Is AI already eliminating jobs? What the data from Europe show
We examined 13 countries and 39 occupational groups: did exposed occupations lose more jobs in places where firms adopted AI faster? The answer for now: no effect can be detected.
In data from 13 European countries, we found no statistically significant relationship between AI adoption by firms and a decline in employment in exposed occupations. The estimate leans toward lower employment, but it is not different from zero. This does not mean there is no effect, but that it is still too early to measure it.
The question
If AI replaces workers, we would expect to see a decline in employment specifically in occupations where AI can perform many tasks, and specifically in countries where firms are adopting AI quickly. That is the comparison we examined: not whether employment declined in general, but whether it declined more where exposure and adoption meet.
What did we examine?
We combined three sources: the share of firms using AI in each country (Eurostat), the exposure score of each occupational group (ILO 2025), and the number of workers in each occupation in each country and year (the European Labour Force Survey). After filtering out observations with missing data or data flagged as unreliable, 848 observations remained from 13 countries and 39 occupational groups, for changes ending in 2024 and 2025. The model controls for fixed differences between occupations and for shocks common to each country in each year.
What did we find?
The estimate is negative: it leans toward lower employment in exposed occupations as adoption increases. But the uncertainty is larger than the estimate itself. The confidence interval includes both a negative and a positive effect, and the probability of obtaining such a result even with no real relationship is high (p=0.32).
Why can't we conclude more?
Three main limitations. First, adoption is measured at the level of the country as a whole, not at the occupation level. Second, there are only two annual changes, and generative AI entered widespread use only recently. Third, AI adoption and employment can affect each other and third factors, so even a significant relationship would not prove causality.
And what about Israel?
This analysis is based on European data only. The ILO exposure scores are global, but we currently have no data on AI adoption and employment by occupation in Israel at a similar level of detail. Therefore, no conclusions should be drawn from this about what is happening in the Israeli labour market.
What is needed to know more?
To answer better, we need adoption data by occupation or by firm, a longer period, verified timing of when the tools were introduced, and placebo tests that check that the method does not find effects where there should be none. Until then, the research decision is not to pass this estimate on to the scenario model.
Sources and transparency
This publication explains the existing research in accessible language. Its accounting example and scenarios are neither official forecasts nor investment recommendations.
What do you think?
In your view, what data would help better answer the question of whether AI is replacing workers?