Since the launch of ChatGPT in late 2022, studies have been multiplying to try to gauge the impact of generative AI on the job market. Most rely on theoretical estimates of the capabilities of large language models (LLMs).
Anthropic, the creator of Claude, offers a different approach, and the results are more nuanced than the catastrophe scenarios often amplified.
A new indicator: the “observed exposure”
The study, titled “Labor Market Impacts of AI: A New Measure and Early Evidence,” introduces a new measure of displacement risk, called observed exposure, which blends the theoretical capability of LLMs with real usage data, giving greater weight to automated uses (as opposed to augmentative uses) and to professional contexts.
The researchers, Maxim Massenkoff and Peter McCrory, begin with a simple observation: purely theoretical measures systematically overstate the real risk. AI is far from reaching its theoretical capabilities because real-world coverage represents only a fraction of what would be technically feasible.
To construct their indicator, they cross three sources: the O*NET database, which catalogs tasks associated with roughly 800 occupations in the United States; Claude usage data (from the Anthropic Economic Index); and the theoretical estimates from Eloundou et al. (2023).
By way of example, Claude currently covers only 33% of tasks in the “Computing & Mathematics” category, whereas theoretical estimates suggest that 94% of these tasks would be technically accessible to an LLM. The gap between potential and reality is thus substantial.
Ten occupations are highly exposed
Among the ten most exposed occupations, developers top the list: 75% of their tasks are already covered by observed AI usage. They are followed by customer service agents, whose interactions are increasingly handled by automated systems, and data entry operators, covered at 67%.
On the other hand, nearly one in three workers is not currently affected by AI. These are everyday, on-the-ground jobs: cooks, mechanics, lifeguards, bartenders, or restaurant dishwashers.
The unexpected profile of exposed workers
The study draws an unexpected portrait of the employees most at risk. It is more often women, older workers, graduates, and higher earners who are exposed than low-skilled workers.
The figures speak for themselves: in the most exposed occupations, women are overrepresented by 16 percentage points compared with non-exposed occupations, white people by 11 points, and people of Asian origin are almost twice as numerous.
In terms of pay, the gap is also notable: exposed workers earn on average 47% more. As for education, holders of a master’s degree or a PhD account for 17.4% of the exposed, versus only 4.5% of the non-exposed.
No measurable impact on unemployment… but
Perhaps the most important conclusion of the study: there is no abnormal rise in unemployment observable in the occupations most exposed to AI since the end of 2022. The researchers note that their method would have detected a shock on the scale of the 2008 crisis, when U.S. unemployment doubled. The absence of a signal is therefore not a statistical artefact.
However, they find signs that hiring of young workers has slowed in exposed occupations. Since 2024, 22- to 25-year-olds have more difficulty securing a job there: the monthly hiring rate has fallen by about half a percentage point, whereas other sectors remain stable. The decline reaches 14% across the entire post-ChatGPT period.
That signal is admittedly fragile but consistent with other recent studies, which attribute the slowdown not to waves of layoffs, but to a simple freeze in recruitment.
A study set to be updated
The authors explicitly present this study as a starting point. Their aim is to develop a robust method before the effects of AI are fully visible, so as to better identify them when they manifest.
They plan to update their indicators as new data come in, and indicate they will soon examine the case of young graduates in the most exposed fields. A population for whom the slowdown in hiring could have lasting consequences on early career development.
