B Explanatory memorandum
by Ms Marietta Karamanli, rapporteurNote
1 Introduction
1. Relations between the Council
of Europe and the Organisation for Economic Co-operation and Development
(OECD) were officially established in 1962. The first Parliamentary
Assembly debate on the activities of the OECD took place in 1963.
Enlarged Parliamentary Assembly debates were introduced in 1993, based
on special rules,
Note to allow delegations from the national
parliaments of OECD member States that are not members of the Council
of Europe and of the European Parliament to participate. Since then,
the enlarged Assembly has operated as a unique platform for parliamentary
scrutiny of the OECD activities.
2. In January 2019, a new methodology for enlarged debates was
agreed upon, on the basis of a Memorandum of Understanding between
the Parliamentary Assembly and the OECD,
Note with the aim of achieving
a stronger and more efficient institutional relationship between
the Assembly and the OECD. It provides for such debates to be held
every two years, based on a report focusing on specific themes,
chosen by common agreement between the rapporteur and the OECD.
Furthermore, it was agreed that during the year when there is no
enlarged Assembly debate on the OECD activities, an Assembly delegation
will participate in the OECD Global Parliamentary Network as an
institutional partner. In December 2020, the Council of Europe and
the OECD also signed a Memorandum of Understanding, aimed at deepening
their co-operation.
3. The last enlarged Assembly debate took place in January 2024
and focused on “Globalisation in times of crisis and war: the role
of the OECD since the Russian Federation's aggression against Ukraine”
(
Resolution 2526 (2024)). The Assembly also established ad hoc committees to
participate in the meetings of the OECD Global Parliamentary Network
in Paris, from 5 to 7 February 2025 and from 11 to 13 March 2026.
4. On 2 July 2025, following my appointment as rapporteur for
the new report on the activities of the OECD, I held a meeting in
Paris with Mr Mathias Cormann, Secretary-General of the OECD, and
other representatives of the OECD’s Secretariat: Ms Audrey Plonk
(Deputy Director for Science, Technology and Innovation); Mr Mark
Pearson (Deputy Director for Employment, Labour & Social Affairs);
Mr Stéphane Levesque (Director for Communications); and Mr James
Mancini (Acting Head of Policy, Secretary-General’s Office). During
this meeting, we agreed that the next report would focus on the
topic of digital transformation and the OECD’s role in evaluating
the impact of artificial intelligence (AI) on the future of work.
5. On 26 November 2025, I visited the OECD headquarters in Paris
again, and I held additional meetings with different Senior officers
of the OECD, to discuss in further details the issues to be covered
in the report. On 10 December 2025, the Committee on Political Affairs
and Democracy held a hearing with the participation of Ms Angelica
Salvi Del Pero, Senior advisor, Directorate for Employment, Labour
and Social Affairs, and Mr Flavio Calvino, Senior Economist, Firms,
Digital Transformation & Technology Diffusion, Directorate for Science,
Technology and Innovation. I am sincerely grateful to all the OECD
staff involved, for their valuable contributions at different stages
during the preparation of this report.
2 Relevant work of the Parliamentary
Assembly
6. In addition to the numerous
reports regularly debated by the enlarged Parliamentary Assembly
on the work of the OECD, the Parliamentary Assembly has recently
adopted various texts that are relevant for this report. It is worth
mentioning the following in particular:
7. In
Resolution 2526
(2024), the enlarged Assembly stated that the OECD and the
Council of Europe “should continue working together in the field
of artificial intelligence (AI)”.
8. Furthermore, the Assembly makes a significant contribution
to the Council of Europe’s efforts to address the impact of AI technologies
on human life. Notably, it has a dedicated Sub-Committee on Artificial
Intelligence and Human Rights, and in October 2020 it adopted a
series of resolutions and recommendations on the matter:
In a common Appendix to these reports, the Assembly set out
the ethical principles that it believes should be applied to AI
systems: transparency; justice and fairness; responsibility; safety
and security; privacy.
9. Through
Opinion 303 (2024), the Assembly welcomed the finalisation of the Framework
Convention on Artificial Intelligence and Human Rights, Democracy
and the Rule of Law (CETS No. 225). The Framework Convention is
the first-ever international legally binding treaty in this field.
Opened for signature on 5 September 2024, it aims to ensure that
activities within the lifecycle of AI systems are fully consistent
with human rights, democracy and the rule of law, while being conducive
to technological progress and innovation. The Framework Convention
complements existing international standards on human rights, democracy
and the rule of law, and aims to fill any legal gaps that may result
from rapid technological advances. In order to stand the test of
time, the Framework Convention does not regulate technology and
is essentially technology-neutral. It is based on the Council of
Europe’s standards on human rights, democracy and the rule of law,
which are also shared by the non-member States that participated
in the negotiations.
Note
10. Furthermore, the Assembly adopted
Resolution 2628 (2025) and
Recommendation
2300 (2025) “Artificial intelligence and migration”,
Resolution 2654 (2026) “Copyright enforcement in the artificial intelligence
environment”, and
Resolution
2662 (2026) “Protecting democracy from disruptions caused by artificial
intelligence”. The Assembly is also currently preparing other relevant
reports on: the use of AI by parliaments; safeguarding human rights
in the AI-driven public sector; AI and gender equality; and safeguarding
creativity and education in the age of generative AI.
11. Finally, in December 2025 the Assembly co-organised in London
a Parliamentary Conference on Artificial Intelligence with the Parliament
of the United Kingdom, during which participants exchanged best practices
and defined parliamentary roles in AI governance.
Note
3 An
overview of the OECD’s work on artificial intelligence
12. The rapid growth of AI and
related technologies is already having a significant impact on people’s
lives, particularly with regard to the ways in which they interact,
study, conduct research and work. According to the OECD’s Employment
Outlook 2023, this rapid progress suggests that OECD countries might
be on the brink of an “AI revolution”.
Note
13. AI has the potential to be a general-purpose technology that
could profoundly change a wide range of sectors and industries,
including finance, health, and security. Furthermore, the control
over data as well as over AI and the related infrastructure is having
increasingly significant implications at geopolitical level.
