Anthropic published a labor market study on March 5, 2026 that does something most AI job-threat research doesn’t: it uses actual Claude usage data to measure what AI is replacing today, not what it could theoretically replace.
The report introduces a metric called “observed exposure” — a way of distinguishing between jobs AI is capable of automating and jobs it’s actually being used to automate. That distinction turns out to matter a lot.
The gap between theoretical and real is wide. Computer and math occupations score 94% exposed under theoretical models — but only 33% when measured by actual Claude usage patterns. Theory says AI can do almost everything in that category. Reality says workers and companies are using it for about a third of it, and the rest is still human.
What Claude is actually being used for most: writing and updating software code, handling customer service inquiries, and processing data entry tasks. According to Axios, Anthropic economists found “limited evidence” that AI has caused a measurable spike in overall unemployment. But there is an early signal, and it’s pointed at a specific group.
Anthropic: The Jobs Most in the Crosshairs Right Now
Computer programmers sit at the top with 74.5% observed exposure — meaning three-quarters of their documented tasks are already being handled or assisted by AI in measurable ways. Customer service representatives follow at 70.1%, data entry keyers at 67.1%, market research analysts at 64.8%. These aren’t future projections. This is what Claude’s traffic logs show is happening right now.
The workers in the most exposed jobs earn 47% more on average than those in unexposed roles. They’re more likely to hold graduate degrees, more likely to be white or Asian, and more likely to be female. The automation wave, at least at this stage, is moving through higher-earning, white-collar knowledge work first — not the trades, not manual labor, not the jobs that have always been the assumed target.
Workers 30% of the labor force remain entirely untouched. As Business Insider noted, cooks, bartenders, dishwashers, and lifeguards have zero observed AI coverage. The robot apocalypse, if it comes, is starting in the office, not the kitchen.
The Young Worker Signal Nobody Should Ignore
Overall unemployment in AI-exposed occupations hasn’t moved since ChatGPT launched in late 2022. That’s the headline finding, and it’s reassuring. But within the data there’s a quieter number: among workers aged 22 to 25, the job-finding rate in highly exposed occupations has dropped 14% relative to 2022 levels. Workers over 25 show no equivalent trend.
That means the AI impact, so far, isn’t showing up as mass layoffs — it’s showing up as reduced hiring at the entry level. Companies appear to be doing more with existing staff rather than replacing senior employees, while quietly shrinking the pipeline for new graduates. Harvard Business Review found similar dynamics — real job loss and slowed hiring driven by AI expectations, even before performance justifies it.
Anthropic’s own CEO Dario Amodei told The Atlantic in May 2025 that AI could push unemployment to 10–20% within years. His own company’s data, published ten months later, shows the wave hasn’t broken yet. But the shore is getting closer.