.edu/wp-content/uploads/2018/03/1_autorsalomons.pdf
(accessed on 20 November 2018).
Behaghel, L., E. Caroli and M. Roger (2011), “Age Biased Technical and Organisational Change, Training [76]
and Employment Prospects of Older Workers”, IZA Discussion Papers,
https://www.iza.org/publications/dp/5544/age-biased-technical-and-organisational-change-training-
and-employment-prospects-of-older-workers (accessed on 10 August 2020).
Berkowitz, M. and K. Miller (2018), AI, Attitudes and Values and Ethics, OECD Publishing, Paris, [126]
http://www.oecd.org/education/2030-project/about/documents/Education-and-AI-preparing-for-the-
future-AI-Attitudes-and-Values.pdf.
Bersin, J. (2018), Employee Engagement 3.0: Humu Launches Nudge Engine, [98]
https://joshbersin.com/2018/10/employee-engagement-3-0-humu-launches-nudge-engine/.
Bessen, J. (2018), “Automation and Jobs: When Technology Boosts Employment”, Boston University [117]
School of Law, Law and Economics Research Paper, Vol. No. 17-09,
https://scholarship.law.bu.edu/faculty_scholarship/815 (accessed on 1 April 2020).
Bessen, J. (2015), “How Computer Automation Affects Occupations: Technology, Jobs, and Skills”, SSRN [120]
Electronic Journal, http://dx.doi.org/10.2139/ssrn.2690435.
Bessen, J. et al. (2018), “The Business of AI Startups”, Law and Economics Research Paper, Boston [11]
University School of Law, https://scholarship.law.bu.edu/faculty_scholarship/537 (accessed on
8 June 2020).
For Official Use

Bodie, M. et al. (2017), “The Law and Policy of People Analytics”, University of Colorado Law Review, [116]
Vol. 88, https://heinonline.org/HOL/Page?handle=hein.journals/ucollr88&id=997&div=&collection=
(accessed on 13 August 2020).
Bostrom, N. (2006), “How long before superintelligence?”, Linguistic and Philosophical Investigations, [17]
Vol. 5/1, pp. 11-20, https://www.nickbostrom.com/superintelligence.html (accessed on 3 April 2020).
Bresnahan, T. and M. Trajtenberg (1992), “General Purpose Technologies “Engines of Growth?””, NBER [4]
Working Paper, NBER, http://dx.doi.org/10.3386/w4148.
Brynjolfsson, E. and A. McAfee (2017), “The Business of Artificial Intelligence - What it can and cannot [31]
do for your organization”, Harvard Business Review, https://hbr.org/2017/07/the-business-of-artificial-
intelligence?ab=seriesnav-bigidea.
Brynjolfsson, E. and A. McAfee (2014), The second machine age : work, progress, and prosperity in a [21]
time of brilliant technologies, WW Norton & Company.
Brynjolfsson, E., T. Mitchell and D. Rock (2018), “What Can Machines Learn and What Does It Mean for [25]
Occupations and the Economy?”, AEA Papers and Proceedings, Vol. 108, pp. 43-47,
http://dx.doi.org/10.1257/pandp.20181019.
Brynjolfsson, E., D. Rock and C. Syverson (2017), “Artificial Intelligence and the Modern Productivity [6]
Paradox: A Clash of Expectations and Statistics”, NBER Working Paper Series, No. 24001, NBER,
http://www.nber.org/papers/w24001 (accessed on 16 January 2020).
Byrne, D. and D. Sichel (2017), “The productivity slowdown is even more puzzling than you think”, [39]
VoxEU, https://voxeu.org/article/productivity-slowdown-even-more-puzzling-you-think (accessed on
31 March 2020).
Cappelli, P. (2020), “The consequences of AI-based technologies for jobs”, R&I Paper Series, No. 4, [42]
European Commission, http://dx.doi.org/10.2777/348580.
Caselli, F. and A. Manning (2017), “Robot Arithmetic: Can New Technology Harm All Workers or the [50]
Average Worker?”, CEP Discussion Paper, No. 1497, Centre for Economic Performance.
CB Insights (2018), Top AI Trends To Watch In 2018, [102]