Note
14. Societies and economies must be prepared for an AI and digitalisation
transition in order to keep up with the pace of technological advancements,
harness their benefits and mitigate their risks. This must go hand
in hand with the green and sustainability transition, to ensure
that technological developments do not come at the expense of the
environment and of the climate, and that they take into account
the rights and needs of future generations. The OECD has highlighted
the positive contribution of digital technologies, including AI,
to advancing environmental goals in the “Recommendation of the Council
on Digital Technologies and the Environment” (adopted in 2010 and
amended in 2025).
Note
15. Recognising that “AI has pervasive, far-reaching and global
implications that are transforming societies, economic sectors and
the world of work, and are likely to increasingly do so in the future”,
in 2019 the Council of the OECD adopted a Recommendation containing
the following “Principles for responsible stewardship of trustworthy
AI” (OECD AI Principles, amended in 2024):
Note
- inclusive
growth, sustainable development and well-being;
- respect for the rule of law, human rights and democratic
values, including fairness and privacy;
- transparency and explainability;
- robustness, security and safety;
- accountability.
16. The OECD AI Principles also include the following recommendations
for governments to implement in their national policies and international
cooperation:
- investing in AI
research and development;
- fostering an inclusive AI-enabling ecosystem;
- shaping an enabling interoperable governance and policy
environment for AI;
- building human capacity and preparing for labour market
transformation;
- international co-operation for trustworthy AI.
17. Following the adoption of the OECD AI Principles, the OECD
has worked in recent years to develop a wide range of tools for
monitoring and reporting on the latest AI-related developments and
their economic and social impacts.
18. The OECD AI Policy Observatory is a large and up-to-date global
repository of national AI policies from over 90 jurisdictions and
international organisations. It also features live data on AI in
a range of areas, including trends in the demand for and supply
of AI talent across countries and sectors over time, as well as insights
into the most in-demand skills.
Note
19. The Catalogue of Tools and Metrics for Trustworthy AI is a
one-stop shop containing over 1000 tools and 130 metrics designed
to help AI actors develop and use trustworthy AI systems. These
include mechanisms, practices and methodologies for developing fair
and transparent AI systems, as well as for measuring and evaluating
AI trustworthiness and AI risks.
Note
20. The OECD AI Incidents and Hazards Monitor (AIM)
Note tracks and documents AI incidents
and hazards in real time, as reported by reputable media outlets.
Based on the OECD’s common reporting framework for AI incidents
Note and the OECD’s definition
of AI incidents,
Note the AIM helps build a common
understanding of AI incidents and hazards and highlights their multifaceted
nature, serving as an important tool for trustworthy AI.
21. The AI Policy Toolkit is a practical resource designed to
help governments and stakeholders map their AI policy landscape,
identify priorities, and explore concrete policy examples from other
countries to inform their own policy design.
Note
22. The OECD AI Capability Indicators provide a monitoring system
for governments to assess the evolving capabilities of AI systems
in relation to core human abilities. These indicators, currently
published in
beta form, provide
clear information to policymakers about current and likely future
AI capabilities on the labour market, informing discussions on what
skills will be needed in the future.
Note
23. Since the release of generative AI in 2023, the OECD has analysed
the different trajectories of countries, regions and cities in the
AI transition, highlighting emerging divides and the importance
of local conditions for seizing the full potential of AI. This work
has helped better understand the geography of generative AI and
job exposure across regions, develop blueprints for AI use in public
services in cities and regional development, and inform place-based
strategies for sustainable, ethical and trustworthy AI, including
through local innovation ecosystems, skills upgrading and capacity
building.
Note
24. In September 2025, the OECD released “Advancing the measurement
of investments in artificial intelligence”, a methodology for estimating
public and private AI investments in European Union (EU) member States,
and benchmark them against other economies, such as the United States,
the United Kingdom, Canada, and Japan.
Note
25. To support the European Commission in monitoring the EU Coordinated
Action Plan on AI, the OECD developed a report assessing EU
member
States’ progress towards its implementation,
Note as well as a report analysing
the uptake of AI in the EU in the agriculture, health, manufacturing
and mobility sectors.
Note
26. As part of the G7 Hiroshima AI Process, the G7 adopted in
2023 a Code of Conduct for organisations developing advanced AI
systems. To support its implementation, the OECD developed a voluntary
Reporting Framework for AI developers to report on their AI risk
management practices
Note and launched a report
Note analysing companies’ first
twenty submissions. A revised version of the reporting framework
was launched in May 2026.
27. The OECD Digital Well-being Hub provides key statistics spanning
digital aspects of work and job quality, health, education, environment,
personal safety, work-life balance, social connectedness, and civic engagement.
This includes selected indicators broken down by gender, age or
income decile. Through its embedded survey, the Hub is also gathering
unique data on how people experience their digital lives, filling
an important evidence gap.
Note
28. The OECD programme on AI in Work, Innovation, Productivity
and Skills, which was launched in 2020 with the support of Germany,
has produced extensive data and policy evidence on the impact of
AI on labour markets, innovation, productivity and skills needs.
Note
29. The 2025/2026 OECD Horizontal Project entitled “Thriving with
AI: Empowering Economies, Societies and Citizens” aims to improve
understanding of AI development and deployment trends, impacts and implications
for the economy and society. The project deliverables cut across
policy areas and sectors and feature analyses of AI’s potential
to drive productivity and output growth across firms, sectors, economies
and places. Deliverables are expected to be finalised by the end
of 2026, including two policy reports summarising key findings.
30. Findings from various OECD research projects and publications
demonstrate the importance of monitoring and analysing the impact
of increased AI usage in the workplace on areas such as employment rates,
wages, job quality, productivity and growth. As not all sectors,
places, occupations or demographic groups will be equally exposed
to or impacted by AI, tailored policies and multilevel governance
will be needed to ensure that no one is left behind in the transition,
and to promote all people’s ability to benefit from AI, while addressing
the risks to their fundamental rights and well-being.
4 The
impact of artificial intelligence on the economy: productivity and
the labour market
4.1 The
impact on productivity
31. Productivity growth at global
level has decreased in recent years. In OECD countries it has gone
from a yearly average of 1.5% in the period 2000-2010, to around
0.5% in 2019-2023. Economists are currently debating what impact
AI will have in terms of productivity, and opinions vary widely.
The most optimist scenarios suggest that AI might contribute up
to 2-3% to annual labour productivity growth over the next decade,
whereas the most pessimist ones indicate that the gains would be
negligible, given that only a limited fraction of the economy can
successfully automatise labour.
