Anthropic Measures AI's Real Impact on Jobs: Small, For Now

Anthropic has published a study that attempts to answer one of the most repeated questions of recent years: is artificial intelligence destroying jobs? Its answer, based on data through early 2026, is nuanced. AI still operates well below its theoretical capacity, there's no detectable systematic rise in unemployment among the most exposed workers, and the only worrying sign points to a slowdown in hiring young workers in occupations where Claude is already widely used.

The paper, titled Labor market impacts of AI: A new measure and early evidence, is neither triumphalist nor doom-laden. It's a methodological attempt to lay the groundwork for properly measuring a phenomenon that, as the authors themselves admit, economics has historically measured poorly.

Why previous predictions failed

The study opens with an unusual dose of humility. Anthropic notes that previous attempts to anticipate labor disruptions tend to age badly.

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 study's contribution is a new metric it calls observed exposure. Most previous analyses measure theoretical exposure: what tasks a language model could do if asked. Anthropic adds a crucial second layer: what tasks AI is actually performing in professional contexts.

To build it, the researchers combine three sources:

• O*NET, the official U.S. database that breaks down roughly 800 occupations into specific tasks.
• Claude usage data, collected via the Anthropic Economic Index, showing what people actually use the model for.
• Estimates from Eloundou et al. (2023), which score each task based on whether a language model can at least double the speed at which it's performed.

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 interesting part is the gap between what's possible and what's real. Eloundou et al., for instance, mark the task of "authorizing prescription refills and notifying pharmacies" as fully automatable (β=1). Anthropic admits it has never seen Claude actually do this. It's possible in theory; it doesn't happen in practice, held back by legal barriers, software requirements, or human verification steps.

Even so, theory and practice largely align: 97% of observed Claude usage falls within tasks that Eloundou et al. consider theoretically feasible.

AI is far from its ceiling

The most striking finding is how much room is left. In the Computer and Mathematical occupation category, theoretical capacity covers 94% of tasks, but actual Claude usage reaches only 33%. In Office and Administrative Support, theory reaches 90%, while real usage falls far short.

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 of occupations with the highest observed exposure confirms what the use of these tools already suggested:

• 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

One finding that breaks stereotypes is the profile of the most exposed worker. According to the study, they tend to be older, female, more educated, and better paid. This isn't the precarious profile usually associated with classic factory automation, but rather that of skilled office work.

This makes sense: language models excel precisely at cognitive, document-based, and writing tasks—the kind that occupy skilled professionals.

And employment? Holding up, for now

Here's the part likely to spark the most debate. Anthropic cross-referenced its exposure measure with U.S. Bureau of Labor Statistics (BLS) employment projections for 2024-2034.

The result: occupations with higher observed exposure show somewhat weaker growth forecasts. For every 10-percentage-point increase in coverage, projected employment growth drops by 0.6 points. It's a real relationship, but a small one. And, tellingly, that correlation disappears if only the theoretical measure from Eloundou et al. is used: it's the combination with real usage that provides predictive power.

Even more important: since late 2022—when ChatGPT popularized these tools—Anthropic finds no systematic rise in unemployment among the most exposed workers. Unemployment hasn't spiked in the professions AI touches most.

The only warning sign is subtle but significant: there are indications that hiring of young workers has slowed in exposed occupations. In other words, the effect wouldn't (yet) show up as layoffs of current employees, but as a closed door for those trying to enter. If confirmed, this would be a quiet but profound shift, since it affects how new generations are trained and brought into the labor market.

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 is that the authors themselves frame it as a starting point, not a closed conclusion. Their stated goal is to establish a method for measuring impact before it becomes obvious, and to revisit it periodically. They acknowledge that their framework doesn't capture every channel through which AI can transform work.

Even with those caveats, the message is useful precisely because it pushes back against both dominant narratives. Against those announcing an imminent labor apocalypse, the data show that AI still operates well below its potential and that aggregate employment is holding steady. Against those dismissing any effect at all, there's that signal in youth hiring that deserves close monitoring.

The real value will come with future updates. As the paper itself acknowledges, this methodology shines when effects are ambiguous: it can detect vulnerable jobs before displacement becomes visible in the numbers. If the gap between what AI can do and what it actually does begins to close—the red area filling in the blue area, as shown in the study's chart—this will be the thermometer to watch.