Anthropic just released a major research report: Labor Market Impacts of AI: A New Measure and Early Evidence.
The moment it dropped, a single chart went viral on Twitter. Not because of how capable AI is — but because of who it's targeting. The workers most at risk from AI are not who you'd expect.
Why Past Predictions Keep Getting It Wrong
Previous research has shared a fundamental flaw: it measured what AI can theoretically do, not what AI is actually doing.
For example, in Eloundou et al.'s widely-cited framework, "authorizing drug refills and transmitting prescriptions to pharmacies" is rated as fully automatable (β=1). Yet in the real world, nobody is actually using Claude for that task.
There's a vast gap between theoretical capability and actual deployment — legal constraints, technical requirements, human verification steps, specialized software dependencies. These are real barriers.
A New Metric: "Observed Exposure"
Anthropic introduces a new measure: "Observed Exposure" — combining two dimensions: whether a task is theoretically feasible for an LLM, and whether that task is actually being handled by AI in Anthropic's usage data. Fully automated uses get full weight; augmentation uses get half weight.
Here's how Claude usage breaks down by exposure level:
Figure 1: 97% of observed tasks fall in the theoretically feasible category — but actual coverage is far lower
The gap between theory and reality by occupational category:
Figure 2: Blue = theoretical coverage, Red = observed coverage — the gap is enormous
Take Computer & Math occupations: 94% of tasks are theoretically automatable. But Claude currently covers only 33% in practice. AI is nowhere near its theoretical ceiling.
The Ten Most Exposed Occupations
Ranked by Observed Exposure:
Figure 3: Computer programmers (74.5%), customer service reps (70.1%), and data entry workers (67.1%) lead the list
Worth noting: 30% of U.S. workers have zero exposure — cooks, motorcycle mechanics, lifeguards, bartenders, dishwashers. Their tasks simply don't appear meaningfully in AI usage data.
Higher Exposure, Lower Employment Growth
Observed Exposure correlates negatively with BLS projected job growth for 2024–2034 — every 10 percentage point increase in exposure is associated with a 0.6 percentage point reduction in projected growth:
Figure 4: Slope = -6.07, R² = 0.027. Crucially, this relationship disappears when using theoretical exposure only.
Key insight: This correlation only appears when using observed exposure data. Theory-only metrics miss it entirely — proving that real usage data is irreplaceable.
The Viral Chart: Who Are the At-Risk Workers?
This is the chart that spread across Twitter — comparing workers in the top quartile of AI exposure vs. workers with zero exposure:
Figure 5: The "striking image" — the profile of high-exposure workers defies conventional wisdom
The findings challenge everything we assumed:
Gender: High-exposure workers are 54.4% female, vs. 38.8% in the zero-exposure group. Women face greater AI risk.
Education: High-exposure workers are nearly 3x more likely to have a bachelor's degree (37.1% vs. 13.3%), and nearly 4x more likely to have a graduate degree (17.4% vs. 4.5%). AI is threatening the most educated workers first.
Wages: High-exposure workers earn an average of $32.69/hour, vs. $22.23 in the zero-exposure group — 47% higher. Higher earners are more at risk, not less.
Race/Ethnicity: Asian workers are nearly twice as represented in the high-exposure group (9.1% vs. 4.7%). Hispanic workers are more concentrated in the zero-exposure group (24.8% vs. 13.8%).
This is what makes the chart so striking: it shatters the narrative that AI will primarily displace low-skill, low-wage workers. The opposite is true — AI is coming first for the highly educated, well-compensated, predominantly female and Asian white-collar workforce.
No Systematic Job Losses Yet — But Early Warning Signs
The report tracks unemployment trends from 2016 to 2025 for high-exposure vs. zero-exposure workers:
Figure 6: Since ChatGPT's launch in late 2022, no systematic rise in unemployment for high-exposure workers
The main finding: Since ChatGPT launched in late 2022, high-exposure workers have not experienced a systematic increase in unemployment. At least for now, AI hasn't left a visible scar on aggregate employment data.
But there's a warning sign worth watching — new hiring of workers aged 22–25 in high-exposure occupations is slowing down:
Figure 7: Young workers facing slower hiring — AI's impact may be more subtle than mass layoffs
AI's impact may not come through mass layoffs. Instead, it may manifest quietly: industries are simply hiring fewer new workers.
Why This Report Matters
This report offers a sharper, more honest lens: don't ask "what can AI theoretically do?" — ask "what is AI actually doing?"
History keeps proving that theoretical automation risk overstates actual impact. Researchers predicted offshoring would devastate U.S. jobs 25 years ago. Industrial robots were supposed to trigger mass unemployment a decade ago. Both turned out more complicated.
AI's disruption is coming — but its path, pace, and targets may all be different from what we imagine. This report is one of the best early warning instruments we have.
Source: Anthropic, "Labor market impacts of AI: A new measure and early evidence," March 5, 2026 Authors: Maxim Massenkoff and Peter McCrory