32. The OECD, in turn, estimates that AI could contribute between
0.4% and 0.9% annually to labour productivity growth in the US,
which could be compared to some of the estimates regarding the contribution
of Information and Communication Technologies (ICT) during the last
technology driven productivity boom in the mid-90s in the country
(around 1-1.5%).
Note A more recent update, in
light of rapidly rising AI adoption rates and expanding capabilities,
considers a more optimistic scenario as well, that leads to a contribution
of 1.2% to annual labour productivity growth in the US.
Note
33. The OECD treats AI in its studies as a production technology
which combines inputs (software, skills, data) and computing capacity
to produce a wide range of outputs (analytical tasks like prediction, recommendations
or optimisation; content generation; physical tasks in association
with robotics), thus increasing the productivity of economic activities.
34. In this framework, the OECD identifies four key factors determining
the impact of AI on productivity.
Note The first factor implies
that efficiency gains in specific tasks can be highly significant:
these include customer service, coding, professional and general
writing, and business consulting; a study of available literature
has shown that, on average, the improvement observed is around a
30% increase in performance.
35. The second factor indicates that highly qualified sectors
are the ones mostly exposed to AI, or in other words, the ones in
which AI can potentially have a stronger impact, because there is
a larger share of tasks that AI can assist with or substitute for:
finance, ICT services, media and professional services. On the other hand,
the least exposed sectors include agriculture, construction, and
sectors with a strong manual component. In many cases there will
be strong complementarity between human workers and AI, for instance where
AI tools require enhanced human supervision and involve liability.
36. The third factor is represented by the speed of business adoption
of AI: to assess different scenarios, the OECD has compared the
speed of adoption of AI at country level with the speed of adoption
of past general-purpose technologies (GPTs), such as electricity,
computers, and the internet. The US data show that the rate of AI
adoption is still in its early stages, but it is progressing in
line with past digital technologies. At regional level, however,
the speed of AI adoption varies markedly between capital and non-capital
regions, between innovation leaders and laggards, or within a same
country and a same sector, between AI hotspots and non-specialised
regions.
Note Even regions at similar levels
of potential job exposure show different speeds of actual business
adoption.
Note
37. The fourth factor that can determine the impact of AI on productivity
results from local conditions and place-based strategies. Beyond
economic specialisation and skills structure, adoption is tied to
local access to AI assets, such as data or skills, or AI compute
and non-AI infrastructure that is essential for operating AI systems,
such as power grid or high-speed broadband networks. Integrating
AI into legacy systems also requires technology convergence across
workplaces and production lines, as well as financial and institutional capacities
for deployment. Local innovation systems play a key role in tailoring
AI solutions and boosting adoption.
Note
38. AI is being more widely used across large firms; the share
of AI users tends to be also higher among “young firms”, including
start-ups.
Note Firms using AI tend to be
more productive than other firms: this seems to be related to the
fact that they have higher complementary human and technological
capital (digital infrastructure, digital and innovative capabilities
of firms, and other digital technologies) that might be a prerequisite
for the use of AI. Furthermore, sectors with high AI intensity are
also the ones shaping the technology, i.e. IT services, telecommunications
and media.
Note
39. The OECD analysis also shows that, in the next 10 years, there
will be significant differences in terms of macroeconomic gains
related to AI among OECD countries, caused by different economic
structures, variable skill levels available (companies adopting
AI require much higher levels of technical and social skills), differing
adoption rates, and the way place-based conditions and strategies
unfold. This might also increase global inequalities, with poorer
countries and places being more penalised, given their slower adoption
rates.
Note
40. Furthermore, the technological development measured through
the number of AI-related patents shows a strong geographical concentration
in the US and China, which have a dominant position, whereas the
EU lags behind, with less than half of the total AI patents. The
US and Chinese AI models currently represent, in fact, the technological
frontier.
41. Europe invests far less than the United States and China in
infrastructure, research and computing capacity, which fuels a structural
dependence on clouds and the models offered by a handful of dominant private
players. During the first three months of 2026, Amazon, Google,
Microsoft and Meta repeatedly broke records for spending on AI;
they invested a total of $130.65 billion in capital expenditure,
mainly in data centres that power AI. This figure – which sets a
new record – is more than three times the cost of the Manhattan Project
to develop nuclear bombs and is 71% higher than what the tech giants
had spent during the same quarter the previous year.
Note Public investment
must therefore not be viewed as a scattered catch-up effort, but as
a strategic lever: sustainable funding for research and skills,
support for shared and sovereign infrastructure, and prioritising
well-considered technologies rather than simply subsidising additional
energy-intensive capacity. This involves making an explicit trade-off
at European level between technological sovereignty, environmental
cost and fiscal sustainability, rather than leaving this choice
solely to large private consortia.
42. At the same time, Europe’s dependence on semiconductors, on
infrastructure dominated by GAFAM (the tech giants: Google, Apple,
Facebook-Meta, Amazon and Microsoft) and on AI models developed
outside Europe undermines technological sovereignty, economic security
and the capacity for democratic control. The response cannot be
solely national: it requires a foundation of infrastructure and
models under European jurisdiction, compatible with the Council
of Europe’s Framework Convention and OECD principles, as well as mechanisms
for pooling investment at EU level. The challenge is to prevent
public administrations and services from becoming structurally dependent
on non-European solutions, while recognising that greater digital strategic
autonomy will require substantial investment but will also bring
significant economic, security and democratic benefits.
43. In order to benefit from the possible significant productivity
gains from AI, countries will need to:
- focus on the acceleration of AI adoption, by strengthening
the relevant infrastructure, data and skills, and the capacity of
local innovation systems to support sectoral transformations;
- promote an efficient reallocation of labour (through reskilling
and flexible labour markets);
- ensure global access to AI, e.g. by ensuring the interoperability
of open standards;
- preserve trade to disseminate technologies.
44. Generative AI (GenAI) tools are capable of creating new outputs
(e.g. text, code, audio, images, video), often in response to prompts,
based on their training data. GenAI has the potential of becoming
a GPT, as it appears to show defining characteristics of GPTs: pervasiveness;
continuous improvement over time; and innovation spawning.
Note The extent to which AI could
boost innovation and the generation of new ideas, and therefore
productivity, is important to understand the long-term transformative
implications of AI and its future returns in terms of economic growth.
