IEB compared 502 occupations with SEPE records from January 2021 to March 2026 and found 1.2% more unemployment per ten points of AI exposure
Key Takeaways
- Audit tasks, not titles, because AI exposure shows up at the workflow level.
- Choose training that produces portfolio evidence of review, judgment, and domain context.
- Treat the unemployment signal as early evidence, not proof of full role replacement.
Why it matters
- learnersLearners should invest in workflows that prove judgment, not credentials that only rename tool use.
- managersManagers should separate automatable task volume from the human accountability needed to ship reliable work.
For programmers and designers, the useful response is a task map: what can be automated, and where judgment still has to show up.
The job title is not the unit of disruption. The task is. That is the uncomfortable lesson in a new Spanish labor market study that should make programmers, designers, analysts, and anyone with a workflow full of text, code, forms, or patterns stop and inventory what they actually do all day.
The headline number is small enough to avoid panic and clear enough to ignore at your peril. Merca2 reports that the Institut d’Economia de Barcelona compared 502 occupations with SEPE registered unemployment data between January 2021 and March 2026, and detected a 1.2% increase in unemployment for each ten point rise in AI exposure. That is not a verdict on any one worker. It is a warning that generic tasks are becoming easier to substitute, while the hard to observe parts of work need to be made visible.
What the Spanish data actually measures
IEB lists the study as working paper 2026/17, titled AI exposure and registered unemployment in Spain: Early evidence from occupation-level administrative data, by Diego González-González, Nicolás González-Pampillón, Héctor Jiménez Portilla, and Javier Vázquez-Grenno. According to CostedeOportunidad, the authors examined 502 Spanish occupations at the four digit CNO level between January 2021 and March 2026. They looked at two monthly SEPE indicators: people registered as unemployed at month end, and new contracts registered during that month.
That matters because this is not a survey of vibes or a recruiter poll dressed up as research. CostedeOportunidad says the authors used a continuous treatment difference in differences model with two way fixed effects to compare more exposed and less exposed occupations around the arrival of generative AI. The descriptive pattern is narrower than many headlines imply: more exposed occupations saw smaller declines in the stock of registered unemployed people after the AI shock, while new contracts did not show a clear shift in pattern.
Xataka reports the central estimate this way: for every ten point increase in AI exposure, registered unemployment rose 1.2% in the 25 months after ChatGPT launched in late 2022. The same report notes a later 1.7% figure, but says the authors treat it as imprecise and not usable for statistical purposes. Translation for career planning: do not treat this as proof that your job disappears. Treat it as evidence that exposure is starting to show up in administrative labor data.
Exposure is not replacement, but it is not nothing
The measurement choice is important. CostedeOportunidad says the study built its exposure score from Anthropic’s observed exposure index for ONET occupations, then mapped those occupations to Spanish roles by task similarity. Anthropic describes its Economic Index as categorized with ONET-SOC codes, and says its percentages reflect Claude.ai conversations tied to tasks, not job titles.
That distinction is where many career decisions go wrong. An “AI Engineer” label can mean model training, app integration, MLOps, internal automation, or someone who became the office prompt person. Likewise, “programmer” is too blunt a category if one worker mostly translates product ambiguity into architecture while another mostly implements predictable tickets from a clear spec. The exposure score points to task coverage, not full professional replacement.
Still, Xataka reports that programmers lead the exposure ranking, with 75% of tasks covered, followed by data entry workers, sales agents, and financial analysts. Xataka also places software and multimedia analysts and designers in the top 10 of exposed professions. If your work sits in those zones, the right question is not whether to learn a chatbot. It is which parts of your output can be generated, checked, or routed by software, and which parts require accountable judgment.
What hiring managers will screen for next
SEPE’s October 2026 note on training needs gives the other half of the story. SEPE says artificial intelligence was the second most cited technical professional training need, and that needs linked to the technology multiplied by 3.4 from the previous year, with an increase of 243.86%. It also says these needs are not limited to technology jobs, appearing in management, administrative, and education profiles.
That should change how learners pick courses. A certificate that teaches vocabulary without workflow evidence is weak signaling. A stronger portfolio shows how you used AI to speed one part of a process, how you checked its output, what domain rules mattered, and where a human decision was still necessary. For a programmer, that might mean documenting design tradeoffs, testing strategy, security review, and maintenance decisions. For a designer, it might mean showing research synthesis, brand constraints, accessibility decisions, stakeholder negotiation, and final accountability for the result.
This is also why age and stage matter. At 25, you may have more room to rebuild your stack around automation heavy workflows and show projects fast. At 45, you may have stronger domain context, client judgment, regulatory awareness, and organizational memory, but less appetite for a costly credential that does not change your daily work. The hype sells one path to everyone. The labor market usually rewards evidence that matches the role.
Build a task map before buying another course
The practical move is boring, which is usually a good sign. Start with your current role and split it into repeatable production tasks, review tasks, judgment tasks, and accountability tasks. Then mark which ones a general purpose model can draft, classify, summarize, translate, or code against a clear instruction. The exposed tasks are not necessarily the tasks to abandon. They are the tasks where you need to show supervision, quality control, and business context.
CostedeOportunidad cautions that the exposure measure does not directly show how many Spanish workers or firms use AI, or when they adopted it. That caveat is useful. It means a worker should not overreact to a ranking, but should not hide behind a job title either. If your résumé says designer, analyst, programmer, or operations specialist, the stronger version now explains the judgment layer: the constraints you resolve, the errors you catch, the users you understand, and the decisions you can defend.
The IEB paper will not settle the labor market debate by itself. It does give learners a better question than “Will AI take my job?” Watch whether new contracts begin to diverge in future SEPE data, and whether employers start asking for concrete AI review workflows rather than generic tool familiarity. In the meantime, map your tasks before you buy the course. The workers with the clearest evidence of judgment will have a better story than the workers with the longest list of buzzwords.
Sources6 sources
The reporting, announcements and research the AI editor worked from. Links open the original publisher.
- Public Policiesieb.ub.edu
- Estudio del IEB sobre empleo e inteligencia artificial: programadores y diseñadores, los más expuestos al paromerca2.es
- ¿Cómo afecta la Inteligencia Artificial al desempleo registrado en España?costedeoportunidad.com
- La IA ya empieza a dejar huella en el paro español. Los más expuestos: programadores y diseñadoresxataka.com
- The Anthropic Economic Index \ Anthropicanthropic.com
