Much of the public debate about AI and jobs relies on a simple question: which occupations could be automated, based on the tasks they involve? Anthropic's economics team argues that question has been answering the wrong thing. In a report published in March 2026, "Labor market impacts of AI: A new measure and early evidence," the company introduced a different metric — observed exposure — built from how people are actually using Claude at work, not just what the model is theoretically capable of doing.

What "observed exposure" measures

The methodology combines three sources: task-level occupational data from O*NET, Anthropic's own usage data from the Economic Index, and the theoretical-exposure framework used in earlier AI-and-labor research. Critically, it weights automated uses (the AI doing the task directly) more heavily than augmentative ones (the AI helping a person do the task), and it only counts work-related usage. That distinction matters, because theoretical exposure and observed exposure turn out to diverge sharply in practice.

The clearest example the report gives is Computer & Math occupations. Theoretical frameworks suggest AI could touch roughly 94% of the tasks in that category. Anthropic's observed data puts actual Claude usage coverage closer to 30-33% of those tasks. In other words, the technical capability exists far ahead of where usage currently is — a gap that should temper predictions built purely on what models can theoretically do.

No broad unemployment spike — yet

The report's central finding is that, as of its publication, there's no systematic increase in unemployment among workers in highly AI-exposed occupations since late 2022. That's a meaningful data point against the sharper "AI is already replacing workers at scale" narratives. But the report is careful not to read that as "nothing is happening" — it also finds suggestive evidence that hiring of younger workers has slowed specifically in the occupations most exposed to AI, even without broad layoffs showing up yet.

Who is actually most exposed

One of the more counterintuitive findings is about who faces the highest observed exposure. It isn't entry-level or lower-paid workers, as the earlier conventional wisdom around automation assumed. Anthropic's data shows highly exposed workers are more likely to be older, more educated, and higher-paid — because AI usage today skews toward knowledge work like analysis, writing, and coding rather than physical or lower-wage service tasks.

What this means if you're planning your next career move

• If you're early-career in a highly AI-exposed field (software engineering, data analysis, research-heavy roles), the hiring-slowdown signal is worth taking seriously even without layoff headlines — competition for entry-level openings in these fields may be tighter than raw job-posting counts suggest.
• Experience isn't automatically a shield. Since exposure skews toward more senior, more educated workers, a long track record doesn't exempt you from needing to actively demonstrate how you use AI tools in your own work, rather than assuming your seniority alone makes you AI-resistant.
• The theoretical-vs-observed gap will likely narrow over time, not stay fixed. A 30% usage rate against a 94% technical ceiling in coding-adjacent work suggests considerable room for AI adoption to keep climbing in exactly the fields where it's already highest — worth factoring into multi-year career planning, even if today's numbers look stable.

The report is a useful corrective to both extremes of the AI-jobs debate: it doesn't support claims of mass displacement already underway, but it also doesn't support the idea that AI's labor effects are purely theoretical. The real story, at least so far, is a slow shift in hiring patterns among the most-exposed, most-educated segment of the workforce — not a wave of layoffs.