Who Is AI Actually Replacing? The Gap Between Theory and Reality

The Gist: A new Anthropic report (5 March 2026) introduces a more honest way to measure AI’s impact on jobs. It distinguishes between what AI could do and what it is actually doing. The gap between those two things matters enormously, and the early employment data are more ambiguous than the headlines suggest.

There is a particular kind of anxiety that follows every major technology. We worry in advance. We map the territory of disruption before disruption has properly arrived. The history of automation research is full of warnings that proved, with hindsight, to be systematically wrong about timing, scale, or the direction of harm.

This does not mean the warnings are always wrong. It means we need better tools for telling the difference.

A new working paper from Anthropic researchers Maxim Massenkoff & Peter McCrory (2026) attempts exactly that. It introduces a measure called observed exposure, and it is worth understanding why that framing matters.

The Problem with Theoretical Exposure

Most of the influential work on AI and jobs has used what we might call theoretical exposure measures. The best-known is Eloundou et al. (2023), published in Science, which scored occupational tasks on whether a large language model could in principle double the speed of completion. Their findings were striking: around 80% of the US workforce could have at least 10% of their work tasks affected by LLMs (Eloundou et al.). These are important baselines.

But theoretical capability is not the same as actual use. A surgeon could theoretically perform a procedure faster with a particular instrument. That does not mean the instrument is in their hand.

Massenkoff and McCrory make this distinction concrete. They combine three data sources: the O*NET occupational task database, the task-level exposure scores from Eloundou et al., and actual usage data from millions of Claude conversations recorded in the Anthropic Economic Index. The result is a measure that weights tasks by whether they are genuinely being automated or merely augmented in professional settings, and by what fraction of a job those tasks represent.

What the Data Show

The gap between theoretical and observed exposure is significant. Take computer and mathematics occupations. Eloundou et al. estimate that 94% of tasks in this category are theoretically within LLM reach. Actual observed coverage, based on real Claude usage, sits at 33%. That is a substantial shortfall. We are not in the world the theoretical models describe.

This matters for how we interpret both fear and reassurance. If we only use theoretical exposure, we may dramatically overstate how much disruption is already under way. If we ignore theoretical exposure entirely, we may miss where disruption is heading.

The ten occupations currently most exposed on the observed measure include Computer Programmers (75% coverage), Customer Service Representatives, and Data Entry Keyers (67%). At the other end, 30% of workers show zero coverage. Their jobs simply do not appear in AI usage data often enough to register. This group includes cooks, motorcycle mechanics, lifeguards, and bartenders.

Does Higher Exposure Predict Worse Job Prospects?

The paper checks this against US Bureau of Labor Statistics employment projections for 2024 to 2034. It finds a modest but real relationship: every 10 percentage point increase in observed exposure corresponds to a 0.6 percentage point drop in projected employment growth. This is a validation that the measure is tracking something meaningful. It is not a large effect, and it rests on projections rather than outcomes, but it gives the framework some empirical grounding.

Interestingly, the theoretical Eloundou et al. measure alone shows no such correlation with BLS projections. Observed exposure appears to capture something the theoretical measure does not.

Who Is Most Exposed?

The demographic profile of highly exposed workers is not what many people expect. Comparing the top quartile of observed exposure to the 30% of workers with zero exposure in late 2022, the exposed group is 16 percentage points more likely to be female, almost twice as likely to be Asian, and earns 47% more on average. Workers with graduate degrees make up 4.5% of the unexposed group but 17.4% of the most exposed.

This is a finding that should give us pause. The standard narrative of automation hitting low-wage, low-skill workers first does not hold here. The workers most in the path of observed AI adoption are better educated and better paid. Whether this means they are better placed to adapt, or that the consequences will be concentrated in a group that has more to lose, remains an open question.

Why This Framework Is Valuable

What Massenkoff and McCrory have built is not a prediction. It is a monitoring framework. They are explicit about this. By establishing the approach now, before large-scale disruption has clearly emerged, they create the conditions for cleaner causal inference later. As AI usage grows and employment data accumulate, the framework will sharpen.

That is good science. Daron Acemoglu & Pascual Restrepo’s work on industrial robots (2020) is instructive here. Even studying a technology whose rollout was already largely complete, economists reached opposing conclusions about the scale of job displacement. The debate over the China trade shock continues to this day (Autor et al., 2013). Post-hoc analysis of major economic disruptions is genuinely difficult.

The point is not that AI will cause no harm. The point is that we need frameworks we can trust to tell us when it is happening, how much, and to whom. That is what this paper starts to build.

Share

Thanks for reading Data in Motion.

Subscribe for free to receive new posts and support my work.