Anthropic Reveals Jobs Most Threatened by AI

A New Metric for Measuring Exposure to AI

The question of whether artificial intelligence can replace our jobs is becoming increasingly urgent. On March 5, 2026, Anthropic published a study titled The Impacts of AI on the Labor Market, which introduces a new indicator called observed exposure. This indicator cross-references the theoretical capabilities of large language models (LLMs) with real usage data from Claude, a model developed by Anthropic. This approach aims to provide nuance to the often alarming conclusions of other recent studies.

How Observed Exposure is Constructed

Observed exposure differs from previous methods by measuring what AI actually automates in professional contexts, rather than limiting itself to what it could theoretically handle. To achieve this, Anthropic uses three main data sources:

• The O*NET database, which catalogs tasks associated with 800 American occupations.
• Usage data from Claude, which relies on the Anthropic Economic Index to determine which tasks are actually performed with an LLM in a professional setting.
• The metric from Eloundou et al. (2023), which assesses whether an LLM can accelerate the execution of a task by at least 50%.

A graph included in the study illustrates the gap between the theoretical potential of AI and its actual usage. In blue, the tasks that can be automated or accelerated by AI; in red, those that Claude is already handling. For example, in the fields of computer and mathematical occupations, 94% of tasks could be automated, but only 33% currently are. Anthropic emphasizes that this gap may narrow as AI progresses and its adoption intensifies. However, Anthropic believes that certain tasks will remain "obviously" beyond the reach of AI, regardless of technological advancements.

Are Highly Skilled Jobs on the Front Line?

Using observed exposure, Anthropic has ranked the occupations most exposed to automation by AI. Computer programmers top the list with a coverage of 74.5%, followed by customer service representatives at 71.1%. At the other end of the spectrum, 30% of workers, whose jobs require physical presence or manual skills, are not affected by automation. These occupations include cooks, mechanics, lifeguards, and divers, whose tasks appear too infrequently in the data to meet the minimum automation threshold.

Anthropic's study also outlines a typical profile of the most exposed workers: they are generally older, female, more educated, and better paid. On average, their salaries are 47% higher than those in the least exposed occupations.

Recruiters Are Already Hesitant About Exposed Junior Profiles

Although Anthropic urges caution in interpreting these results, the study does not observe a significant increase in unemployment since late 2022 in the professions most exposed in the United States. However, a concerning signal is emerging: hiring of young individuals aged 22 to 25 in these occupations has declined by about 14% since the emergence of generative AI tools. This suggests that companies are already less inclined to hire junior profiles for roles likely to be automated.