What the Data Actually Says About AI and Jobs Right Now

Anthropic just published one of the more rigorous attempts to measure what AI is actually doing to the labor market — not what it might do, not what CEOs are predicting, but what the employment data shows right now. The findings are more nuanced than either the panic or the reassurance camp wants them to be.

The short version: no measurable spike in unemployment for AI-exposed workers yet. But hiring into those roles is slowing for workers aged 22 to 25, and the occupations most exposed to AI are projected by the Bureau of Labor Statistics to grow less through 2034. The effects are real, early, and uneven — and the researchers are careful to say they're building a framework for catching disruption before it becomes undeniable, not declaring it hasn't arrived.

A New Way of Measuring Exposure That's Worth Understanding

Most prior research on AI job displacement relied on theoretical capability — essentially asking whether an LLM could perform a given task. This paper introduces something more grounded: observed exposure, a measure that combines theoretical feasibility with actual usage data from Claude interactions in professional settings.

The distinction matters. Theoretical capability is broad. Actual usage is considerably narrower. The paper finds that AI covers just 33% of tasks in the Computer and Math occupational category — despite the fact that, in theory, 94% of tasks in that category could be sped up by an LLM. The gap between what AI can do and what it's actually doing at scale is substantial, driven by legal constraints, software requirements, human verification steps, and plain old slow diffusion.

The 10 most exposed occupations under this framework are led by Computer Programmers, with 75% task coverage, followed by Customer Service Representatives and Data Entry Keyers, at 67%. Roughly 30% of all workers have zero observed exposure — cooks, bartenders, motorcycle mechanics, lifeguards — roles where the physical or interpersonal nature of the work keeps AI at arm's length for now.

What the Employment Data Shows

The researchers matched their exposure measures to Current Population Survey data and examined unemployment trends since 2016. The finding: no statistically significant increase in unemployment for workers in the most AI-exposed occupations since ChatGPT launched in late 2022. The gap between high-exposure and low-exposure worker unemployment is small and not distinguishable from zero.

That's not a full exoneration. The paper is explicit about why. AI labor market effects may operate more like the internet or the China trade shock than like COVID-19, gradual, diffuse, and easily obscured by other economic forces like trade policy and the business cycle. An effect would need to be substantial to be clearly detectable with current data. The researchers estimate that a "Great Recession for white-collar workers" — a doubling of unemployment in exposed occupations from 3% to 6% — would be visible in this framework. Nothing approaching that has appeared.

The Young Worker Signal Is the Part Worth Watching

The more pointed finding concerns workers aged 22 to 25. Drawing on panel data from the Current Population Survey, the paper tracks monthly job-finding rates for young workers entering high-exposure versus low-exposure occupations. The series visually diverges in 2024. Entry into the most exposed occupations drops by roughly half a percentage point per month, representing a 14% decline in the job-finding rate compared to 2022 — a result that is just barely statistically significant. No equivalent decline is observed among workers aged 25 or older.

This echoes separate findings from Brynjolfsson, Chandar, and Chen (2025), who identified a 6 to 16 percent fall in employment in exposed occupations among workers aged 22 to 25, attributed primarily to slowed hiring rather than increased layoffs. The mechanism matters: if companies are using AI to avoid opening entry-level positions rather than displacing existing workers, the disruption shows up in who doesn't get hired — labor market entrants who may leave the workforce entirely rather than appear in unemployment statistics.

The paper acknowledges that young workers who are not hired into exposed roles may remain in their existing jobs, take different positions, or return to school. The data doesn't resolve which. But the signal is consistent enough across two independent datasets to warrant attention.

Who Is Most Exposed — And It's Not Who Most People Assume

One of the more striking findings concerns the demographic profile of highly exposed workers. Looking at pre-ChatGPT data from August to October 2022, workers in the top quartile of AI exposure are 16 percentage points more likely to be female, 11 percentage points more likely to be white, and nearly twice as likely to be Asian compared to unexposed workers. They earn 47% more on average and hold significantly higher levels of education — people with graduate degrees are nearly four times as likely to be in the high-exposure group as in the zero-exposure group.

This cuts against the narrative that AI primarily threatens low-wage, low-skill work. The occupations most exposed to actual AI usage in professional settings are educated, relatively well-compensated, and disproportionately female. The policy and organizational implications of that demographic profile have not received nearly enough attention.

What This Means for Marketing and Growth Teams

For marketing professionals tracking AI's operational impact, this research offers something rare: a disciplined, data-grounded baseline rather than a forecast built on extrapolation. The authors are building a framework designed to detect displacement before it becomes undeniable — a canary metric for the labor market the same way OpenAI's CoT controllability research functions as a canary for AI safety.

The occupations showing early stress — programmers, customer service, data entry, financial analysis — map directly onto roles that marketing and growth teams rely on, outsource to, or are actively automating with AI tools. The slowing of entry-level hiring into these fields has downstream consequences for team composition, institutional knowledge, and the pipeline of experienced practitioners five years from now.

None of that is an argument against using AI. It is an argument for being clear-eyed about what the data is actually showing, rather than what any given stakeholder needs it to say.

The research paper is available directly from Anthropic at anthropic.com/research/labor-market-impacts.

If you want help thinking through how AI adoption is reshaping your team structure and growth operations in ways that hold up over time, Winsome Marketing's strategists can help you build a plan grounded in what's actually happening — not what's projected.