Anthropic's new labor measure finds AI hasn't moved the unemployment needle

Anthropic published new research in March introducing a measure of AI displacement risk that blends theoretical capability with real-world usage data. The headline finding: no systematic increase in unemployment for highly exposed workers since late 2022. But the paper also finds suggestive evidence that hiring of younger workers has slowed in exposed occupations, and that the Bureau of Labor Statistics projects weaker growth through 2034 for jobs with higher observed exposure.

The paper is a joint effort by Anthropic’s economic research team. It builds on the company’s Economic Index, which tracks how Claude is used in professional settings. The new metric, called observed exposure, weights automated use patterns more heavily than augmentative ones, and only counts work-related usage. The goal is to measure not just what AI could do, but what it is actually doing in the labor market right now.

The gap between theory and reality is large. Eloundou et al. (2023) rated 94% of tasks in Computer and Math occupations as theoretically feasible for an LLM. Claude’s actual coverage in that category: 33%. Across all occupations, 97% of observed Claude usage falls into tasks rated as theoretically feasible, but the red area on Anthropic’s chart — actual usage — covers far less ground than the blue area of theoretical capability. The paper notes that “AI is far from reaching its theoretical capability.”

That gap matters for the debate about AI and jobs. Many forecasts of mass displacement assume that capability equals deployment. Anthropic’s data suggests the timeline is longer. Legal constraints, software requirements, human verification steps, and model limitations all slow diffusion. The paper gives an example: “Authorize drug refills and provide prescription information to pharmacies” is rated fully feasible, but Claude has not been observed performing it.

The ten most exposed occupations under the new measure are Computer Programmers (75% coverage), Customer Service Representatives, Data Entry Keyers, and others in routine cognitive work. At the bottom, 30% of workers have zero coverage — cooks, motorcycle mechanics, lifeguards, bartenders, dishwashers, and dressing room attendants. The exposed group is 16 percentage points more likely to be female, 11 points more likely to be white, and almost twice as likely to be Asian. They earn 47% more on average and are nearly four times as likely to hold graduate degrees.

This demographic profile is unusual. Past waves of automation — industrial robots, offshoring — hit less-educated, lower-paid workers hardest. AI, at least in its current form, concentrates exposure among higher-paid, more-educated, disproportionately female and Asian workers. That does not mean those workers will lose jobs. It means the shape of disruption, if it comes, will look different from what economists are used to.

The paper focuses on unemployment as its priority outcome. That is the right call: unemployment captures direct economic harm. But the data so far shows no systematic increase. The authors are careful about causality. AI’s effects may be “less like COVID and more like the internet or trade with China” — gradual, confounded by business cycles and policy, hard to isolate in aggregate numbers.

The most provocative finding is the slowdown in hiring of younger workers. The paper calls it “suggestive evidence,” not conclusive. But if AI reduces demand for entry-level work in exposed occupations, the effect would show up first in hiring, not firing. Companies stop posting jobs before they lay people off. A hiring chill for younger workers in exposed occupations would be consistent with firms substituting AI for tasks that junior employees traditionally do.

The BLS projection correlation adds another layer. For every 10 percentage point increase in observed exposure, the BLS’s growth projection drops by 0.6 percentage points. The relationship is slight but statistically significant. Notably, the theoretical measure alone shows no such correlation. The usage data adds signal.

The gap between what AI could do and what it actually does in the labor market is large, and that gap is the most important number in the paper.

What the paper does not address is the quality of work that remains. Even if employment holds steady, the nature of work can change. Workers in exposed occupations may find their roles narrowed to tasks AI cannot handle, their autonomy reduced, their wages stagnant. The paper’s framework is designed to be revisited periodically. That is the right approach. The effects of AI on labor are not a single event. They are a process that will unfold over years.

For AI builders, the paper offers a concrete benchmark. Observed exposure is a measure of real deployment, not theoretical potential. It tracks how fast the red area of actual usage spreads toward the blue area of capability. If that gap narrows quickly, the labor market effects will follow. If it narrows slowly, the doomsday forecasts will look as overblown as the offshoring predictions of the 2000s.

The paper ends on a note of humility. The track record of past economic forecasts gives reason for caution. A prominent measure of job offshorability identified a quarter of US jobs as vulnerable. A decade later, most of those jobs maintained healthy employment growth. AI may be different. But the data so far does not show the disruption that headlines predict.

What it shows is a labor market that has absorbed AI without measurable damage to aggregate employment, but with early signs that the entry points for younger workers may be narrowing. That is not a crisis. It is a signal worth watching.