Summary: Anthropic’s March 5, 2026 study introduced observed exposure, measuring tasks people actually use AI to perform rather than tasks AI could theoretically perform. It found 94% theoretical feasibility but 33% observed use in computer and math tasks, and no systematic climb in unemployment among highly exposed workers. Job-finding by 22–25-year-olds in exposed occupations fell about 14%, a barely significant result, while the Claude-only data does not prove displacement.
Anthropic released a study titled “Labor market impacts of AI: A new measure and early evidence” on March 5, 2026, written by Maxim Massenkoff and Peter McCrory. Its central contribution is a new metric, “observed exposure”: rather than asking what AI could do, it tallies what people actually use AI to do. It lands in a 2026 where, as the opening outlook for 2026 shows, everyone expected agents to take over whole workflows — which makes this a rare labor-market measurement grounded in real usage data, not capability demos.
The paper’s core result fits in one line: between theoretical capability and actual use sits a gap of roughly three to one.
How “Observed Exposure” Is Measured
Earlier work on AI and jobs rarely got past “theoretical exposure”: list an occupation’s tasks, ask whether a model could handle each one, and score it. Critics have long flagged that this inflates the estimated impact — a model being capable of a task is not the same as anyone using it that way. Anthropic instead layers three datasets: task breakdowns for roughly 800 US occupations drawn from the O*NET database, the record of real Claude conversations accumulated by the Anthropic Economic Index, and the task-level exposure scores of Eloundou et al. (2023).
Intent shapes the weights: a task the AI carries out end to end counts at full weight, augmentative assistance counts at half, and on-the-job usage outweighs personal use. Coverage at the task level is then rolled up to occupations, each task weighted by the time spent on it. The result is an index of automation that has already happened, rather than automation that might.
The Threefold Gap Between Feasible and Used
The paper’s strongest contrast involves computer and math occupations: LLM capability theoretically covers 94% of their tasks, yet real usage reaches only 33% — a gap close to threefold. Across the board, 97% of Claude usage lands inside theoretically feasible task categories, but actual coverage swings widely from one occupation to the next.
Treat that gap as the labor market’s pressure gauge for the years ahead. Capability is already out in front of adoption; when the impact lands depends on how fast the gap closes, not on the next model’s capability. For product teams, the same numbers size up an opening: two-thirds of the computer-related tasks that could theoretically be automated still have no product catching them.
Who Tops the List, and the Three in Ten With No Exposure
The leaderboard was picked up by both CBS News and Investopedia: computer programmers lead at roughly 75% task coverage, customer service representatives follow at about 70%, and data entry keyers register 67%. At the other end, roughly 30% of US workers register zero exposure — cooks, bartenders, dishwashers, lifeguards, motorcycle mechanics, dressing room attendants. AI’s footprint sits almost entirely in work that can happen through a screen.
The demographic profile of the exposed group goes against a few instincts: highly exposed workers are likelier to be female by 16 percentage points, out-earn the rest by 47%, and hold graduate degrees at a rate of 17.4% against 4.5% for the unexposed. AI moves through high-education, higher-paid white-collar occupations first, not the other way around. Two more demographic figures: exposed workers are nearly twice as likely to be Asian and 11 percentage points more likely to be white. As for method, the authors apply a difference-in-differences design to the Current Population Survey, setting workers in the top exposure quartile against the roughly 30% with none.
The Data So Far: Flat Unemployment, a Narrower Door for Young Workers
Bad news first — for now, there is none. In the period since ChatGPT arrived in late 2022, unemployment among highly exposed workers has shown no systematic climb — the study’s design would have caught a differential rise of about 1 percentage point, and it caught nothing.
Two directional signals do stand out, though. The first concerns young workers: once ChatGPT appeared, the rate at which 22–25-year-olds found jobs in exposed occupations dropped by about 14% — an effect the authors call “just barely statistically significant” — and workers over 25 show no such slide. The direction matches Brynjolfsson et al. (2025), whose analysis of ADP payroll data put the employment decline for ages 22–25 in exposed occupations at 6–16%. Put another way, the evidence reads less like incumbents being replaced and more like the entrance narrowing: those trying to get in are shut out, while the people already seated are untouched. The second signal sits in the BLS ten-year projections: each additional 10 percentage points of observed exposure lines up with projected 2024–2034 job growth that is 0.6 points lower — whereas the identical regression run on conventional theoretical-exposure scores turns up nothing at all.
Three Caveats Before Citing It
This study comes with three limitations worth remembering. First, the usage data is Claude-only: white-collar professionals and developers are naturally heavy users, so what it captures is automation among Claude users, not automation by AI as a whole. Second, exposure and displacement are different things — within the 33% coverage, augmentative uses make up the bulk and fully automated uses a smaller slice. Third, usage data inevitably trails frontier capability: the study records what has already been deployed while capability moves forward every quarter, so real exposure can only be understated, never overstated.
Even with those caveats, this is the first framework to separate capability from usage and hook directly into official labor statistics, handing policymakers and product teams a dashboard worth revisiting. What deserves watching is not the unemployment rate but the pace at which 33% converges on 94%, and whether the job-finding curve for 22-to-25-year-olds keeps bending downward.