Programmers, First Exposed to AI According to Anthropic: What the Measurement of "Observed Exposure" Reveals Across 800 American Occupations

On March 5, 2026, Anthropic published a study titled Labor market impacts of AI: A new measure and early evidence, authored by economists Maxim Massenkoff and Peter McCrory. This study introduces a new concept into the debate on AI and employment: observed exposure. The idea is simple yet powerful: instead of measuring what AI could theoretically do, it measures what it actually does in professional contexts. The result is a ranking of 800 American occupations according to their actual degree of exposure to automation by large language models.

Observed Exposure: A Metric That Changes the Conversation

Since 2023, the debate on AI's impact on employment has largely relied on the work of Eloundou et al. (OpenAI), who estimated the proportion of professional tasks theoretically accelerable by an LLM. Anthropic proposes to bridge the gap between potential and reality by cross-referencing three data sources: the O*NET database from the U.S. Department of Labor (800 occupations), real Claude usage data from the Anthropic Economic Index, and the theoretical metric from Eloundou et al.

Observed exposure works as follows: a task is considered "covered" if it is theoretically achievable by an LLM and if it appears with sufficient frequency in Claude's professional usage data. Automated use receives full weight; augmentative use receives half weight.

The Ranking: Who Is Exposed, Who Is Not

In the "computer and mathematical" category, 94% of tasks are deemed theoretically automatable. But observed exposure is only 33%. AI is far from having reached its potential.

At the other end, 30% of American workers have an observed exposure of zero. Occupations requiring physical presence, manual expertise, or direct human interaction remain beyond AI's reach.

The Typical Profile of an Exposed Worker: A Social Paradox

Workers in the most exposed quartile are 16 percentage points more likely to be women, 11 points more likely to be white, and almost twice as likely to be Asian. Their wages are on average 47% higher. Postgraduate degree holders represent 17.4% of the most exposed group, compared to 4.5% of the non-exposed group.

This finding has profound implications for public policy. Professional retraining programs are traditionally designed for low-skilled workers. If AI primarily affects educated white-collar workers, these programs will need to be rethought.

No Mass Unemployment, But a Warning Signal for Young People

Comparing employment trends since late 2022, the authors find no systematic increase in unemployment in the most exposed occupations. However, they observe a 14% decrease in the hiring rate for 22-25 year olds in the most exposed occupations. This decrease is not observed among those over 25.

AI does not destroy existing jobs — it slows the creation of new positions for labor market entrants.

Criticisms

Forbes criticized the fact that the data comes exclusively from Claude's usage. Ethan Batraski argues that the relevant distinction is not between automatable and non-automatable tasks, but between time-billed jobs and results-based jobs. The Brookings Institution reminds us that research on AI and the labor market is still in its early stages.

The Theoretical-Real Gap: A Window for Adaptation

In computer science, the gap is 61 points (94% theoretical, 33% observed). These gaps represent a window for adaptation — the time workers, businesses, and governments have to prepare. The question is not whether the red zone will catch up to the blue zone, but how quickly.

The Anthropic study answers a precise question: in March 2026, which occupations are actually exposed to automation by LLMs? It does not answer the long-term question. Cataclysmic predictions and reassuring denials are equally premature. What is needed are regular, transparent, and methodologically rigorous empirical measurements.