45. The role of policy-makers in this sense is crucial, to achieve
an inclusive digital transition that can lead to significant social
and economic benefits, while respecting human rights and democratic
values. In order to do so, particular care should be given to cultivating
inclusive AI innovation ecosystems; supporting firms’ digitalisation;
developing human capital for AI adoption and use (technical and
non-technical AI skills), to achieve an equitable labour market
transformation; and co-operating at international level for a trustworthy
AI governance, in particular to ensure the interoperability of policy
and regulation frameworks.
Note
4.2 The
impact on the labour market
46. OECD countries are experiencing
a rapidly changing labour market:
Note overall, employment rates
have been on the rise since 2014, but there are fewer low-skilled
jobs. On average, the occupations at the highest risk of automation
account for 27% of employment, with some regions displaying shares
well above this average: this might lead in the future to a mass
need for redeployment of workers to other jobs in potentially different
regions within their country, with increased demand for training
and support services for mobility. The risk of automation is not
equally distributed across groups in the population: a concerning
decrease of 30% in employment rates for entry-level jobs among young
people has been observed in 2025.
NoteHowever,
the worsening labour market situation of young people started well
before the COVID pandemic, and so AI cannot be the only reason.
47. That being said, workers are worried about their jobs. Among
workers in manufacturing and finance sectors, as shown by a survey
conducted by the OECD, three in five are worried about losing their
jobs entirely to AI in the next ten years, particularly those who
actually work with AI. On the other hand, most employers in the
same sectors stated that the adoption of AI has not had an impact
on workforce numbers so far.
Note
48. It should be noted that the major trade unions rightly point
out that it is not abstract technological forces that determine
whether jobs are lost or created, but the choices made by companies
regarding how they use AI. In this regard, trade unions emphasise
that social dialogue on the use of AI remains limited; they call
for the resulting decisions affecting employees to be the subject
of social dialogue, for AI to form part of a collective company-wide
strategy, and for the necessary training to be provided by the employer.
49. A survey conducted among SMEs in 2024 indicated that the adoption
of generative AI did not have an impact on staffing needs in most
cases (83%).
Note
50. According to the OECD, high-skilled workers are currently
most exposed to AI – on the other hand, employment growth has been
greatest in the most exposed occupations: AI could at the same time
increase productivity and, under certain conditions, increase the
demand for the related services and therefore for the kind of workers
with the skills to deliver them.
Note
51. AI is reshaping skill needs in the labour market. One of the
decisive factors determining whether economies benefit from AI is
likely to be the skills of the people using it. Skills gaps are
already holding back AI adoption, while AI is simultaneously raising
the skills bar, shifting demand toward higher-level skills. Governments
must scale up access to AI-relevant training, particularly for SMEs
and workers at risk of being left behind.
Note
52. In fact, a large majority of workers in manufacturing and
financial services have confirmed that AI has improved their performance
at work. In turn, the most reported benefit from surveyed SMEs is
indeed employee performance. While different levels of skills among
workers can become a barrier and a discriminant in the adoption
of AI, a number of SMEs also reported that using generative AI is
actually helping compensate the lack of skills or experience among
their staff.
53. More importantly, workers are reporting improvements in their
performance, enjoyment and mental health at work: AI can reduce
physical strain, automate routine and tedious tasks, and free up
time for more interesting ones. At the same time, workers are worried
about privacy and excessive data collection by AI (e.g. increased
pressure to perform; worry that collection of data will lead to
decisions biased against them).
Note
54. An additional in-depth study of the OECD has surveyed over
6000 mid-level managers in 6 countries to explore the prevalence
of algorithmic management, i.e. the use of software, which might
include AI, to fully or partially automate tasks traditionally carried
out by human managers.
Note
55. The findings show that algorithmic management tools are already
widespread in the US (90% of managers responded that their firm
provides at least one algorithmic management tool) and in the European countries
surveyed (79%), but less prevalent in Japan (40%). The intensity
(i.e. number of tools used) also varies: in the US, more than three-quarters
of firms use ten or more tools, whereas the intensity is more moderate
in Europe and low in Japan.
56. Furthermore, the types of tools used are different: in the
US, monitoring tools and evaluation tools are very largely used
(adoption rate of 90%); 55% of firms also monitor the content and
tone of employees’ voice calls or emails, against only 6% in Europe
and 8% in Japan. European firms most often use instruction tools (intended
to give instructions to employees, with an adoption rate of 69%)
and basic monitoring tools (e.g. to track working time, with an
adoption rate of 33%). Japanese firms mostly use monitoring tools.
57. A large share of managers reports a positive impact of algorithmic
management tools in their decision-making (60%), and many of them
reported changes in their job satisfaction (the largest share in
the US), mostly linked to reductions in stress and in repetitive
work. On the other hand, two in three managers reported having at
least one concern regarding the trustworthiness of algorithmic management
tools used, in particular for what concerns unclear accountability
in case of a wrong decision; the inability to follow the logic of
algorithmic decisions or recommendations; and inadequate protection
of workers’ physical and mental health.
58. The most important area of concern regards the impact on the
quality of work: a trustworthy use of AI in the workplace means
recognising and addressing risks regarding privacy but also transparency
and explainability; robustness, safety and security; and accountability.
59. Overall, it is perhaps too early to make clear estimations
on the impact of AI on the labour market, which is likely to be
mixed: a generalised adoption of AI will probably lead to both job
creation and job losses. Some occupations might see a reduction
in the demand, whereas most exposed workers could benefit from it
on average. There is a significant risk of widening inequalities,
as certain groups (such as older workers or low-skilled workers)
might be left behind in the transition, but at the same time AI
could create new opportunities for certain categories, e.g. disabled
people.
5 Education
and skills: how to prepare citizens for the AI transition
60. The rapid changes in the labour
market will likely have a significant impact in the kind of skills
demanded for certain jobs. This involves ensuring that the adoption
of AI and the resulting change in demand for skills is accompanied
by strategies to update curricula and qualifications, building foundational
AI and digital literacy, critical thinking and socio-emotional skills,
while also preparing high-skilled pathways, investing in teacher training
and in lifelong learning. Importantly, AI literacy of 15-year-olds
at the end of upper secondary education will be assessed internationally
for the first time in the OECD's Programme for International Student Assessment
(PISA) in 2029, marking a significant step in enabling countries
to benchmark their students’ preparedness against peers globally.
