Most research on AI and jobs asks what AI could do to employment. This study asks what it’s actually doing — right now. That distinction matters more than it might seem.
Anthropic researchers Maxim Massenkoff and Peter McCrory developed a new metric called observed exposure, which combines theoretical AI capability with real-world usage data from Claude. A task only counts as covered if it’s both feasible for an LLM and already appearing in professional, automated workflows.
The findings are grounding. Even in Computer & Math — where 94% of tasks are theoretically feasible for AI — Claude currently covers only 33% of them in practice. The gap between potential and reality is wide across every occupational category.
The most exposed occupations today include computer programmers (74.5%), customer service representatives (70.1%), and data entry keyers (67.1%). Notably, these aren’t low-wage roles. Workers in the most exposed occupations tend to be more educated, better paid, and more likely to be female — earning about 47% more per hour than workers with zero AI exposure.
As for actual unemployment: no significant impact has been detected since ChatGPT’s release in late 2022. The effect on exposed workers is statistically indistinguishable from zero.
There is one early signal worth watching. Hiring of workers aged 22–25 into AI-exposed occupations appears to have slowed by around 14% in the post-ChatGPT era — a result that is barely statistically significant, but echoes findings from other recent studies.
What makes this research valuable isn’t just the data — it’s the framework. By establishing a clear methodology before major disruptions emerge, it gives future analysts a way to separate genuine AI-driven effects from ordinary economic noise.
The clock isn’t ticking toward apocalypse. But it is ticking. And now we have a way to watch it in real time.
Source: Massenkoff, M. & McCrory, P. (2026). Labor market impacts of AI: A new measure and early evidence.
Anthropic. anthropic.com/research/labor-market-impacts