For more than a year, the loudest voice warning that AI was about to gut the white-collar job market belonged to Dario Amodei, the chief executive of Anthropic. This week the most detailed pushback against that forecast came from inside his own company.
On July 24, Peter McCrory, Anthropic’s head of economics, published a long essay on X laying out eighteen months of work on one question: has AI started to cost Americans their jobs? After combing through federal labor statistics and Anthropic’s own records of how people actually use Claude at work, his answer was no. Not yet, and not in any way the headline numbers can pick up.
What the labor data actually shows
McCrory started with the broad picture. The US unemployment rate sat at 4.2 percent in June, a level the Federal Reserve treats as full employment. Job openings roughly matched the number of people looking for work, and employment among prime-age adults was close to a multi-decade high. None of that looks like an economy quietly shedding workers to software.
Then he ran a sharper test. He isolated the occupations where the daily tasks line up most closely with what Claude is used to automate, and set their unemployment rates against roles with far less exposure. If AI were hollowing out jobs, the damage should surface there first. It did not. The more-exposed group showed no relative deterioration at all.
His explanation leans on a word from Wharton professor Ethan Mollick, who calls AI’s abilities “jagged.” A model can draft a passable memo in seconds and then flub something a first-week intern would handle without thinking. When McCrory looked across the Labor Department’s entire catalogue of occupations, not one had all of its tasks handed over to Claude. People slot the tool into their workflow to draft and revise, then keep the parts it cannot do.
A public disagreement with the boss
The essay is striking partly because of who signs McCrory’s paychecks. Amodei has spent the past year pressing the opposite case, and loudly. In May 2025 he told Axios that AI could wipe out half of all entry-level white-collar jobs and drive unemployment to somewhere between 10 and 20 percent within one to five years, and he accused other executives of sugarcoating the risk. By January he was describing AI as a general substitute for human labor. By June he was floating universal basic income and wage insurance, and suggesting that heavy job loss might be built into the technology.
McCrory’s numbers do not prove any of that wrong. They show that the crisis Amodei keeps describing has not turned up in the aggregate figures, and that McCrory does not expect the unemployment rate to be meaningfully higher a year from now for reasons tied to AI. The two of them even agree on who is most at risk. Their fight is over how hard the hit lands, and how soon.
The warning signs under the surface
This is no victory lap. McCrory flags hiring that has already cooled for young workers in AI-exposed roles over the past year, the same trend Stanford researchers have started calling “canaries in the coal mine.” The Bureau of Labor Statistics projects slower employment growth through 2034 for technical writers, data-entry clerks and customer-support staff, the very jobs his analysis marks as most exposed.
He also sees a divide opening between workers who have built AI into the center of how they operate and everyone else. The power users are measurably faster and more productive. The rest are roughly where they were a year ago. That gap tends to show up in hiring and pay well before it reaches the unemployment rate, which is part of why the topline figure can read calm while pressure builds underneath.
Where the gains are going
The more uncomfortable question in the essay is about distribution. If a slice of workers is producing more without their employers adding headcount, the benefit flows to the earnings line and to the people who already know how to drive the tools. Most of them already sit at the higher end of the income scale. Left alone, that widens the distance between the AI-fluent and everyone else, and a widening gap at the bottom eventually drags on consumer spending, a slower and more corrosive problem than a burst of layoffs.
The practical point is about timing. If the disruption is real but still gathering, then companies and policymakers have more runway than Amodei’s timeline suggests. Retraining and safety-net changes are far easier to build before unemployment climbs than after, and that window is open right now. Nothing in McCrory’s data promises it stays open.
For more coverage of AI and the future of work, visit Mylistingo.
Source: Why AI Hasn’t Increased Unemployment, According to Anthropic, The AI Daily Brief.