5.1 The
impact on learning and teaching
61. The OECD is implementing with
the European Commission the joint initiative “Empowering Learners
for the Age of AI: An AI Literacy Framework for Primary and Secondary
Education (AILit Framework),
Note with the support of Code.org. The
framework is designed for teachers, education leaders, education
policymakers, and learning designers. Its purpose is to equip students
with the knowledge, skills, and attitudes to understand and use
AI safely and effectively.
62. The framework defines AI Literacy as “the technical knowledge,
durable skills, and future-ready attitudes required to thrive in
a world influenced by AI. It enables learners to engage, create
with, manage, and design AI, while critically evaluating its benefits,
risks, and ethical implications.”
63. The Framework evolves around four domains of AI Literacy:
Engaging with AI, which implies using AI as a tool to access new
content, information, or recommendations; Creating with AI, which
consists of collaborating with an AI system in a creative or problem-solving
process; Managing AI, which requires intentionally choosing how
AI can support and enhance human work; and Designing AI, which empowers learners
to understand how AI works and connect it to its social and ethical
impacts by shaping how AI systems function.
64. The Framework is an interdisciplinary tool which covers technical
skills and attitudes, provides best practices and is designed to
stand the test of time, despite future technological developments.
It also develops most of the skills that will be evaluated in the
PISA 2029 exercise.
65. Furthermore, the OECD has been producing dedicated Digital
Education Outlooks since 2021,
Note in order to provide education policy
makers and researchers with insights concerning the latest trends
and policies internationally related to the increased use of digital
technology and data in education.
66. The Digital Education Outlook 2026 focuses on “Exploring Effective
Uses of Generative AI in Education”.
Note GenAI tools are widely accessible
and used outside institutional control by students, teachers and
researchers. They can be general-purpose tools (designed to be performance-enhancers)
or educational tools (designed to be learning-enhancers).
67. As shown in the Digital Education Outlook 2026, using GenAI
tools to augment performance does not automatically enhance students
learning. As an example, a study of 1000 high school students in
Türkiye has shown that while having access to GPT-4 tools improved
students’ performance during practice, once that access was removed,
students who had used the tool performed 17% worse than those who
never had access.
68. There is therefore a crucial distinction to be made between
immediate performance and actual learning. The over-reliance on
GenAI tools can lead to negative cognitive effects, such as “metacognitive
laziness” (avoiding the diagnosis of a problem as well as the evaluation
and iteration of possible solutions), with the risk of longer-term
cognitive consequences.
69. On the other hand, there are possible positive impacts when
used with pedagogical purpose: a study in Indonesia focusing on
the use of ChatGPT as a pedagogical support tool showed that it
improved critical thinking awareness and collaborative tendencies
among students. Another study conducted in the United Kingdom showed
that students using GenAI to draft stories produced more creative
outputs, but less diverse ideas.
70. An interesting possible development mentioned in the Digital
Education Outlook 2026 is the evolution of GenAI tools into AI tutors,
following the longstanding tradition of intelligent tutoring systems:
AI tutors powered by generative AI offer new educational possibilities,
not so much by providing answers but rather by engaging users in
deep dialogue and providing adaptive guidance. A key for the success
of these systems will be the integration of the “pedagogy-first”
principle, the prioritisation of inclusivity and ethics, and the
adoption of safeguard mechanisms.
71. For what concerns teaching, GenAI tools can contribute to
boost productivity and the quality of teaching, in particular of
online tutoring; the risk is, however, that over-reliance on them
could lead to loss of skills and teaching expertise.
72. A case study presented in the Digital Education Outlook 2026
is the one of JeepyTA, from the United States: an AI teaching assistant,
which was developed to provide students with answers related to
logistic questions; feedback on their essays; clarifications on
course readings and lectures; and other additional support. The
tool was rated by students as comparable to human teacher assistants
in clarity, accuracy and professionalism, and it raised the proportion
of students achieving top grades on essays from 64% to 95% – even
though students still preferred to interact with human teaching
assistants.
73. Human-AI interactions can happen in three ways: through replacement,
when the AI tool accomplishes a task; through complementarity, when
the AI tool pairs human judgment; and through augmentation, when
the human-AI system accomplishes a task outperforming what the AI
or the human could have produced on their own. GenAI tools provide
promising possibilities to augment teachers’ professional abilities.
74. It is important to ensure a pedagogical validation of the
tools: generic AI tools have proven effective at boosting short-term
performance, but less so at supporting long-term learning. As suggested
in the Digital Education Outlook 2026, a possible way to tailor
GenAI tools to educational needs is to co-design them with teachers
and students, allowing them to control how they behave and interact
with students.
75. GenAI can also be applied to improve system and institutional
management, by reshaping administrative tasks (making them faster
and more accurate); standardising assessment (generating exam items
at scale); and expanding research possibilities.
76. Ultimately, the challenge for policymakers will be to ensure
that GenAI tools are developed to be learning tools and not shortcuts.
The human relationship between teacher and students should not be
lost, also because machines cannot replace human critical thinking
nor some motivational aspects of the human relationship that are
key to learning.
5.2 The
impact on skills
77. AI adoption by firms is changing
how workers perform their jobs, how work is organised, and ultimately, what
skills are demanded. Furthermore, against the backdrop of an ageing
population, AI can help address labour and skills shortages, as
already reported by SMEs in the above-mentioned study conducted
by the OECD.
Note On the other hand, 50% of
SMEs also reported that lack of skills among employees was one of
the major barriers to adoption of GenAI. A recent study explores
the relation between AI and ageing, confirming that AI has the potential
to ease labour shortages and support economic growth. However, unlocking
these benefits will require investing in reskilling and upskilling,
as well as supporting labour market reallocation and fostering business
dynamism and innovation.
Note
78. Employers in the finance and manufacturing sectors involved
in the study expressed that AI would increase demand for highly
skilled workers; skills that are perceived as more relevant include
data analysis and interpretation, as well as creativity and innovation.
79. The requested skills are indeed evolving: while there will
be no need to be specialised AI experts to use the technology, an
analysis of job advertisements conducted in 2023 has shown that
leading employers seeking technical AI skills are also looking for
broader, high-level soft skills such as problem-solving, leadership,
and innovation.
