What Happened
Anthropic published a research paper on March 6, 2026 titled "Labor market impacts of AI: A new measure and early evidence," authored by Maxim Massenkoff and Peter McCrory. The paper introduces a new metric called "observed exposure" that compares what AI can theoretically do against what people are actually using it for, based on real-world Claude interaction data.
The headline numbers are striking. Computer and math workers face 94% theoretical AI capability exposure, but only 33% is showing up in actual usage today. Office and administrative roles sit at 90% theoretical capability. Lawyers, financial analysts, software developers, customer service reps, and data entry workers all rank high on the exposure list.
About 30% of workers have essentially zero AI exposure - cooks, mechanics, bartenders, dishwashers, and others whose jobs require physical presence.
The demographic breakdown is worth noting: the most AI-exposed workers are 16 percentage points more likely to be female, earn 47% more on average, and are nearly four times more likely to hold a graduate degree than the least exposed group.
On early labor market signals, the researchers found a 14% drop in the job-finding rate in AI-exposed occupations compared to 2022 (post-ChatGPT launch), though they note the findings are "barely statistically significant." No systematic unemployment increase has appeared yet.
Why It Matters
This paper is significant because it comes from Anthropic itself - a company that directly profits from AI adoption. When the maker of Claude publishes data showing the gap between AI capability and actual workplace adoption, it tells us two things: the current disruption is real but early, and the ceiling is much higher than where we are now.
The "Great Recession for white-collar workers" framing puts concrete numbers on the risk. During the 2007-2009 financial crisis, U.S. unemployment doubled from 5% to 10%. The researchers suggest a comparable doubling in the top-quartile AI-exposed occupations would push unemployment from 3% to 6% in those fields. That is not hypothetical catastrophe - it is a plausible scenario backed by trend data.
The February 2026 jobs report already showed employers shedding 92,000 jobs, with unemployment rising to 4.4%. Separate research found a 16% employment drop in AI-exposed jobs specifically among workers aged 22 to 25. Entry-level knowledge workers are getting hit first.
Our Take
The gap between theoretical capability (94%) and actual adoption (33%) in computer and math work is the most important number in this paper. It means we are roughly a third of the way through the disruption curve in the most exposed fields. The barriers - legal constraints, model limitations, need for human review - are real but temporary. Every model update chips away at them.
What is genuinely useful here is the "observed exposure" metric. Previous AI job studies relied on expert opinion about what AI could theoretically do. Anthropic used actual Claude usage data to measure what people are actually doing with AI right now. That is a better signal.
The demographic skew deserves attention. Higher-paid, more-educated, more-female workforces face the most exposure. This inverts the usual automation narrative where factory and service workers bear the brunt.
For anyone working in AI-exposed fields, the practical takeaway is clear: the tools are not replacing jobs overnight, but the job-finding rate is already declining in these categories. Learning to work with AI tools effectively is not optional career development - it is baseline job security. The 14% drop in job-finding rates is a leading indicator, not a lagging one.