Anthropic researchers Maxim Massenkoff and Peter McCrory published a study measuring AI's labour market footprint through actual professional Claude usage rather than theoretical capability assessments1. They introduced "observed exposure" — a metric comparing what AI is actually being used for in workplaces against what it could theoretically perform. The headline numbers: computer programmers face 75% task coverage; computer and maths occupations 35.8%; office and administrative roles 34.3%. The demographic profile of the most-exposed workers overturns the popular image of displacement running downhill to the least skilled. They are "older, female, more educated and higher-paid."
The methodology fills a gap that has weakened earlier research. The most widely cited exposure studies — including the Eloundou et al. GPT-4 assessment published in 2023 — measured what AI could theoretically do if deployed at full capacity. Massenkoff and McCrory measured what is actually happening, using anonymised professional usage data from Claude. The distance between theoretical and observed exposure is where companies, workers, and policymakers need to focus. LLM adoption among US workers rose from 30.1% in December 2024 to 38.3% by December 2025, but adoption and displacement are different phenomena. Oxford Economics concluded in January 2026 that AI's role in layoffs may be "overstated," finding firms do not appear to be replacing workers with AI at scale. The Anthropic data reframes the question: the issue is less whether displacement is happening than where it is concentrated and who bears it.
No systematic unemployment increase has appeared among heavily exposed occupations since late 2022. That finding sits in tension with the drumbeat of corporate layoff announcements — 45,363 confirmed global tech layoffs in Q1 2026, of which 9,238 cite AI explicitly. But the "suggestive evidence" of slowing hiring among younger workers aligns with the Dallas Fed's data on collapsed job-finding rates for under-25s2. The pattern across both data sets is consistent: AI is restructuring work at the hiring margin — fewer new positions, changed job descriptions, shifted task allocations — rather than generating mass terminations. Harvard Business Review research by Thomas H. Davenport and Laks Srinivasan found only approximately 2% of organisations reported layoffs tied to actual AI implementation; the remainder cut in anticipation of capability that does not yet exist3.
The demographic skew carries policy implications that current proposals do not address. Bipartisan AI disclosure requirements and Senator Bernie Sanders' proposed robot tax both implicitly frame displacement as a problem for lower-wage, less-educated workers — the constituency that previous automation waves hit hardest. If AI's actual exposure falls most heavily on workers who already hold degrees and earn above-median wages, retraining programmes built around upskilling the least educated will miss the population bearing the greatest impact. The Anthropic data does not settle the displacement debate, but it does something more useful: it grounds the debate in what AI is actually doing, rather than what it might theoretically do.
