Spanish AI-exposed occupations saw more registered unemployment. Hiring did not clearly fall.

What was counted

The outcome is the monthly count of people registered as unemployed in an occupation, not that occupation’s unemployment rate. The paper also counts new contracts registered each month and finds no persistent fall as exposure rises. It cannot tell whether more people entered unemployment or whether registered jobseekers took longer to leave it.

The authors compare more- and less-exposed occupations within the same broad occupational family and month. Their exposure score maps Claude-use patterns and task descriptions onto Spanish occupations. It does not record AI adoption at Spanish employers.

Why this is not an AI layoff count

The full pre-2022 trend test flags differences between occupations. Excluding the unusual 2021 recovery leaves similar post-release estimates, but that does not remove every competing cause. The later unemployment estimate is imprecise, and the two periods are not statistically different. The study does not identify an aggregate AI effect or individual job losses.

A separate FEDEA–BBVA analysis finds employment grew in all Spanish AI-exposure groups through 2025, with faster growth in the highest group. It measures employment with an ILO-based potential-exposure score, not registered-unemployment counts with the IEB score. Both results can hold at once; neither establishes AI causation.

What would settle the mechanism

Linked worker spells and employer AI-use records could show whether the registered-unemployment difference comes from entries, exits, changes in occupation or adoption. Until then, do not read this percentage as your chance of losing a job or as a reason to buy training.