AI is Reshaping the Labor Market, but Not How People Think

By Dr. Gleb Tsipursky

The AI jobs debate has been stuck in two bad extremes. One side insists a white-collar wipeout is already underway. The other waves away every concern because the unemployment rate has not exploded. In Anthropic’s new report, “Labor market impacts of AI: A new measure and early evidence,” Maxim Massenkoff and Peter McCrory cut through that noise with something far more valuable: a way to track where AI is actually entering work, where it still falls short, and where the earliest damage may appear first.

They argue that labor market analysis needs to move beyond abstract capability and toward observed use. Their central contribution is a metric called “observed exposure,” which combines theoretical task feasibility with real-world Claude usage in professional settings, then weights automated use more heavily than simple assistance. That framing matters because AI disruption will not arrive as a single cinematic layoff event. It will show up through specific tasks, specific occupations, and specific hiring bottlenecks long before the broad labor market fully registers it.

Thus, AI is already reshaping the labor market, but the strongest signal today is weaker entry-level hiring in exposed occupations, not a mass unemployment shock. The evidence is serious. The panic is premature. The implications are immediate.

The Most Important Shift Is Measuring Actual Exposure, Not Hypothetical Potential

The report’s biggest achievement is methodological. Instead of asking only what large language models could do, it asks what they are actually doing in work-related settings. The authors build observed exposure from three ingredients: O*NET task data, Anthropic usage data from the Economic Index, and the theoretical exposure framework from “GPTs are GPTs”. That move instantly makes the discussion more grounded.

The report explains that a task counts as covered when it is theoretically feasible with an LLM and has seen sufficient work-related usage in Claude traffic. Fully automated implementations receive full weight, while augmentative use receives half weight, and those task-level values are then averaged up to occupations using time shares. This is a much stronger lens than simple technical possibility because firms do not adopt tools just because a benchmark says they can. Legal risk, workflow friction, verification needs, and software integration all slow deployment in the real world.

That distinction helps explain why theoretical exposure still overstates immediate labor market risk. In “GPTs are GPTs”, Tyna Eloundou, Sam Manning, Pamela Mishkin, and Daniel Rock estimated that around 80 percent of the U.S. workforce could have at least 10 percent of their tasks affected by LLMs, while roughly 19 percent could see at least 50 percent of tasks affected. The paper mapped enormous potential, but it did not claim that firms had already deployed those capabilities at scale. The Anthropic report shows that this gap between capability and deployment is still large. In Computer & Math occupations, theoretical exposure is 94 percent, but observed current coverage in Anthropic’s data is just 33 percent.

That gap is the story executives should be watching. AI is powerful enough to matter now, but still constrained enough that labor market effects will emerge unevenly. That gives employers and policymakers a narrow but real window to respond before abstract risk turns into entrenched displacement.

The Jobs Under Pressure Are Real, but Broad Unemployment Still Has Not Broken

The report’s occupational ranking is striking because it shows where AI is already most present in real workflows. Computer programmers top the list at 75 percent observed coverage. Customer service representatives come next, followed by data entry keyers at 67 percent. At the other end of the spectrum, 30 percent of workers are in occupations with zero measured coverage, including cooks, mechanics, lifeguards, bartenders, and dishwashers. AI is not sweeping across the labor market in one uniform wave. It is advancing first through digital, language-heavy, structured work.

That pattern lines up with outside research. The online vacancies study by Daron Acemoglu, David Autor, Jonathon Hazell, and Pascual Restrepo found that establishments adopting AI reduced hiring in non-AI positions and changed skill requirements in the remaining postings, even while aggregate employment effects remained too small to detect clearly. The message is simple: labor market change often appears in hiring behavior and job design before it appears in top-line unemployment numbers.

That is exactly what the Anthropic report finds. Using Current Population Survey data, the authors report no systematic increase in unemployment for workers in the most exposed occupations since late 2022. Their difference-in-differences estimate is “small and insignificant,” and they note that a sizable white-collar shock should have been detectable in this framework if it had already occurred. That is a major result because it directly challenges the claim that generative AI has already produced broad labor market collapse.

Other recent work points in the same direction. The Budget Lab at Yale’s running labor market tracker says current measures of AI exposure, automation, and augmentation show no clear relationship with broad changes in employment or unemployment so far, and its later updates through late 2025 continue to describe the patterns as flat or modest. Together, these findings support a more disciplined conclusion: AI pressure is real, but the economy has not yet absorbed it as a broad unemployment shock.

The Clearest Early Warning Signal Is Entry-Level Hiring

The sharpest finding in the report is about younger workers. For workers ages 22 to 25, entry into highly exposed occupations fell by about half a percentage point, which the authors translate into a 14 percent drop in job-finding rates in the post-ChatGPT period relative to 2022. They are careful with the claim and note that the estimate is only barely statistically significant, but it is the strongest sign in the paper that disruption may already be visible in the hiring pipeline.

That result becomes much more persuasive when placed alongside outside evidence. In “Canaries in the Coal Mine?”, Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen use ADP payroll data and find that early-career workers ages 22 to 25 in the most AI-exposed occupations experienced a 13 percent relative decline in employment, with the adjustment driven primarily by slower hiring rather than a surge in separations. The symmetry between the two studies matters. Different data sources. Similar warning sign.

This is where the labor market story gets serious. A stable unemployment rate can hide a shrinking career ladder. If firms keep senior workers, automate pieces of junior work, and hire fewer newcomers, then the damage appears first in the on-ramp. That hurts recent graduates, career switchers, and the long-run development of expertise inside firms. It also fits the logic in “Expertise”, where David Autor and Neil Thompson argue that the labor impact of automation depends heavily on what happens to the expertise value of the tasks that remain. When AI strips away the tasks that once trained junior workers, the labor market does not just lose jobs. It loses apprenticeship.

That concern also helps explain why some markets show faster pain than the broader economy. In Organization Science, Xiang Hui, Oren Reshef, and Luofeng Zhou find that freelancers in highly affected occupations suffered reductions in both employment and earnings after the release of generative AI tools. Online freelance markets are contestable, transparent, and easy to benchmark, so substitution shows up faster there. Traditional employment markets move more slowly, but they are not immune.

The strongest reading of the report is neither doom nor denial. AI has not yet triggered a broad unemployment crisis. That is genuine good news. But exposed occupations are already measurable, projected growth is somewhat weaker in those jobs, and the earliest visible strain is showing up where careers begin, not where they peak. That is the signal leaders should respect. The winners will be the organizations that redesign work around human judgment before the hiring pipeline thins out beneath them.

About the Author