A vast, dimly lit industrial control room filled with rows of glowing blue computer screens showing abstract AI neural network visualizations. In the foreground

Workforce Risk · March 2026 · Anthropic Research

The Automation Gap

AI can theoretically perform 94% of tasks in Computer & Math occupations. In practice, Claude covers just 33%. That chasm — between what AI could do and what it actually does at work — is the story no existing measure has been able to tell. Until now.

Observed Exposure at a Glance

Data Sources: O*NET (~800 occupations) · Claude usage (Anthropic Economic Index) · Eloundou et al. (2023) β scores · BLS employment projections · Current Population Survey
β Scale: 1 = LLM doubles task speed alone · 0.5 = requires additional tools · 0 = not feasible
Coverage Weights: Automated use → 1× · Augmentative use → 0.5×
Theoretical exposure alone: Zero correlation with BLS 2024–2034 job-growth projections — only Observed Exposure correlates
Illustrative gap: 'Authorize drug refills' rated β=1 by Eloundou et al. — yet never observed in Claude traffic

68%

of observed Claude usage falls on β=1 tasks — fully feasible for an LLM alone. Tasks rated β=0 account for just 3%.

Computer & Math: 94% theoretically feasible — yet only 33% observed coverage

75%

Computer Programmers — highest task coverage of any occupation

30% of all workers have zero observed exposure: Cooks, Bartenders, Lifeguards, Dishwashers & more

Conventional wisdom frames automation as a threat to low-wage, low-skill workers. The demographic data tell a different story. Using Current Population Survey data from August–October 2022 — before ChatGPT's release — workers in the top quartile of observed exposure look nothing like the assumed victim of technological displacement.

The most-exposed workers earn 47% more on average than the zero-exposure group. They are 16 percentage points more likely to be female, 11 percentage points more likely to be white, and nearly twice as likely to be Asian. Graduate degree holders make up just 4.5% of the zero-exposure group — but 17.4% of the most-exposed group, a nearly fourfold difference.

The workers most at risk are older, more educated, higher-paid, and more likely to be female — upending conventional assumptions about automation and low-wage labor.

Most-Exposed vs. Zero-Exposure Workers (CPS, Aug–Oct 2022)

Avg. wages: +47% higher
More likely female: +16 pp
More likely white: +11 pp
More likely Asian: ~2× more likely
Graduate degree holders: 17.4% vs. 4.5%

Measured unemployment increase for high-exposure workers since late 2022

There is, however, a subtler early warning: hiring of younger workers has slowed in occupations with high observed exposure post-ChatGPT. Not a collapse, but a detectable softening that wasn't there before.

The authors caution against reading the silence as safety. AI's labor-market impact may resemble the internet or the China trade shock — effects that were real but invisible in unemployment figures for years, obscured by the business cycle and trade policy noise. By establishing rigorous measurement now, before meaningful effects have fully emerged, the authors aim to build an early warning system capable of identifying disruption before post-hoc analysis becomes the only option.

–0.6pp

projected employment growth drop per 10pp rise in Observed Exposure (BLS 2024–2034)

Theoretical exposure alone: r = 0, no correlation

Three Takeaways for Workforce Planners

1 · Use Observed, Not Theoretical, Exposure: Tasks rated β=1 account for 68% of Claude usage — yet even in Computer & Math, observed coverage is only 33% of a 94% theoretical ceiling. Workforce plans built on theoretical ceilings misallocate resources.
2 · Reskilling Must Follow Exposure, Not Wage Tier: The most-exposed workers earn 47% more, are 16 pp more likely to be female, and are nearly 4× more likely to hold a graduate degree (17.4% vs. 4.5%). Risk is concentrated at the top of the credential ladder, not the bottom.
3 · Watch Talent Pipelines, Not Just Headcount: No unemployment spike has emerged yet — but hiring of younger workers has already slowed in exposed occupations. The leading edge of disruption will likely appear in graduate hiring and early-career intake first.

This framework is most useful when the effects are ambiguous — and could help identify the most vulnerable jobs before displacement is visible.
— Massenkoff & McCrory, Anthropic (2026)