Anthropic Looks for Early AI Signals in Jobs and Hiring

On March 5, 2026, Anthropic published a study of AI exposure and labor markets. Its design does not causally identify how many jobs AI created or destroyed. It compares occupations using a Claude-usage measure: it detects no differential rise in unemployment among the most exposed and finds a barely significant signal of lower entry by 22-to-25-year-olds into those occupations.

Labor market impacts of AI: A new measure and early evidence proposes a method for observing early signals. The authors say it does not capture every channel through which AI may reshape employment and is most useful when effects are ambiguous. That limitation comes before the percentages: no detectable difference is not the same as no effect.

Why previous predictions failed

The study begins by examining why technical capability or estimated exposure does not automatically become lost employment.

It cites the example of offshoring: one influential analysis estimated that a quarter of U.S. jobs were vulnerable to being sent abroad. A decade later, most of those positions maintained healthy employment growth. Even the U.S. government's own forecasts, the paper admits, have added little predictive value beyond linearly extrapolating past trends.

The lesson is clear: separating a technology's effect from the noise of the business cycle is devilishly hard. Studies on the impact of industrial robots reach opposite conclusions, and the debate over how many jobs the China trade shock destroyed remains open.

Anthropic places AI in that ambiguous category. It doesn't expect a sudden, obvious blow like the pandemic—when unemployment spiked within weeks and no sophisticated statistics were needed to see the cause—but something more like the internet or trade with China: a slow, underlying shift that's hard to isolate in aggregate figures.

The core idea: observed exposure

The paper contributes a metric called observed exposure. Theoretical exposure asks which tasks a model could accelerate; the new measure adds which tasks appear in Claude traffic, in conversations classified as work-related, with greater weight when use appears to automate rather than assist. “Observed” means observed on that platform and sample, not across the entire economy.

To build it, the researchers combine three sources:

• O*NET, the U.S. occupational database that breaks roughly 800 occupations into tasks.
• Claude usage data from the August and November 2025 Anthropic Economic Index samples; these do not measure ChatGPT, Gemini, or internal software that does not pass through Claude.
• Eloundou et al. (2023) estimates, scoring whether an LLM of that period could cut task time by at least half, alone or with tools.

That theoretical score, called β, is simple: it equals 1 if the model can speed up the task on its own, 0.5 if it needs additional tools built on top of it, and 0 if it can't.

The gap between capability and use is central. Eloundou et al. assign β=1 to “authorizing prescription refills and notifying pharmacies,” while Anthropic says it has not observed Claude doing it. The article lists possible explanations—model limits, law, software, or human verification—but does not identify which explains that case. The example therefore establishes absence in the sample, not practical automability.

Even so, 97% of observed Claude use falls within tasks Eloundou et al. classified as theoretically feasible. The overlap validates that the measures are related; it does not show that 97% of all jobs can be automated.

AI is far from its ceiling

In Computer and Mathematical occupations, the theoretical measure covers 94% of tasks while observed Claude coverage reaches 33%. In Office and Administrative Support, theory reaches 90% and observed coverage remains much lower. These are task-weighted percentages within the method, not shares of workers replaced or hours already automated.

In other words: even though headlines claim "AI can do almost everything," its actual penetration into work is a fraction of what's technically possible.

The most exposed occupations

The ranking produced by that measure places these occupations near the top:

• Computer programmers, leading the pack with 75% task coverage.
• Customer service representatives, whose functions increasingly show up in Claude's API traffic.
• Data entry operators, at 67%, whose core task—reading documents and transcribing them—is easily automated.

At the opposite end, 30% of workers have zero coverage: their tasks barely appear in Claude's data. That includes cooks, motorcycle mechanics, lifeguards, bartenders, and dishwashers. Physical or in-person jobs that, for now, remain beyond a language model's reach.

Who's in the line of fire

After linking its measure to the Current Population Survey, the study finds that the most exposed group is older on average and contains more women, more education, and higher earnings than the zero-exposure group. This describes groups with different occupational composition; it does not mean age, sex, education, or pay causes exposure.

The comparison needs a baseline. The more exposed group contains more office work; the zero-exposure group contains more physical and in-person tasks. Separating composition from effect prevents a correlation between occupation and exposure from becoming an individual risk claim.

And employment? Holding up, for now

Anthropic cross-referenced its measure with the BLS 2024-2034 occupational projections. Those projections are another estimate, not observed future employment, and the BLS program publishes assumptions and occupation-level data for inspection.

In the weighted regression, ten percentage points more coverage is associated with 0.6 points less projected employment growth. The paper calls the relationship slight and says it does not appear with the theoretical measure alone. This is a correlation between two constructed measures; it does not prove Claude use caused the BLS projection or that the metric has already predicted future growth.

Using CPS data since 2016, the average post-2022 change in the unemployment gap between the top exposure quartile and zero-exposure group is small and indistinguishable from zero. The authors estimate that their design would detect roughly a one-percentage-point differential increase. Failure to detect less than that does not establish an exactly zero effect.

For 22-to-25-year-olds, the study estimates a 14% post-ChatGPT decline relative to 2022 in the rate of starting exposed jobs, barely statistically significant. The authors warn that these people may remain in existing jobs, enter other occupations, return to school, or leave the labor force, and that surveys may mismeasure transitions. On March 8 they also corrected Figure 7 because its labels were reversed. This is a signal to replicate, not a settled causal attribution.

What this study means

The study should be read with two caveats. The first concerns its source: it's published by Anthropic, a company whose business is selling AI models, and its usage data comes exclusively from Claude. That leaves out ChatGPT, Gemini, and the rest of the market, and limits the view to what happens on its own platform.

The second caveat is methodological. The appendix exposes judgment calls about what counts as sufficient use, how similar tasks are grouped, and how much more weight automation receives than augmentation. Occupational rankings are robust to several alternatives, but each choice defines “exposure” and must remain visible when editions are compared.

The result sits between two claims the data do not support: it neither demonstrates a labor apocalypse nor proves aggregate employment is protected. It shows a gap between theoretical capability and Claude use, no unemployment difference detectable with this design, and a fragile young-hiring signal. Each conclusion belongs to a specific variable and population.

The transferable skill is to read a labor claim in four layers: assumed technical capability, observed use, measured labor outcome, and the counterfactual needed for attribution. Exposure can guide where to look, but it cannot replace unemployment, hiring, wages, or causality. A useful update must preserve definitions, show intervals, log corrections, and compare Claude with other tools and labor sources.