Note
80. The OECD also conducted a literature review covering more
than 80 experimental studies which investigate the effects of GenAI
on productivity, innovation and entrepreneurship. The results show
that GenAI produces relevant gains across the field reviewed, but
its impact appears to be highly dependent on how it is utilised,
for which purpose, by whom, and on their level of expertise and
trust in GenAI. Individuals with less experience and skills tend
to benefit more from GenAI when tasks are well-defined; in order
to unlock significant gains for experienced workers, the technology
needs to complement their expertise. Understanding how and for which
purpose GenAI is used, critically assessing outputs and considering
the fit between the tasks undertaken and GenAI’s capabilities is,
however, critical for benefits to materialise.
81. The technology can support personalised learning experiences,
but overreliance on AI-generated content may reduce critical thinking:
careful integration of GenAI, including through training and support, remains
crucial.
Note In fact, worker’s training
and consultation regarding the adoption of new technologies seem to
make them more positive about the impact of AI on their work, and
more prepared. Training and consultation become therefore important
not only for upskilling, reskilling and adaptation, but also to
improve their enjoyment in the workplace.
6 Ensuring
inclusiveness and equal opportunities
82. One of the major risks related
to the adoption of AI tools in the workforce is the possibility
that it leads to greater inequalities among workers. Countries will
need to develop policies that support different groups and capture
the benefits of AI (in terms of increased productivity and economic
growth), without exacerbating inequalities and societal resistance
to technological progress. It will be particularly important to
focus on policies that enable both women and men to benefit from
the AI transition equally.
83. Data shows that women are underrepresented in occupations
with the very highest exposure to AI. Furthermore, women are highly
represented in clerical occupations, which could be at particular
risk of automation given the developments related to AI technology.
Note
84. Women participate less in science, technology, engineering
and mathematics (STEM) fields, limiting their opportunities to access
newer, specialised AI positions (which are also better paid).
85. Overall, it appears that high-skilled male workers are overrepresented
in the AI workforce (those who develop AI technology) as well as
among AI users: women and lower-educated workers would therefore
have less access to both specific AI jobs, and to AI tools in the
workplace.
Note
86. Inequalities can also appear among different age groups. Young
people are more at risk, as entry-level jobs are more easily automated.
On the other hand, even though older people are generally less exposed
to automation, as they tend to have skills complementing AI,
Note some are less likely to easily
adapt to use AI tools. Policy-makers will need to carefully assess
these issues in order to counter the risk of age discrimination.
87. On the other hand, a very significant contribution of AI technologies
to increase inclusiveness is represented by their potential for
significantly improving the access to labour market and the quality
of the work environment for people with disabilities (for example,
through speech or image recognition, conversational chatbots, remotely
operated machinery).
Note
88. It is also important to mention that labour markets in different
countries are and will be affected by AI in different ways. Southern
and Eastern Europe seem to be the most vulnerable regions on the
European continent, based on their employment structure and on the
higher amount of low-skilled workers exposed to AI. This will be
an important indicator to be considered in order to define where
training investments can potentially have the highest impact.
7 Implications
for intellectual property
89. The rise of GenAI tools, and
the increasing demand of data to train them (text, audios, images
and videos) is also accompanied by concerns regarding intellectual
property (IP), especially regarding certain data collection methods.
90. The OECD has produced a publication on one of these methods:
“data scraping”, which refers to the automated extraction of public
or non-public data and information from third-party sources (websites, databases,
social media). This practice is a common method to train AI models,
occurring on a very large scale. It relies on automated means, and
it can severely affect creators and owners of IP-protected works,
as often there is no coordination with them, meaning that they are
not informed, and most importantly, not paid.
Note
91. Data scraping is used for commercial purposes, but also to
support academic research. The IP implications are different, in
terms of copyright, database rights, trademarks, trade secrets,
publicity, and moral rights. Possible public policy tools should
be tailored to the different uses, and the OECD has developed some suggestions
in this regard:
- awareness raising
on data scraping and the related legal implications would empower
stakeholders and provide them with information on how to protect
their rights;
- a “data scraping code of conduct” would encourage responsible
practices and provide specific guidelines for different actors,
including mechanisms for monitoring adherence and recommendations for
transparency practices;
- standard technical tools would protect IP rights and enable
rights holders to manage access to their data;
- standard contract terms would address legal and operational
issues and allow the harmonisation of contractual terms.
92. IP legislation will probably need to be revised, updated and
possibly harmonised among countries, to reflect the ongoing technological
developments. The impact of GenAI technologies on IP also has additional implications,
as restricting access to data could limit the development of AI
technologies and hamper innovation.
8 Adoption
by small and medium-sized enterprises and creative industries
93. Across the OECD, small and
medium-sized enterprises (SMEs) represent around 99% of all firms:
they are therefore a crucial source of employment, and on average
they generate between 50% to 60% of value added.
Note Beyond the mass effect, SMEs are
key actors for the diffusion of sustainable, ethical and trustworthy AI,
through the role they play in tailoring innovation to markets, including
in niches or remote areas, and their integration in supply chains.
Note
94. The adoption of AI by SMEs has risen sharply since 2023, with
its use by small businesses (from 7.1% to 17.4%) and medium-sized
enterprises (from 13.6% to 29.6%) more than doubling between 2023
and 2025.
Note However, this increase was from
a low base, and large firms extended their lead in AI adoption over
the same period (the use of AI by larger firms grew from 30.4% to
52.1%). There are also wide gaps between sectors, and in the uptake
of more mature digital technologies. As these technologies continue
to evolve, it is essential to put in place policies that help small
and medium-sized businesses keep pace and access complementary assets
to successfully adopt AI, including broadband connectivity, compute,
software, data and skills.
Note
95. A recent OECD survey of 2,000 SMEs across 10 OECD countries
showed that among AI users more than half of surveyed SMEs (54%)
report deriving at least moderate value from their AI use, with
33% reporting a moderate impact on their business, 15% a significant
impact, and 6% reporting a transformational impact. At the same
time, 15% cite a significant positive impact on efficiency, productivity
or decision making. Although this sample was not representative,
it does provide useful indicative evidence that SMEs using AI are
already experiencing tangible business benefits, particularly in
terms of efficiency, productivity and decision making.