https://www.cbinsights.com/research/report/artificial-intelligence-trends-2018/.
Cockburn, I., R. Henderson and S. Stern (2018), The Impact of Artificial Intelligence on Innovation, [7]
NBER, http://dx.doi.org/10.3386/w24449.
Cowen, T. (2016), “Why Innovation Isn’t Over”, Foreign Affairs, [41]

https://www.foreignaffairs.com/reviews/review-essay/2016-02-15/innovation-over (accessed on
31 March 2020).
Dastin, J. (2018), “Amazon scraps secret AI recruiting tool that showed bias against women”, Reuters, [114]
Reuters, https://www.reuters.com/article/us-amazon-com-jobs-automation-insight/amazon-scraps-
secret-ai-recruiting-tool-that-showed-bias-against-women-idUSKCN1MK08G (accessed on
13 August 2020).
Daugherty, P. and J. Wilson (2018), “Collaborative Intelligence: Humans and AI Are Joining Forces”, [85]
Harvard Business Review, https://hbr.org/2018/07/collaborative-intelligence-humans-and-ai-are-
joining-forces.
For Official Use

Daugherty, P. and J. Wilson (2018), Human + Machine: Reimagining Work in the Age of AI, Harvard [84]
Business Review Press, Boston.
Davenport, T. and R. Kalakota (2019), “The potential for artificial intelligence in healthcare”, Future [33]
Healthcare Journal, Vol. 6/2, pp. 94-98, http://dx.doi.org/10.7861/futurehosp.6-2-94.
De Stefano, V. (2018), ““Negotiating the algorithm”: Automation, artificial intelligence and labour [106]
protection”, ILO Employment Working Paper Series, No. 246,
https://www.ilo.org/wcmsp5/groups/public/---ed_emp/---
emp_policy/documents/publication/wcms_634157.pdf (accessed on 4 February 2020).
Edelman (2019), 2019 Edelman AI Survey, https://www.edelman.com/sites/g/files/aatuss191/files/2019- [55]

03/2019_Edelman_AI_Survey_Whitepaper.pdf (accessed on 18 February 2020).
Ernst, E., R. Merola and D. Samaan (2018), The economics of artificial intelligence: Implications for the [58]
future of work, International Labour Organization, http://www.ilo.org/publns. (accessed on
25 May 2020).
EU-OSHA (2018), Key trends and drivers of change in information and communication technologies and [81]
work location, http://dx.doi.org/10.2802/515834.
European Commission (2020), On Artificial Intelligence - A European approach to excellence and trust, [112]
https://ec.europa.eu/info/sites/info/files/commission-white-paper-artificial-intelligence-feb2020_en.pdf
(accessed on 17 July 2020).
European Commission (2018), A new era for data protection in the EU - What changes after May 2018, [113]
European Commission, Brussels, https://ec.europa.eu/info/sites/info/files/data-protection-factsheet-
changes_en.pdf.
Felten, E., M. Raj and R. Seamans (2019), “The Occupational Impact of Artificial Intelligence on Labor: [24]
The Role of Complementary Skills and Technologies”, NYU Stern School of Business,
http://dx.doi.org/10.2139/ssrn.3368605.
Fossen, F. and A. Sorgner (2019), “New Digital Technologies and Heterogeneous Employment and Wage [32]
Dynamics in the United States: Evidence from Individual-Level Data”, IZA Discussion Paper Series,
No. 12242, https://www.iza.org/publications/dp/12242/new-digital-technologies-and-heterogeneous-
employment-and-wage-dynamics-in-the-united-states-evidence-from-individual-level-data (accessed
on 18 May 2020).
Frey, C. and M. Osborne (2018), Automation and the future of work, [27]
https://www.oxfordmartin.ox.ac.uk/blog/automation-and-the-future-of-work-understanding-the-
numbers/ (accessed on 9 April 2020).
Frey, C. and M. Osborne (2017), “The Future of Employment: How susceptible are jobs to [26]
computerisation”, Technological forecasting and social change, Vol. 114, pp. 254-280,
https://doi.org/10.1016/j.techfore.2016.08.019 (accessed on 8 April 2020).
Gal, P. et al. (2019), “Digitalisation and productivity: In search of the holy grail – Firm-level empirical [38]
evidence from EU countries”, OECD Economics Department Working Papers, OECD Publishing,
Paris, https://www.oecd-ilibrary.org/economics/digitalisation-and-productivity-in-search-of-the-holy-
grail-firm-level-empirical-evidence-from-eu-countries_5080f4b6-en (accessed on 19 August 2020).
For Official Use

