Anthropic's landmark study measuring not just what AI could do to jobs - but what it's actually doing right now. The data challenges almost every assumption about who's at risk, and why.
Of computer programmer tasks can be performed by Claude today
Fewer 22–25 year-olds hired into AI-exposed fields since late 2022
Claude's real task coverage in Computer & Math - vs 94% theoretically possible
Workers in highly exposed roles earn more on average than unexposed peers
AI's theoretical capabilities dramatically outpace its actual adoption. The gap between what LLMs could do and what Claude is doing across occupations is surprisingly wide - and it's the most important number in this study.
The upshot: 97% of Claude's actual task usage falls within theoretically feasible categories - confirming the methodology is sound. The bottleneck isn't capability; it's adoption speed, trust, and workflow integration.
Computer & Math figures confirmed from paper. Other sectors show approximate patterns described in the research.
Exposure cuts cleanly along cognitive vs. physical lines - but the divide isn't what most people expect.
~30% of these occupations have zero AI coverage. All require physical presence, manual skill, or real-time human judgement.
AI can't cook your food, fix your motorcycle, pull you out of a pool, or pour you a drink. The embodied world remains stubbornly human.
Compared to workers with zero AI exposure - pre-ChatGPT, August–October 2022. This baseline reveals who the technology is already reaching.
The inversion: This isn't the automation story of the 20th century, where machines displaced low-wage, physically demanding work. AI is targeting educated, higher-paid, predominantly white-collar workers - a historic reversal of a century of labour economics.
The clearest early evidence isn't a surge in unemployment - it's a quiet narrowing of the entry-level pipeline for workers aged 22–25.
Hiring of 22–25 year-olds into exposed fields vs 2022 baseline.
No statistically significant hiring change for more experienced workers.
No increase in unemployment - displacement isn't happening yet. Just fewer seats.
The mechanism: Companies in AI-exposed fields appear to be hiring fewer junior workers - using AI to bridge the capability gap instead. The jobs aren't disappearing yet, but the traditional path in is getting narrower. This is what early displacement looks like before it shows up in the data.
Four takeaways - whether you're hiring, job-seeking, or building with AI.
The 61-point spread between theoretical (94%) and actual (33%) coverage in Computer & Math tells the real story: AI is far from its perceived potential. Disruption will happen - but more slowly and unevenly than the headlines suggest.
Prior automation waves displaced low-skill, blue-collar workers. AI inverts this completely: graduate-degree holders are 3.5× more represented in highly exposed fields. The cognitive economy is now the exposed economy.
A 14% hiring drop for 22–25 year-olds in AI-exposed fields has emerged with zero unemployment spike. This is what early displacement looks like: quiet narrowing of entry points, not mass layoffs.
Anthropic built this framework before large-scale disruption occurs so we can measure it clearly when it does. This is innings one of a very long game - and we finally have a scoreboard.
800+ US occupations and their associated tasks - the foundation for mapping AI capability to real jobs.
Real usage data from Claude - measuring actual automated, work-related tasks vs augmentative uses.
Task exposure estimates measuring theoretical LLM speedup potential, providing the capability ceiling.
Exposure scoring: β=1 (task completable by LLM alone at 2× speed) · β=0.5 (requires additional tools) · β=0 (not feasible)
Original Source
"A New Measure and Early Evidence"
Published March 5, 2026 · Anthropic · Based on O*NET, Anthropic Economic Index & Eloundou et al.