The European Central Bank (ECB) published survey results on 26 August 2026 showing the share of euro-area workers using AI at work rising from 26% in 2024 to 41% in 2025 and 52% in 20261. The education split is sharp: 61% among the university-educated against 37% among those with less education. Median self-reported time saved runs to three hours a week, 7.7% of working time, which the ECB reads as an economy-wide efficiency gain of 3.8%. Positive sentiment about AI fell over the year, from 43% to 41%.
Fedea and BBVA Research found the same population from the other direction. Their quarterly labour bulletin of 17 September reports that more than 40% of Spain's higher-educated workers hold occupations with high exposure to generative AI, against roughly 15% of workers with lower-secondary education or less2. Those occupations now account for more than 30% of Spanish employment and have grown faster than employment as a whole since 2022. Adoption has tracked education in every dataset anyone has published, including the US figures Federal Reserve governor Michael Barr gave in July.
So the occupations most exposed to AI are adding jobs, and the individuals most exposed to AI are finding it harder to get hired into them. Both readings come from first-party releases and neither has been challenged. The likeliest reconciliation is that the adjustment falls on new entrants rather than on people already in post, which would fit Korea's pension register, Japan's graduate hiring plans and the Canadian separation rate at once.
We cannot prove it. No dataset published in this window splits the job-finding differential by age or by first-job status, and the exposure indices built by the BIS, the World Bank and the International Labour Organization score occupations rather than the people queueing to enter them. Until somebody publishes that cut, the mechanism stays an inference and we will keep labelling it one.