Gartner (2017), Gartner Says By 2020, Artificial Intelligence Will Create More Jobs Than It Eliminates, [65]
https://www.gartner.com/en/newsroom/press-releases/2017-12-13-gartner-says-by-2020-artificial-
intelligence-will-create-more-jobs-than-it-eliminates (accessed on 6 January 2020).
Geroski and P. A. (2000), “Models of technology diffusion”, Research Policy, Vol. 29/4-5, pp. 603-625, [124]
https://doi.org/10.1016/S0048-7333(99)00092-X (accessed on 9 April 2020).
Goldberg, K. (2019), “Robots and the return to collaborative intelligence”, Nature Machine Intelligence, [87]
Vol. 1/1, pp. 2-4, http://dx.doi.org/10.1038/s42256-018-0008-x.
Gombolay, M. et al. (2015), “Decision-making authority, team efficiency and human worker satisfaction [93]
in mixed human–robot teams”, Autonomous Robots, Vol. 39/3, pp. 293-312,
http://dx.doi.org/10.1007/s10514-015-9457-9.
Good, I. (1966), “Speculations Concerning the First Ultraintelligent Machine”, Advances in Computers, [18]
Vol. 6/C, pp. 31-88, http://dx.doi.org/10.1016/S0065-2458(08)60418-0.
Gordon, R. (2018), “Why has Economic Growth Slowed When Innovation Appears to be Accelerating?”, [37]
NBER Working Paper Series, No. 24554, NBER, http://www.nber.org/papers/w24554 (accessed on
31 March 2020).
Grace, K. et al. (2017), “When Will AI Exceed Human Performance? Evidence from AI Experts”, Journal [22]
of Artificial Intelligence Research, Vol. 62, pp. 1-48, http://arxiv.org/abs/1705.08807 (accessed on
9 June 2020).
Grennan, J. and R. Michaely (2017), “Artificial Intelligence and the Future of Work: Evidence from [57]
Analysts”, NBER Economics of Artificial Intelligence Conference Paper,
https://conference.nber.org/conf_papers/f130049.pdf (accessed on 29 May 2020).
Gries, T. and W. Naudé (2018), “Artificial Intelligence, Jobs, Inequality and Productivity: Does Aggregate [28]
Demand Matter?”, IZA Discussion Paper Series, Vol. 12005, http://ftp.iza.org/dp12005.pdf (accessed
on 30 March 2020).
Griliches, Z. (1957), “Hybrid Corn: An Exploration in the Economics of Technological Change”, [8]
Econometrica, Vol. 25/4, p. 501, http://dx.doi.org/10.2307/1905380.
Guszcza, J. et al. (2018), “Why We Need to Audit Algorithms”, Harvard Business Review, [63]
https://hbr.org/2018/11/why-we-need-to-audit-algorithms.
Hancock, P. et al. (2011), “A meta-analysis of factors affecting trust in human-robot interaction”, Human [95]
Factors, Vol. 53/5, pp. 517-527, http://dx.doi.org/10.1177/0018720811417254.
HRWins (2018), Q4 HR Tech VC Update, https://larocqueinc.com/q4-hr-tech-vc-update/ (accessed on [101]

19 August 2020).
ILO (2016), Workplace Stress: A collective challenge, https://www.ilo.org/wcmsp5/groups/public/--- [105]

ed_protect/---protrav/---safework/documents/publication/wcms_466547.pdf (accessed on
13 August 2020).
For Official Use

