Occupations With the Highest Observed AI Exposure: Computer Programmers Lead at 74.5%

What the score counts

The score comes from Anthropic's March 5, 2026 paper "Labor market impacts of AI: A new measure and early evidence", which combines three sources: the O*NET database of tasks for around 800 US occupations, usage data from the Anthropic Economic Index, and the task-level estimates of Eloundou et al. (2023) of whether a language model could theoretically make a task at least twice as fast.² A task counts as covered when it is theoretically feasible and has seen enough work-related usage in Claude traffic; automated implementations count in full and augmentative use at half weight; and task scores are averaged to the occupation using the fraction of time spent on each task.² The usage behind this release covers August and November 2025 data from the Economic Index.² The authors say the construction involves "judgment calls at every step", and report that rankings of occupations are highly correlated across many alternative choices.² This is the task-based approach to AI exposure, using O*NET tasks; for how MetaRoles compares such measures, see how we measure how much of a role AI does.

Anthropic published the paper and the data, and the usage is from its own product, so this is a lab-run measure of one platform. MetaRoles does not average it with other exposure indices or rank it against them; where indices disagree we show them side by side.

Why 411 occupations score zero

In Anthropic's file, 411 of the 756 occupations score exactly 0 (our count).¹ Anthropic explains that 30% of workers are in occupations with zero coverage because their tasks "appeared too infrequently in our data to meet the minimum threshold", and names Cooks, Motorcycle Mechanics, Lifeguards, Bartenders, Dishwashers and Dressing Room Attendants as examples.² A zero therefore means "not seen often enough in this dataset", not that AI cannot do the work. The file scores Taxi Drivers at 0¹, for example, although driverless taxis carry paying passengers in several cities: the measure only sees language-model usage on Claude, not robots or autonomous vehicles. See robotaxis and the taxi driver's job for graded cases.

Position is 1 plus the number of occupations with a higher score, so every occupation scoring zero shares the same position. Medicine rows use the closest occupations to those in the linked posts, not every physician type. A score is not a measure of what AI does in the role: the posts linked here grade the actual cases.¹

What it says about jobs

Anthropic itself calls observed exposure "a new measure of AI displacement risk", and its paper is a test of whether exposure shows up in employment.² What it reports so far: no systematic increase in unemployment for highly exposed workers since late 2022; suggestive evidence that hiring of younger workers has slowed in exposed occupations; and, from a regression weighted by employment, that the Bureau of Labor Statistics' 2024 to 2034 growth projection falls by 0.6 percentage points for every 10-point rise in coverage, a relationship the authors call slight.² Those are Anthropic's findings under its stated methods and dates, not established effects, and a high score is not a forecast that jobs will disappear: the paper itself notes that earlier attempts to identify vulnerable jobs, such as offshorability, have a mixed record.²