Anthropic's New "Observed Exposure" Metric Reveals Which Jobs AI Is Actually Replacing
A groundbreaking research paper from Anthropic introduces a new way to measure AI's real-world impact on jobs — not what AI could do theoretically, but what it's actually doing. The results are both reassuring and alarming.
What Is "Observed Exposure"?
Previous attempts to measure AI's impact on jobs relied on theoretical assessments — asking whether an AI could perform a given task. Anthropic's new metric, "observed exposure," goes further by combining three data sources:
- O*NET database — 800+ occupations and their constituent tasks
- Anthropic Economic Index — real-world Claude usage data from millions of conversations
- Eloundou et al. (2023) — theoretical LLM capability ratings per task
The key innovation: observed exposure weights automated (rather than augmentative) and work-related uses more heavily. A task gets full weight if it's being fully automated via API, but only half weight if humans are using AI as an assistant.
Key Insight
AI is far from reaching its theoretical capability. In the Computer & Math category, LLMs could theoretically handle 94% of tasks, but actual coverage is just 33%. The gap between what AI can do and what it's actually doing remains enormous.
Top 10 Most Exposed Occupations
| # | Occupation | Observed Exposure | Coverage |
|---|---|---|---|
| 1 | Computer Programmers | 75% | |
| 2 | Customer Service Representatives | 70% | |
| 3 | Data Entry Keyers | 67% | |
| 4 | Financial Analysts | 55% | |
| 5 | Technical Writers | 52% | |
| 6 | Market Research Analysts | 48% | |
| 7 | Editors | 45% | |
| 8 | Interpreters & Translators | 43% | |
| 9 | Accountants & Auditors | 41% | |
| 10 | Paralegals & Legal Assistants | 39% |
30% of Workers Have Zero AI Exposure
On the other end of the spectrum, nearly a third of all US workers are in occupations where AI task coverage is effectively zero. These are jobs where tasks appeared too infrequently in Claude's usage data to meet even the minimum threshold.
These roles share a common thread: they require physical presence, manual dexterity, or real-time human interaction that current AI systems simply cannot replicate.
No Unemployment Crisis — Yet
Critical Finding
There is no systematic increase in unemployment for workers in highly AI-exposed occupations since late 2022. The unemployment rate gap between the most and least exposed workers is "small and insignificant."
However, there's an early warning signal that shouldn't be ignored:
While older workers in these roles aren't being fired, young workers (ages 22-25) are finding it harder to get hired into AI-exposed occupations. The researchers describe this as "suggestive evidence that hiring of younger workers has slowed" — a potential leading indicator of broader displacement to come.
Who's Most at Risk? The Demographics
Workers in the top quartile of AI exposure look very different from those with zero exposure:
This inverts the typical automation narrative. Unlike previous waves of automation that primarily affected blue-collar manufacturing workers, AI displacement risk is concentrated among higher-educated, higher-paid, white-collar workers.
BLS Projections Confirm the Trend
The Bureau of Labor Statistics' own employment projections correlate with Anthropic's observed exposure measure:
While the relationship is "slight," it provides independent validation that Anthropic's measure tracks real labor market dynamics. Notably, there is no such correlation when using purely theoretical AI capability measures alone — only when combining theory with actual usage data.
What This Means
The paper's authors are careful to note limitations: their data comes from one AI platform (Claude), and the absence of unemployment effects doesn't mean they won't emerge. They frame this as establishing a baseline:
"By laying this groundwork now, before meaningful effects have emerged, we hope future findings will more reliably identify economic disruption than post-hoc analyses."
The message is clear: the red area is growing. As AI capabilities advance and adoption spreads, the gap between what AI can theoretically do and what it's actually doing will narrow. The question isn't if but when the labor market effects become unmistakable.
Original Research
"Labor market impacts of AI: A new measure and early evidence"
Maxim Massenkoff and Peter McCrory • Anthropic • March 5, 2026