Jaehrling, K. et al. (2018), “The digitisation of warehousing work. Innovations, employment and job [107]
quality in French, German and Dutch retail logistics companies”, in Jaehrling, K. (ed.), Virtuous circles
between innovations, job quality and employment in Europe? Case study evidence from the
manufacturing sector, private and public service sector,
https://www.soc.lu.se/en/sites/soc.lu.se.en/files/quinne_working_paper_no._6.pdf (accessed on
22 July 2020).
Jha, S. and E. Topol (2016), “Adapting to artificial intelligence: Radiologists and pathologists as [83]
information specialists”, JAMA, Vol. 316/22, pp. 2353-2354,
http://dx.doi.org/10.1001/jama.2016.17438.
Julia, L. (2019), L’intelligence artificielle n’existe pas, FIRST. [20]
Karasek, R. (1979), “Job Demands, Job Decision Latitude, and Mental Strain: Implications for Job [91]
Redesign”, Administrative Science Quarterly, Vol. 24/2, p. 285, http://dx.doi.org/10.2307/2392498.
Karasek, R. and T. Theorell (1990), Healthy work : stress, productivity and the reconstruction of working [92]
life, Ghent University Library.
Kark, R., D. Van Dijk and D. Vashdi (2018), “Motivated or Demotivated to Be Creative: The Role of Self- [97]
Regulatory Focus in Transformational and Transactional Leadership Processes”, Applied Psychology,
Vol. 67/1, pp. 186-224, https://doi.org/10.1111/apps.12122.
Knudsen, M. and J. Kaívo-Oja (2020), “Collaborative Robots: Frontiers of Current Literature”, Journal of [86]
Intelligent Systems: Theory and Applications, pp. 13-20, http://dx.doi.org/10.38016/jista.682479.
Korinek, A. and J. Stiglitz (2017), “Artificial Intelligence and Its Implications for Income Distribution and [77]
Unemployment”, NBER Working Paper Series, No. 24174, NBER,
https://www.nber.org/system/files/working_papers/w24174/w24174.pdf (accessed on
4 February 2020).
Landi, C. et al. (2018), “Relieving operators’ workload: Towards affective robotics in industrial [109]
scenarios”, Mechatronics, Vol. 54, pp. 144-154, http://dx.doi.org/10.1016/j.mechatronics.2018.07.012.
Lasota, P. and J. Shah (2015), “Analyzing the effects of human-aware motion planning on close-proximity [90]
human-robot collaboration”, Human Factors, Vol. 57/1, pp. 21-33,
http://dx.doi.org/10.1177/0018720814565188.
Leka, S. and A. Jain (2010), Health Impact of Psychsocial Hazards at Work: An Overview, World Health [104]
Organization,
https://apps.who.int/iris/bitstream/handle/10665/44428/9789241500272_eng.pdf;jsessionid=915757B9
AE66F6E057EC0D3409D6689B?sequence=1 (accessed on 13 August 2020).
Li, D., L. Raymond and P. Bergman (2020), Hiring as Exploration, NBER, [115]
https://www.nber.org/papers/w27736 (accessed on 27 October 2020).
Mathis, J. (2018), “Innovations are Coming to the Human Resource”, The Journal of Private Equity, [100]
Vol. 21/4, pp. 14-17, http://dx.doi.org/10.2307/26497439.
McElheran, K. (2018), “Economic Measurement of AI”, NBER Conference Paper, [53]

http://conference.nber.org/conf_papers/f114642.pdf (accessed on 18 May 2020).
For Official Use