Note
96. Recent technological developments are likely to have particularly
significant implications for SMEs in the cultural and creative sectors
(CCS). The sector is dominated by very small firms: across OECD
countries, 99.9% are micro enterprises or SMEs, with 96.1% employing
fewer than 10 people.
97. GenAI tools are having a particularly high impact on the sectors
of advertising, film and TV production, music, publishing, video-games
and visual arts. The main applications include the generation of
scripts, still and moving images, music, news stories, computer
codes, as well as video and image editing, and targeted advertising.
These applications can have implications in terms of copyright violation,
and can lead to the substitution of creators. Furthermore, the rise
in GenAI use can lead to a loss of cultural diversity due to the standardisation
of products, as well as to the dissemination of deep-fakes and disinformation.
98. Upcoming OECD work on the impact of AI on labour markets in
CCS shows that cultural and creative occupations are more exposed
to AI than the average occupation (more than 80% of tasks can be
performed 50% faster with AI, compared to 44% of tasks on average
across all occupations). It also shows that demand for AI skills
is roughly twice as high in cultural and creative occupation job
postings compared to average (4% against 2%) and that uptake of
AI is generally higher in CCS businesses compared to average (56%
vs 20% average), though this varies by subsector.
99. Evidence points towards potential disruption in CCS as a result
of generative AI specifically. For example, in 2025, around 9.5%
of businesses in the EU used AI to generate pictures, videos or
sound/audio, and 34.9% of SMEs in 2024 had used AI to generate images,
16.9% to generate video, and 14.7% to generate audio. The most common
use of GenAI was in marketing and sales (50% of SMEs using GenAI),
which may suggest a reduction in need for freelance designers and
marketing consultants.
100. The findings suggest that a proactive policy approach will
be needed to help CCS adapt to the growing use of GenAI. Given the
sector's wider cultural and social value, high levels of self-employment
and project-based work, and the continued importance of human creativity,
policy should support workers and businesses in adapting to changing
skill demands while mitigating risks to employment quality, creative
careers and cultural diversity. Priorities include investing in
hybrid skills that combine creative, digital and AI capabilities,
supporting AI adoption among SMEs and freelancers, ensuring fair
governance of rights and remuneration, strengthening cultural ecosystems,
and improving the monitoring of AI's labour market impacts through
more timely and granular data.
9 The
use of artificial intelligence by the public sector
101. In order to meet increasing
demands from citizens and strengthen trust in public institutions,
national and local governments will need to adopt AI ensuring that
it is well-governed, targeted, and human‑centred. Ignoring the AI
transformation or waiting for all uncertainties to be resolved risks
turning governments into technology‑takers rather than shapers of
public values, with significant long‑term costs and disadvantages. Without
sustained investment in internal AI capacity and governance, national
and local governments risk lock‑in, dependency, and loss of strategic
autonomy.
102. The full potential of AI tools in different public services
(such as employment, social and health services) remains untapped
and will require careful design, implementation and monitoring.
AI has multiple potential uses by the public sector, with benefits
ranging from improved efficiency of public administration, increased tailoring
and effectiveness of public policies, strengthened integrity and
transparency in public action, and greater responsiveness in public
services. Opportunities are to be found across all levels of government,
and the public services delivered at each level, including in and
within regions and cities.
103. The OECD has recently conducted research
Note on the use of AI in 11 core
government functions, across 200 use cases. The research provides
a rich collection of examples and good practices from different
OECD countries. The results indicate that AI is most prevalent in
core functions of public service, justice and civic participation,
and less common in policy evaluation, tax administration and civil
service reform. Furthermore, AI use is more prevalent for improving
internal administration or public facing service delivery, but less
in government oversight and policymaking. Similar results emerge
from recent analysis of use cases for regional development by national
and subnational governments.
Note
104. In June 2026, the OECD published its first Digital Government
Outlook,
Note a comparative analysis of digital
government across OECD member and accession candidate countries.
It draws on two OECD benchmarking indexes, the 2025 Digital Government
Index and the Open, Useful and Re-usable Data Index, which provide
comparative metrics of countries' digital maturity by assessing
the strength of the capabilities that underpin digital government,
from data governance and infrastructure to skills and the use of
AI, rather than whether policies or online services simply exist.
In a dedicated chapter on AI, the Outlook finds that the technology
is already used in at least one area of government in almost every
OECD country, and that most countries have established dedicated
strategies, oversight bodies and training programmes, marking a
shift from early experimentation towards integration. It demonstrates
that governments have now built many of the foundations for digital
and AI-enabled government, such as shared infrastructure, interoperable
systems and open data, but that realising impact at scale depends
on closing gaps in institutional capabilities, governance mechanisms
and workforce skills, and on creating the conditions for AI to deliver
reliable results in day-to-day operations.
105. While government use of AI is increasing, closing these gaps
is also critical to public trust in how it is used. The latest OECD
Survey on Drivers of Trust in Public Institutions, also published
in June 2026, includes for the first time a dedicated chapter on
trustworthy AI in the public sector, and finds that most people
remain sceptical, with only around four in ten holding positive
views about the potential of AI use by government institutions.
Note Familiarity with AI is strongly
associated with more positive expectations, and people who trust government
to handle their personal data responsibly are also more optimistic
about its use in the public sector. The report highlights appropriate
regulations, transparency, meaningful human oversight and strong
data protection as important levers for building public confidence.
106. Other OECD studies have focused on specific AI use by public
entities. One of these focuses on the use of AI to strengthen public
employment services: already in 2023, almost half of public employment
services in OECD countries were utilising AI, most commonly to match
job-seekers with vacancies and to aid the design of vacancy postings
(including classifying occupations).
Note In particular, public employment
services in OECD countries are using AI across all key areas of
their activity, to understand job-seeker needs and targeting support;
for labour market matching and employer services; and for administrative
activities and knowledge generation.
Note
107. AI has also multiple current and potential uses in social
protection, to improve policy design; support claims administration;
enhance service delivery; reduce “non-take-up”; and enhance monitoring
and evaluation. Opinions among citizens are however divided on governments
using AI in social services and to process benefits.
Note
108. Furthermore, governments at all levels are starting to use
AI to improve and enhance civic participation in policy making,
for instance to better analyse citizens’ inputs received through
public consultations; lower the barriers to participation opportunities
through automated translation, transcription, virtual assistance,
and targeted communication; and facilitate deliberative processes
at scale.
