AI and work, country by country

Most of what circulates about AI and work is American evidence with the nationality removed. The national statistical offices, labour ministries and research institutes that have measured this in their own economies reach different conclusions, and they differ in a pattern: adoption is far lower than the debate assumes, exposure is set by what a country's people do for a living rather than by what the technology can do, and where effects appear at all they fall on the young and the educated rather than on the low-skilled.

International is a position, not a category

A word about the label before the evidence, because the label does some quiet work.

This research is written from the United Kingdom. Filing everywhere else under international is a view from one place wearing a neutral badge, and it produces a specific error: the United States and the United Kingdom stop being countries with particular labour markets, particular demographics and particular blind spots, and become the default against which everywhere else is a variation.

They are not the default. The United States has unusually weak employment protection, unusually high wage dispersion and an unusually large technology sector. Those are three good reasons to expect American findings to travel badly, and the evidence below suggests they do.

One further commitment, because it changes what this page can see. The sources here are read in the language they were published in. The German, French, Spanish, Japanese, Korean and Chinese material is cited from the original rather than from English reporting about it, which matters because the English-language summaries of these studies are frequently thinner than the studies, and occasionally say something the study does not.

Almost nobody is using it yet

The single most consistent finding across national surveys is that adoption is a fraction of what the discourse implies.

Japan's Institute for Labour Policy and Training surveyed 22,000 employees, stratified on the 2020 Census, with the OECD involved in the design. It found 12.9 per cent reporting any AI use by their employer and 8.4 per cent using it themselves.

The Korea Development Institute put the two halves of this side by side and the gap is the finding. Its assessment scored 38.8 per cent of Korean jobs as technically automatable across more than 70 per cent of their tasks. Its September 2023 survey of 800 firms found that 2.7 per cent of firms with ten or more staff had adopted AI.

Thirty-nine per cent possible against under three per cent actual. Almost every widely circulated number in this field measures the first quantity and gets discussed as though it described the second. The same distinction is examined at the most-quoted AI statistics, checked.

Germany complicates the picture usefully. The DiWaBe 2.0 survey of roughly 9,800 employees found more than half already using AI at work, though largely informally, ranging from about a third of unqualified workers to around 80 per cent of those with a degree or Meister qualification. Informal, unmeasured, personally adopted use is a different phenomenon from the sanctioned deployment the Japanese and Korean surveys asked about, and it is almost certainly the bigger one.

What a country does for a living decides its exposure

Funcas remapped the Felten occupational exposure index onto Spanish occupational classifications and combined it with the Q4 2025 Labour Force Survey. Spain shows medium-high exposure at 27.4 per cent and low automation risk at 5.9 per cent, against an OECD average near 12 per cent.

The reason is structural. Spain's employment is weighted towards interpersonal and physical work, and an economy built on those is less automatable whatever the models can do. Exposure is a property of a country's occupational mix rather than of the technology, which means every global exposure figure conceals enormous national variation, and applying an OECD average to a specific country is close to meaningless.

Germany inverts the assumption that automation reaches the least skilled first. The IAB's substitutability assessment, in which three independent coders score more than 9,000 tasks across roughly 4,600 occupations, found substitutability rising about ten percentage points for degree-level expert occupations between 2019 and 2022, and roughly flat for helper occupations. The IAB frames AI as relief for skills shortages rather than as displacement, which is a reasonable reading in an economy with Germany's demographics.

Where effects appear, they fall on the young and the educated

This is the finding that recurs across otherwise dissimilar economies, and it is the opposite of what twenty years of automation commentary trained everyone to expect.

Korea's realised effects showed no aggregate employment change, lower earnings, and the impact concentrated on younger, tertiary-educated workers and women.

France's national statistics office reached the same shape independently. INSEE found employment of 15 to 29 year olds, excluding apprentices, falling 7.4 per cent year on year in IT services, 5.8 per cent in publishing and 3.7 per cent in management consulting in Q4 2025, against minus 0.7 per cent across the market sector overall. INSEE explicitly cautions against attributing the fall to AI alone, and that caution is part of the finding rather than a footnote to it.

A European national-statistics office finding the same entry-level pattern reported in American payroll data makes the signal considerably harder to dismiss as an artefact of the US technology sector. The consequences for the talent pipeline are set out at missing rungs and will AI replace entry-level jobs.

Denmark, and the slowest indicator available

Humlum and Vestergaard linked adoption surveys to administrative records for roughly 25,000 workers across 7,000 Danish workplaces in eleven exposed occupations. Two years after ChatGPT they found precise null effects on earnings and hours, ruling out effects larger than 2 per cent, alongside substantial task reorganisation and entirely new tasks in AI oversight and integration.

Read those two results together and they are not in tension. The structure of work moved and pay did not. Pay is the slowest indicator anyone has, and a study that looks only at earnings will report nothing happening for years after a great deal has happened.

Denmark is high-trust, high-wage and heavily unionised, which the authors flag as a limit on generalisation. It is also, for the same reasons, close to a best case: if capability erosion shows up even there, the institutional protections are not sufficient.

The same technology, opposite anxieties

In rich economies the worry is that AI will do too much. Across most of the world it inverts.

CEPAL's modelling for Latin America concludes that the gains run through skilled labour and that the binding constraint across the region is human-capital formation and low investment. The regional risk is under-adoption rather than displacement. A debate conducted entirely in the vocabulary of protecting jobs from automation has nothing to offer a country whose problem is that the productivity gains are not arriving.

India is different again, and is the most useful corrective on this page. Azim Premji University's analysis of Periodic Labour Force Survey data finds roughly five million graduates entering the labour market each year against about 2.8 million finding work, with graduate unemployment near 40 per cent among 15 to 25 year olds. The report explicitly declines to attribute this to AI, because it is a demand-side bottleneck that long predates it.