McKinsey (2019), Survey: AI adoption proves its worth, but few scale impact, [12]
https://www.mckinsey.com/featured-insights/artificial-intelligence/global-ai-survey-ai-proves-its-
worth-but-few-scale-impact (accessed on 9 June 2020).
McKinsey (2018), “Notes from the AI Frontier: Modeling the Impact of AI on the World Economy”, [44]
Discussion Paper, https://www.mckinsey.com/featured-insights/artificial-intelligence/notes-from-the-
ai-frontier-modeling-the-impact-of-ai-on-the-world-economy (accessed on 14 January 2020).
MMC Ventures (2019), The State of AI 2019: Divergence, https://medium.com/mmc-writes/introducing- [68]

the-state-of-ai-2019-divergence-14d69cb3b16c (accessed on 3 January 2020).
Moore, P. (2019), “OSH and the Future of Work: benefits and risks of artificial intelligence tools in [75]
workplaces”, EU-OSHA Discussion papers, EU-OSHA, Bilbao,
https://osha.europa.eu/en/publications/osh-and-future-work-benefits-and-risks-artificial-intelligence-
tools-workplaces/view.
Moore, P., M. Upchurch and X. Whittaker (eds.) (2018), Digitalisation of Work and Resistance, Palgrave [111]
Macmillan.
Moretti, E. (2012), The new geography of jobs, Houghton Mifflin Harcourt. [67]
Moretti, E. (2010), “Local multipliers”, American Economic Review, Vol. 100/2, pp. 373-377, [66]
http://dx.doi.org/10.1257/aer.100.2.373.
Naudé, W. (2019), The Race against the Robots and the Fallacy of the Giant Cheesecake: Immediate and [47]
Imagined Impacts of Artificial Intelligence, https://www.iza.org/publications/dp/12218/the-race-
against-the-robots-and-the-fallacy-of-the-giant-cheesecake-immediate-and-imagined-impacts-of-
artificial-intelligence (accessed on 18 June 2020).
Nedelkoska, L. and G. Quintini (2018), “Automation, skills use and training”, OECD Social, Employment [15]
and Migration Working Papers, OECD Publishing, Paris, https://doi.org/10.1787/1815199X (accessed
on 13 February 2020).
Nielsen, P. et al. (2016), “Are Transformational and Transactional Types of Leadership Compatible? A [96]
Panel Study of Work Motivation”, Paper presented at 74th Annual Midwest Political Science
Association Conference, Chicago, United States.,
https://ps.au.dk/fileadmin/Statskundskab/Billeder/Forskning/Forskningsprojekter/LEAP/Dokumenter/
Nielsen_et_al..pdf.
Nordhaus, W. (2015), “Are We Approaching an Economic Singularity? Information Technology and the [19]
Future of Economic Growth”, Cowles Foundation Discussion Papers, No. 2021,
http://cowles.yale.edu/sites/default/files/files/pub/d20/d2021.pdf (accessed on 30 March 2020).
OECD (2021), Ethical issues arising from AI implementation at the workplace and associated policy [1]
challenges, OECD Publishing, Paris.
OECD (2019), Artificial Intelligence in Society, OECD Publishing, Paris, [3]

https://dx.doi.org/10.1787/eedfee77-en.
OECD (2019), OECD Employment Outlook 2019, OECD Publishing, Paris, [16]
https://doi.org/10.1787/19991266.
For Official Use

OECD (2019), OECD Future of Education and Skills 2030 Conceptual learning framework, OECD [73]
Publishing, Paris, https://www.oecd.org/education/2030-project/teaching-and-
learning/learning/skills/Skills_for_2030.pdf (accessed on 4 May 2020).
OECD (2019), Scoping the OECD AI Principles: Deliberations of the Expert Group on Artificial [2]
Intelligence at the OECD (AIGO), OECD Publishing, Paris, https://doi.org/10.1787/d62f618a-en
(accessed on 6 January 2020).
OECD (2018), Good Jobs for All in a Changing World of Work: The OECD Jobs Strategy, OECD [78]
Publishing, Paris, https://dx.doi.org/10.1787/9789264308817-en.
OECD (2018), Job Creation and Local Economic Development 2018, OECD Publishing, Paris, [125]
https://doi.org/10.1787/26174979 (accessed on 16 April 2020).
OECD (2018), OECD Science, Technology and Innovation Outlook 2018: Adapting to Technological and [64]
Societal Disruption, OECD Publishing, Paris, https://dx.doi.org/10.1787/sti_in_outlook-2018-en.
OECD (2017), OECD Guidelines on Measuring the Quality of the Working Environment, OECD [79]
Publishing, Paris, https://dx.doi.org/10.1787/9789264278240-en.
Olmstead, A. and P. Rhode (2002), “Reshaping the landscape: The impact and diffusion of the tractor in [121]
American agriculture, 1910-1960”, The Journal of Economic History, Vol. 61/03,
https://www.researchgate.net/publication/4734094_Reshaping_the_landscape_The_impact_and_diffusi
on_of_the_tractor_in_American_agriculture_1910-1960 (accessed on 16 October 2020).
Polanyi, M. (2009), The Tacit Dimension, University of Chicago Press, Chicago. [29]
PWC (2018), Artificial Intelligence in HR: a No-brainer, [99]