Note
109. However, using AI in government can presents significant risks,
often similar to those associated with diffusion in industry, but
also with some unique aspects for the public sector. Because AI
is deployed within broader public sector systems and processes,
incidents or harms may not remain isolated, but can propagate across
service delivery, decision-making and institutional relationships.
If significant incidents or harms emerge, they can have far-reaching
impacts on government operations and broader societal outcomes.
Note Known risks consist of:
- ethical risks, when AI outputs
undermine human rights, resulting from biased algorithms, poor-quality data
or misuse, such as invasive surveillance, abusive inference or privacy
infringements;
- operational risk, including technical and operational
failures, mistakes in training, too weak oversight or deficient
infrastructure;
- exclusion risks, related to gaps among citizens in terms
of access to the technology or digital literacy, or inadequate representation
in training data;
- public resistance risks, related to citizens’ distrust
in government AI systems or processes, (as explained above) and
lack of transparency or avenues for redress;
- inaction risks, related to government inaction or delays
in adopting AI technology to yield positive benefits.
110. Governments at national and sub-national levels, face several
challenges in AI implementation. These include the difficulties
in procuring AI technologies; moving from pilot phases to scaling
up successful AI applications; filling skills gaps in the public
sector, including non-expert AI skills; obtaining reliable and quality data;
lack of concrete and actionable guidance to transform strategies
in practice; insufficient monitoring and evaluation mechanisms to
gauge progress and detect risks; risk aversion and regulation uncertainty;
and financial costs.
111. The use of AI by the public sector also raises specific challenges
and opportunities at local and regional level, including concerning
the implementation of smart city initiatives, urban digital twins
and AI‑enabled strategic planning. In particular, effective multilevel
governance and coordination frameworks are required to ensure interoperability
and coherent public action and investment, e.g. through joint public
procurement or financing models; knowledge, data and infrastructure
sharing; or AI governance standards, tools and mechanisms. At the
same time, measures need to be taken to ensure that local and regional
governments have the required capacities and skills to integrate
AI systems into existing processes and services, and manage its associated
risks.
112. In order to seize the benefits of AI, mitigate the risks and
overcome the challenges, the OECD suggests that governments put
in place “enablers” to facilitate a trustworthy AI adoption; “guardrails”
to guide the use of AI, through rules, policies and frameworks;
and “engagement approaches” to involve key actors and shape user-centred
and responsive AI-based services. These comprise the elements of
the OECD Framework for Trustworthy AI in Government, which helps
governments operationalise implementation of the OECD AI Principles.
Human determination and the role of parliaments in terms of regulating
and overseeing AI will also be crucial.
113. The OECD currently has ongoing work that seeks to help governments
close the gaps identified above, while also helping them prepare
for the next generation of AI technologies. In particular, a forthcoming
report on measuring and evaluating AI investments in government,
expected later in 2026, responds to the finding that only around
a quarter of OECD countries currently measure the impact of their
AI use cases, which helps explain why many projects stall at the
pilot stage. Related work under way addresses other challenges highlighted
here, including how governments can experiment with AI safely, build
AI skills across the public service and strengthen the public procurement
of AI. In addition, in promoting readiness for the future, a short series
of reports in preparation examines agentic AI in government (i.e.
systems that can act on a government's behalf rather than only answer
questions), and the use cases, readiness and safeguards governments
will need to adopt agentic AI systems in a trustworthy way.
10 Conclusions
114. The rise of AI and related
technologies is opening a new era of possibilities for OECD countries,
but these are accompanied by a series of challenges and risks that
policymakers need to carefully consider.
115. While there is a growing literature on the impact of AI on
the economy and on labour markets, more must be done in order to
keep up with the fast-paced developments: this will be a challenge
both in terms of research as well as decision-making and implementation.
116. AI can lead to significant productivity gains, as it can improve
the efficiency of work in certain tasks. The degree to which these
gains can be achieved depends on adoption, and therefore on the
data, digital infrastructure, technical, human and managerial skills
and auxiliary technologies available for integrating AI with work.
117. Countries will need to take measures to ensure that the digital
transformation does not leave anyone behind. While it might be too
early to clearly estimate the overall impact of AI on the labour
markets in terms of occupation, it will probably lead to both the
creation and the destruction of jobs, or their deep reconfiguration, with
large differences in impact across territories.
118. Significant investments in training, upskilling and reskilling
(with particular emphasis on digital skills as well as on critical
thinking) will therefore be needed in order to accompany the groups
of workers and regions which are most vulnerable and exposed, either
for mitigating the impact of replacements or for ensuring proper AI
oversight and governance. Policies will also need to be tailored
to consider gender, age, and geographic location of workers, and
will need to be clear and certain in order to facilitate their implementation,
especially by SMEs.
119. Europe needs to boost and innovate its digital ecosystem,
to reduce the gap with countries like the United States or China
in terms of digital infrastructure (data centres; hardware; computing
capacity; advanced models) and investments, and to avoid growing
divides across firms, sectors and places. This has implications in
terms of regulations and economic competitiveness, but also in terms
of digital sovereignty.
120. At the same time, to ensure a trustworthy adoption of AI by
both the private and the public sectors, decision-makers will need
to be guided by transparency, explainability and the respect of
human rights and democratic values. Given the global nature of this
challenge, the importance of international co-operation to ensure
the harmonisation of regulatory frameworks is more relevant than
ever. The oversight role of parliaments will be crucial in this
sense.
121. Finally, it will be equally important to ensure that digital
transformation is properly linked with the net-zero transition:
AI technologies are increasing energy and water consumption exponentially,
but they can also be employed to envisage new methods to increase
environmental protection and energy and material efficiency. Indeed,
AI relies on hardware infrastructure that is highly energy-intensive
and consumes water and critical raw materials, with the carbon footprint
of large models still being significantly underestimated. This environmental cost
manifests itself in the form of increased computing requirements,
the accelerated replacement of equipment, and conflicts over land
use. We must align the digital transition with the ecological transition
by mandating environmental transparency, setting efficiency targets
and promoting suitable architecture, in order to distinguish between
costs that are simply borne and those that can become a future investment
for a sustainable digital technology.
Note
