IMPLICATOR
Anthropic published a new measure of AI's labor market effects on Thursday, combining theoretical LLM capability with real-world Claude usage data to track which occupations face the most displacement risk. The framework, called "observed exposure," found that computer programmers top the list at 75% task coverage, followed by customer service representatives and data entry workers. No measurable increase in unemployment showed up for exposed workers since ChatGPT launched in late 2022.
Anthropic economists Maxim Massenkoff and Peter McCrory wrote the paper. It lands in the middle of a nervous corporate season. CEOs from Jack Dorsey to Mark Zuckerberg have blamed AI for mass layoffs while posting record profits. Actual evidence of AI-driven job losses remains thin.
The Breakdown
- Anthropic's "observed exposure" measure combines LLM capability with real Claude usage data across 800 occupations
- Computer programmers top the list at 75% task coverage; 30% of workers have zero AI exposure
- No measurable unemployment increase for exposed workers since ChatGPT's late 2022 launch
- Entry-level hiring for workers aged 22-25 in exposed jobs dropped 14%, echoing Stanford findings
What the measure actually tracks
Most AI exposure indexes work from theory. Researchers look at job descriptions, estimate which tasks a language model could speed up, and assign a score. Anthropic's contribution is bolting actual usage data onto that framework.
Massenkoff and McCrory pulled anonymized Claude conversation data from Anthropic's Economic Index reports and matched it against roughly 800 occupations in the Department of Labor's O*NET database. Tasks got weighted more heavily when they showed up in automated, work-related contexts rather than casual augmentation. A customer service script running through an API counts for more than a marketing manager brainstorming taglines.
What jumped out was the gap between theoretical capability and actual usage. Take Computer and Math occupations. Existing measures from Eloundou et al. suggest LLMs could handle 94% of their tasks. Claude currently covers 33%. Office and Admin jobs show a similar spread. The blue area on Anthropic's charts, representing theoretical reach, dwarfs the red sliver of actual deployment.
Thirty percent of American workers registered zero coverage on this measure. Their jobs, cooks, motorcycle mechanics, lifeguards, bartenders, never showed up in Claude traffic at volumes high enough to count. That tracks with what you'd expect. Physical work stays physical.
The unemployment signal that isn't there
Massenkoff and McCrory compared unemployment trends for workers in the top quartile of observed exposure against those with zero exposure, using Current Population Survey data going back to 2016. The result was flat. No statistically significant divergence since November 2022.
Yale's Budget Lab reached the same conclusion last October, finding that 33 months of labor data showed no meaningful acceleration in occupational churn since ChatGPT's release. Anthropic's framework adds a different lens, one built on actual product usage rather than theoretical task matching, and arrives at the same place. The aggregate labor market looks stable.
Stable doesn't mean static. Look closer.
Stay ahead of the curve
Strategic AI news from San Francisco. No hype, no "AI will change everything" throat clearing. Just what moved, who won, and why it matters. Daily at 6am PST.
No spam. Unsubscribe anytime.
Young workers feel it first
Workers aged 22 to 25 showed a different pattern. Anthropic tracked monthly job-finding rates, measuring how often young workers started new jobs in high-exposure versus low-exposure occupations. The two lines ran roughly parallel until 2024. Then they split. Entry into exposed occupations dropped by about half a percentage point per month, a 14% decline relative to the pre-ChatGPT baseline. "Just barely statistically significant," the authors noted. Not nothing. Not proof either. Workers older than 25 showed nothing comparable.
If you graduated last spring with a computer science degree and the entry-level market felt tighter than your professors promised, this data offers one possible explanation.
Brynjolfsson and colleagues at Stanford's Digital Economy Lab documented a 6 to 16% employment decline for young workers in AI-exposed occupations using ADP payroll data. Same signal, different dataset. Anthropic's paper cites Brynjolfsson directly, but cautions that slowed hiring doesn't necessarily mean unemployment. The CPS data can't tell whether those young workers are staying at existing jobs, switching fields, or going back to school.
Who faces the most risk
The demographic profile of exposed workers cuts against the layoff headlines. People in the top quartile of Anthropic's measure take home 47% more than those with zero exposure. More likely to be female, by 16 percentage points. More likely to be white, by 11. Almost four times as likely to hold a graduate degree.
Not the portrait of a vulnerable workforce. The workers most exposed to AI displacement are also the ones best positioned to absorb it. Brookings research from January found that 26.5 million of the 37.1 million most-exposed workers had above-median adaptive capacity, meaning savings, transferable skills, and access to dense labor markets. The 6.1 million who lacked that cushion clustered in clerical and administrative roles, the kind of offices where a single receptionist handles the front desk and the filing system nobody else understands. Eighty-six percent of them were women.
What the gap between blue and red means
Massenkoff told Axios that the "China shock" of the early 2000s took years to show up clearly in employment statistics. AI could follow the same pattern. The paper's most striking visual is that enormous uncovered area between what LLMs could theoretically do and what they're actually doing right now.
Anthropic is betting that gap will close. The company designed this framework to detect displacement before it becomes obvious in aggregate data, a cautious hedge from a company whose own product sits at the center of the disruption it's measuring. Whether the red area creeps forward slowly or lurches, the tracker will show it.
For now, the data says what multiple independent research teams keep saying. AI is not yet destroying jobs at scale. The word "yet" carries the weight.
Frequently Asked Questions
What is 'observed exposure' and how does it differ from other AI job measures?
Most AI exposure indexes estimate which tasks a language model could theoretically speed up. Anthropic's measure adds real-world Claude usage data, weighting automated and work-related tasks more heavily than casual augmentation. It tracks what AI is actually doing, not just what it could do.
Which jobs are most exposed to AI displacement?
Computer programmers top the list at 75% task coverage, followed by customer service representatives and data entry keyers at 67%. Thirty percent of workers, in jobs like cooking, bartending, and motorcycle repair, registered zero coverage.
Has AI caused measurable job losses so far?
No. Unemployment rates for workers in highly exposed occupations have not diverged from those in unexposed jobs since ChatGPT launched in late 2022. Yale's Budget Lab reached the same conclusion using different methodology.
Why are young workers showing different results than older workers?
Workers aged 22 to 25 saw a 14% drop in job-finding rates for exposed occupations after 2024. The effect was barely statistically significant. Anthropic cautions this could reflect young workers switching fields or returning to school, not just losing jobs.
Who are the 6.1 million workers with high exposure but low adaptive capacity?
Brookings research found these workers cluster in clerical and administrative roles with limited savings, narrow skill sets, and fewer local job alternatives. Eighty-six percent are women. They're concentrated in college towns and state capitals in the Mountain West and Midwest.
Maria Garcia
Tech culture and generative AI reporter covering the intersection of AI with digital culture, consumer behavior, and content creation platforms. Focusing on technology's beneficiaries and those left behind by AI adoption. Based in California.