https://www.pwc.nl/nl/assets/documents/artificial-intelligence-in-hr-a-no-brainer.pdf (accessed on
28 July 2020).
PWC (2018), The macroeconomic impact of artificial intelligence, https://www.pwc.co.uk/economic- [43]

services/assets/macroeconomic-impact-of-ai-technical-report-feb-18.pdf (accessed on
5 February 2020).
Ransbotham, S. et al. (2019), “Winning With AI”, MITSloan Management Review, [71]
https://sloanreview.mit.edu/ai2019 (accessed on 11 June 2020).
RSA (2017), The Age of Automation Artificial intelligence, robotics and the future of low-skilled work, [118]
https://www.thersa.org/globalassets/pdfs/reports/rsa_the-age-of-automation-report.pdf (accessed on
18 March 2020).
Sage-Gavin, E., M. Vazirani and F. Hintermann (2019), “Getting Your Employees Ready for Work in the [72]
Age of AI”, MIT Sloan Management Review, https://sloanreview.mit.edu/article/getting-your-
employees-ready-for-work-in-the-age-of-ai/.
Saint-Martin, A., H. Inanc and C. Prinz (2018), “Job Quality, Health and Productivity: An evidence-based [82]
framework for analysis”, OECD Social, Employment and Migration Working Papers, No. 221, OECD
Publishing, Paris, https://doi.org/10.1787/a8c84d91-en.
Schwellnus, C. et al. (2018), “Labour share developments over the past two decades: The role of [40]
technological progress, globalisation andwinner-takes-most” dynamics”, OECD Economics
Department Working Papers, No. 1503, OECD Publishing, Paris, https://doi.org/10.1787/3eb9f9ed-en.
For Official Use

Seamans, R. and M. Raj (2018), “AI, Labor, Productivity and the Need for Firm-level Data”, NBER [52]
Working Paper Series, No. 24239, http://www.nber.org/papers/w24239 (accessed on 18 May 2020).
Sengenberger, W. (1987), Struktur und Funktionsweise von Arbeitsmärkten: Die Bundesrepublik [59]
Deutschland im internationalen Vergleich, https://www.isf-muenchen.de/pdf/isf-archiv/1987-
sengenberger-struktur-arbeitsmarkt.pdf (accessed on 19 August 2020).
TNO (Netherlands Organisation for Applied Scientific Research) (2018), Emergent risks to workplace [94]
safety; Working in the same space as a cobot, Report for the Ministry of Social Affairs and
Employment, The Hague, https://publications.tno.nl/publication/34627026/je8DYe/TNO-2018-
R10742.pdf.
Toney, A. and M. Flagg (2020), U.S. Demand for AI-Related Talent, Center for Security and Emerging [69]
Technology, https://cset.georgetown.edu/research/u-s-demand-for-ai-related-talent/ (accessed on
28 October 2020).
Ultimate Software (2018), Sentiment Analysis and The Future of Work, Ultimate Software Group Inc., [62]
https://www.ultimatesoftware.com/Contact/sentiment-analysis-and-future-of-workforce-whitepaper.
Van den Broek, E. (2017), “Monitoring technology in the workplace”, EU-OSHA Discussion papers, EU- [108]
OSHA, Bilbao.
Villani, V. et al. (2018), “Survey on human–robot collaboration in industrial settings: Safety, intuitive [88]
interfaces and applications”, Mechatronics, Vol. 55, pp. 248-266,
http://dx.doi.org/10.1016/j.mechatronics.2018.02.009.
Webb, M. (2020), “The Impact of Artificial Intelligence on the Labor Market”, [23]
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3482150 (accessed on 4 February 2020).
Wilson, H., P. Daugherty and N. Morini-Bianzino (2017), “The Jobs That Artificial Intelligence Will [61]
Create”, MITSloan Management Review, http://mitsmr.com/2odREFJ (accessed on 17 January 2020).
Woodcock, J. (2016), Working the Phones: Control and resistance in call centres, Pluto Press, [110]
https://www.plutobooks.com/9780745399065/working-the-phones/ (accessed on 30 July 2020).
Wright, R. (2019), “Workplace automation: How AI is coming for your job”, Financial Times, [74]
https://www.ft.com/content/c4bf787a-d4a0-11e9-a0bd-ab8ec6435630.
Yamamoto, I. (2019), “The impact of AI and information technologies on worker stress”, VoxEU, [80]
https://voxeu.org/article/impact-ai-and-information-technologies-worker-stress.
Zeng, M. (2018), Everything Alibaba Does Differently — and Better, https://hbr.org/2018/09/alibaba-and- [34]
the-future-of-business (accessed on 18 August 2020).
Zhang, H., S. Feinzig and H. Hemmingham (2018), Making Moves Internal Career Mobility and the Role [103]
of AI, IBM Smarter Workforce Institute, https://www.ibm.com/downloads/cas/X2O7V1OD.
For Official Use

