Anthropic’s head of economics, Peter McCrory, published a lengthy X essay this week arguing AI has caused no material rise in US unemployment—a data-driven rebuttal that lands in sharp tension with CEO Dario Amodei’s repeated warnings of an imminent white-collar bloodbath.
The prediction that started it
Amodei has staked out some of the industry’s most alarming public positions on AI’s labor impact, though his framing has shifted considerably over the past year.
In May 2025, he told Axios AI could wipe out half of all entry-level white-collar jobs and spike unemployment to 10%-20% within one to five years, urging companies and policymakers to stop “sugarcoating” the risk. He doubled down in a January 2026 essay, “The Adolescence of Technology,” warning AI functions as a “general labor substitute for humans” that will displace work from the lower tiers of skill levels to the upper, potentially creating a lasting underclass of unemployed or very-low-wage workers.
By this May, he had moderated somewhat, reframing automation as a multiplier of output and invoking the Jevons paradox recently repopularized by Apollo Global Management’s Torsten Slok: “If you automate 90% of the job, then everyone does the 10% of the job,” he said, explaining “the 10% kind of expands to be 100% of what people do and kind of 10-times their productivity.” The following month he reescalated, arguing in June significant, enduring job loss might be “an intrinsic property of the technology” itself, and calling for government responses including wage insurance and universal basic income.
What the company’s own data shows
McCrory’s analysis, framed as a synthesis of 18 months of Anthropic’s internal economic research, tells a markedly different story. We “don’t see significant impact of AI on the U.S. labor market,” he wrote, at least not yet. The unemployment rate stood at 4.2% in June—a level the Federal Reserve associates with full employment—while job openings roughly matched the number of unemployed workers and prime-age employment sat near multi-decade highs.
Most pointedly, McCrory says updated analysis using more recent Bureau of Labor Statistics data shows no relative deterioration in unemployment among workers whose jobs contain a large share of tasks that Claude is used to automate, compared with workers in less-exposed roles.
“I don’t expect unemployment to be noticeably higher a year from now—at least not because of AI,” he wrote.
McCrory’s explanation for the gap centers on what he calls AI’s “stubbornly jagged” capability profile, borrowing the term made famous by Wharton’s Ethan Mollick: No job in the Labor Department’s O*NET taxonomy has all of its tasks handled by Claude, and complex work still depends on human oversight to direct systems and catch their errors. He pointed to evidence Claude usage correlates with users acting as “thought partners” rather than replacements, and people with more domain expertise succeed more often and recover better when the AI stumbles—the opposite of a scenario in which AI simply substitutes for human labor.
McCrory’s data most directly undercuts Amodei’s “general labor substitute” theory. His finding that unemployment among highly AI-exposed workers shows no relative deterioration compared to less-exposed workers is hard to square with a scenario in which AI is already acting as a wholesale replacement for human labor. If entry-level consultants, lawyers, and financial analysts were being systematically substituted out at the pace Amodei described to 60 Minutes in November 2025, McCrory’s occupation-level unemployment data should already show some divergence. By his account, it doesn’t.
The Jevons-based multiplier theory is actually closer to McCrory’s own “skill-biased, labor-augmenting” framework. Both describe AI expanding what a smaller number of workers can accomplish rather than deleting jobs wholesale. But there is an unresolved tension here: The Jevons paradox requires time for markets and workers to adjust, yet Amodei has repeatedly said AI is moving faster than any past general-purpose technology, which is precisely the condition under which the Jevons rebalancing breaks down. McCrory’s essay implicitly acknowledges the same soft spot. He notes hiring has already softened for young workers in AI-exposed roles, exactly the population that a slow, aggregate-level “bigger pie” effect wouldn’t necessarily protect.
McCrory notes hiring for young workers in highly AI-exposed roles has softened over the past year, consistent with Stanford research on “canaries in the coal mine,” and flags the Bureau of Labor Statistics projects slower growth through 2034 for occupations like technical writers, data entry workers, and customer support reps.
So the fairer story isn’t that McCrory proved Amodei “wrong” outright: It’s that the data undermines the specific magnitude and timeline of Amodei’s original crisis scenario, while leaving open (and even lending some support to) his more recent, more moderate multiplier framing. Both men agree early-career, entry-level workers in highly exposed roles are the most vulnerable group right now. They just disagree about whether that’s evidence of a coming catastrophe or a normal adjustment period within a still-healthy labor market.
McCrory also conceded the future is uncertain: If AI begins automating innovation itself through recursive self-improvement, standard economic models allow for a scenario resembling the “singularity” that Amodei fears—just not, in McCrory’s read, on the near-term timeline or scale that his boss has publicly forecast.
That leaves Anthropic publicly straddling two positions: an in-house data scientist insisting the disruption hasn’t shown up yet, and a chief executive who has said many times now the disruption is coming fast and could be severe enough to warrant a policy response on the scale of universal basic income.
For this story, Fortune journalists used generative AI as a research tool. An editor verified the accuracy of the information before publishing.
