The AI jobs debate has been trapped between two lazy extremes. One camp insists a white-collar collapse is already underway. The other dismisses every concern because the unemployment rate has not surged.
But 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.
The authors argue that labor market analysis needs to move beyond abstract capability and focus on observed use. They use a measure called observed exposure, which combines theoretical task feasibility with real-world Claude usage in professional settings. The measure also gives more weight to automated use than to simple assistance, which makes it much more useful for evaluating substitution risk.
That distinction matters, because AI disruption will not arrive as one dramatic layoff event. It will spread through specific tasks, occupations, and hiring decisions long before broad labor market indicators fully reflect it.
Many studies have focused on what large language models could theoretically do, and that work remains important. In “GPTs are GPTs,” for example, researchers estimated that about 80 percent of the U. S. workforce could see at least 10 percent of their tasks affected by large language models, while roughly 19 percent could see at least half of their tasks affected.
That paper mapped a huge field of potential exposure, but it did not show that firms had already adopted those tools at scale across real workflows. The Anthropic report closes that gap by asking a tougher question: Where is AI already showing up in actual work in ways that resemble labor substitution?
The answer is revealing. A task counts as covered only when it is both theoretically feasible and sufficiently present in work-related Claude activity. Fully automated use gets full weight. Augmentative use gets half weight. Those task-level measures are then rolled up to occupations using time-shares, which creates a more realistic picture of current exposure. A benchmark result does not equal workplace deployment, since multiple factors slow adoption: legal risk, compliance requirements, software integration, quality control, and the stubborn messiness of real operations.
That helps explain why the report’s current exposure estimates remain well below the theoretical ceiling. In computer and math occupations, theoretical exposure reaches 94 percent, but observed current coverage in Anthropic’s data is only 33 percent. That is a crucial reality-check. AI is powerful enough to matter now, but its labor market effects are still constrained by real-world deployment limits.
The occupational rankings in the report make that precision concrete. Computer programmers top the list, at 75 percent observed coverage. Customer service representatives follow close behind, along with data entry keyers at 67 percent. At the other end, 30 percent of workers are in occupations with zero measured coverage, including cooks, bartenders, dishwashers, lifeguards and mechanics.
AI is clearly not moving through the labor market evenly. It is concentrated first in digital, language-heavy, structured forms of work where outputs are easier to generate, evaluate and integrate.
Importantly, the authors detect no systematic increase in unemployment for workers in the most exposed occupations since late 2022. They note that if a large white-collar employment shock were already underway, their framework should be able to detect it. That directly challenges the popular claim that generative AI has already produced a broad labor market collapse.
Outside research points in the same direction. The Budget Lab at Yale has tracked AI exposure, automation and augmentation against employment outcomes and found no clear relationship with broad changes in overall employment or unemployment so far. Earlier work by Acemoglu, Autor, Hazell, and Restrepo found that AI adoption affected hiring patterns and skill requirements at the establishment level even when aggregate labor market effects remained difficult to detect. The labor market is shifting but not yet breaking broadly.
The strongest warning sign appears somewhere more subtle and more serious: entry-level hiring. For workers ages 22 to 25, the report finds that entry into highly exposed occupations fell by about half a percentage point, which translates to a 14 percent drop in job-finding rates in the post-ChatGPT period relative to 2022. The authors are careful not to oversell this result; they note that it is only barely statistically significant. But it is still the clearest signal in the paper that disruption may already be affecting the hiring pipeline.
That finding becomes more compelling when paired with outside evidence. In “Canaries in the Coal Mine?” authors Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen use ADP payroll data to show that workers ages 22 to 25 in the most AI-exposed occupations experienced a 13 percent relative decline in employment, driven mainly by weaker hiring rather than a spike in separations. Different data, similar signal — the early damage appears to be hitting the career ladder first.
This is where the report becomes especially relevant for employers and policymakers. A stable unemployment rate can mask a shrinking on-ramp into professional work. If firms retain experienced workers, automate pieces of junior work, and reduce entry-level hiring, the labor market can look healthy while access to skill-building quietly erodes. That hurts recent graduates, career-switchers, and firms themselves because apprenticeship is how expertise gets built.
AI has not triggered a broad unemployment crisis. That is good news. But exposed occupations are now measurable, projected growth in those jobs is somewhat weaker, and the earliest visible strain is showing up where careers begin, not where they peak. The labor market impact of AI is no longer hypothetical — it is already sorting workers, tasks and opportunities in real time. The organizations that respond best will be the ones that protect the pipeline of human judgment before it thins out beneath them.
Gleb Tsipursky, Ph. D., serves as the CEO of the future-of-work consultancy Disaster Avoidance Experts and wrote “The Psychology of Generative AI Adoption” (Georgetown University Press, 2026) and “Returning to the Office and Leading Hybrid and Remote Teams” (Intentional Insights, 2021).
Copyright 2026 Nexstar Media Inc. All rights reserved. This material may not be published, broadcast, rewritten, or redistributed.