Copyright of OECD Social, Employment & Migration Working Papers is the property of
Organisation for Economic Cooperation & Development and its content may not be copied or
emailed to multiple sites or posted to a listserv without the copyright holder's express written
permission. However, users may print, download, or email articles for individual use.

Teh Impact of AI On The Labour Market

Uploaded by

Copyright:

Available Formats

Teh Impact of AI On The Labour Market

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Teh Impact of AI On The Labour Market

Uploaded by

Copyright:

Available Formats

Organisation for Economic Co-operation and Development

Cancels & replaces the same document of 12 January 2021

The impact of Artificial Intelligence on the labour market: What do we know so

JEL Codes: J20, J81, J31, O14, O33.

Marguerita Lane, [email protected]

Anne Saint-Martin, [email protected]

OECD Social, Employment and Migration Working Papers

For Official Use

This publication contributes to the OECD’s Artificial Intelligence in Work, Innovation,

For Official Use

For Official Use

For Official Use

For Official Use

OECD Social, Employment and Migration Working Papers ............................................................................ 2

For Official Use

For Official Use

For Official Use

For Official Use

For Official Use

Certaines catégories de travailleurs seront peut-être mieux à même, en raison de leurs

For Official Use

For Official Use

For Official Use

For Official Use

For Official Use

1. Artificial intelligence (AI) is reshaping economies and societies, offering new

For Official Use

For Official Use

2. What can AI do?

2.1. AI can be considered a general purpose technology

For Official Use

Box 2.1. Recent trends in AI and its adoption in the workplace

2.2. AI can be considered an automation technology

For Official Use

case with a technology such as machine learning, which is specifically designed to

Box 2.2. AI and the singularity

For Official Use

2.3. Certain occupations are more exposed to AI

2.3.1. High-skilled occupations are among those most exposed to AI

For Official Use

Table 2.1. Occupations most and least exposed to AI

(Felten, Raj and Seamans, (Brynjolfsson, Mitchell and

For Official Use

2.3.2. Some bottlenecks to the development of AI remain

For Official Use

For Official Use

For Official Use

3. Impact of AI on productivity, employment and wages

34. Discussions about the predicted impact of AI on productivity, employment and

For Official Use

technologies rather than true technological breakthroughs resulting in large productivity

For Official Use

Box 3.1. Estimates of the economic potential of AI

Studies by consultancies suggest that AI has enormous potential to contribute to global

3.2. Theoretical models are ambiguous on the impact of AI on employment and

For Official Use

3.2.1. AI-facilitated automation is expected to reduce labour demand and

For Official Use

3.2.3. How does this framework apply to AI specifically?