That refusal is worth borrowing. Every graduate hiring problem now gets read as an AI story, and in the world's most populous labour market the people closest to the data say it is not one.

China provides the fourth position. A Chinese Academy of Sciences analysis finds that between 2018 and 2023 substitution outweighed complementarity, with a one per cent rise in industrial robots reducing firm labour demand by 0.18 per cent, and notes roughly 200 million people, 27 per cent of employment, in flexible work with 37 per cent social-insurance coverage. The policy response centres on social security redesign rather than on retraining, which is a different answer from the one every Western government has reached for. Note the caveat: this is largely industrial robotics rather than service-sector generative AI, and the two do not transfer cleanly.

What Germany and France measured that nobody else did

Two findings deserve separating out, because they go to how AI is introduced rather than to how much of it there is.

The German DiWaBe survey found no difference in training participation between AI users and non-users. That is adoption without redesign, measured at national scale on a representative sample of 9,800 employees. It is the clearest empirical statement of the pattern this research calls drift rather than design.

France's LaborIA study, run by the Ministry of Labour with Inria, named something the survey literature otherwise misses. It identifies a conflit de rationalité, a clash of rationalities: managers justify AI by error reduction (81 per cent), performance (75 per cent) and removing drudgery (74 per cent), while the ethnographic fieldwork shows workers becoming the system's de facto trainers. What management believes the tool is doing and what workers experience it doing diverge systematically inside the same organisation. The qualitative core rests on six sites and ten repeated interviews, so it is a sharp hypothesis rather than a measured rate.

Japan supplies the constructive counterpart. Among Japanese users, reports of improved job quality outweighed reports of decline, and the gain was markedly larger where the employer had consulted staff and funded training. From a 22,000-person sample, that is direct support for the argument that how AI is introduced determines its effect on people more than the technology does.

What this research covers, and what it does not

Set out in enough detail to be argued with, because uneven coverage presented as a world view is its own kind of error.

• Graded national evidence, read in the original language: Japan, Germany (two sources), France (two), Korea, China, India, Spain, Latin America regionally, and Denmark. Eleven entries across nine countries and one region.
• Covered but without a country page: the United States and the United Kingdom. Both are heavily represented in the evidence base through labour-market and experimental work, and neither has been written up as a national profile. On the argument at the top of this page, both should be.
• Not covered, and it is a real gap: the Gulf. The UAE, Qatar and Saudi Arabia have substantial national AI strategies and investment commitments, and they do not yet have the measured labour-market studies that Germany, Japan and Denmark do. A page built on strategy documents and investment announcements would carry a materially weaker evidence grade than the rest of this estate, and it will say so when it is built.
• Thin everywhere: Africa, southeast Asia, and central and eastern Europe. The absence here reflects what this research has read rather than what exists.
• No cross-national study has yet compared capability retention. Every source on this page measures adoption, exposure, employment or earnings. Not one measures whether people can still do the work unaided, which is the question this research regards as the important one.

How to read a national claim about AI

• Ask whether the number is potential or realised. Korea's 38.8 per cent and 2.7 per cent are both true and they describe different worlds. Most quoted figures are the first kind.
• Ask what the country does for a living. Spain's low automation risk is an occupational-mix fact. An OECD average tells you very little about any member of it.
• Ask what the institutions do. Denmark's null result on pay is not evidence that nothing happened; it is evidence about Danish wage-setting.
• Ask whether pay was the only outcome measured. If so, expect it to show nothing for years while task structure moves underneath.
• Ask whether the authors attributed the effect to AI, or whether someone else did it for them. INSEE and Azim Premji University both declined. Their findings are quoted as AI evidence anyway.

Key research and primary sources

• Japan Institute for Labour Policy and Training (2025). Survey on the impact of workplace AI adoption on working styles, Research Series No. 256, in Japanese.
• IAB (2024). Folgen des technologischen Wandels fur den Arbeitsmarkt, IAB-Kurzbericht 5/2024, in German, and BAuA, ZEW, IAB and BIBB (2025). DiWaBe 2.0, in German.
• LaborIA (2024). Etude des impacts de l'IA sur le travail, in French, and INSEE (2026). Note de conjoncture, March 2026, in French.
• Korea Development Institute (2023). Changes in the labour market due to artificial intelligence, Research Report 2023-03, in Korean.
• Lu, Y. and Gui, L. (2025). Analysis of the impact of artificial intelligence technology on employment and income in China, Bulletin of the Chinese Academy of Sciences, 40(4), in Chinese.
• Azim Premji University (2026). State of Working India 2026, and Funcas (2026). Inteligencia artificial y mercado de trabajo en Espana, in Spanish.
• CEPAL (2026). Impacto economico de la inteligencia artificial en America Latina, in Spanish.
• Humlum, A. and Vestergaard, E. (2025). Still Waters, Rapid Currents, NBER Working Paper 33777.

Related SuperSkills research

The country study in depth, AI and work in Japan. On the numbers, the most-quoted AI statistics, checked and what we actually know. On the pattern these findings describe, drift versus design and measuring adoption properly. On the early-career signal, entry-level jobs and missing rungs. On working across markets, global adaptability.

About this research

Rahim Hirji is the author of SuperSkills (Kogan Page, 2026) and founder of The SuperSkills Intelligence Company. Every source on this page carries a graded entry in the evidence base with its method, sample and limits recorded, and the German, French, Spanish, Japanese, Korean and Chinese material is cited from the original publication rather than from English coverage of it. Where an author declined to attribute a finding to AI, that refusal is reported rather than quietly dropped. The coverage gaps are listed above rather than left for a reader to discover. Reviewed quarterly.