Anthropic’s head of economics spent 18 months on a question his chief executive had already answered, and on 24 July he published a different answer. Peter McCrory’s essay, posted to X rather than to Anthropic’s research site, concluded that AI has caused no material rise in US unemployment, directly against Dario Amodei’s repeated warnings of a white-collar jobs crisis.

The claim carries weight because of where it comes from, and needs careful reading for the same reason. Anthropic sells the technology whose labour-market effects it is measuring, and the exposure measure underpinning the finding is built partly on how people use Claude. For marketing managers being asked to justify AI spending or defend a junior vacancy, the gap between what this evidence shows and what it is being used to argue is the whole of its value.

What did McCrory’s essay actually claim?

McCrory opened on aggregate labour data rather than on AI. US unemployment stood at 4.2% in June, a level the Federal Reserve associates with full employment; job openings roughly matched the number of unemployed workers; and prime-age employment sat near multi-decade highs. He then reported that updated Bureau of Labor Statistics figures show no relative deterioration in unemployment among workers whose jobs contain a large share of tasks Claude is used to automate, compared with workers in less exposed roles. He does not expect AI to push unemployment noticeably higher within a year.

His explanation rests on what he called AI’s “stubbornly jagged” capability profile, borrowing a term popularised by Wharton’s Ethan Mollick. No occupation in the Labor Department’s O*NET taxonomy has all of its tasks covered by Claude, and complex work still depends on people to direct the system and catch its errors. McCrory pointed to usage patterns in which people fold Claude into a drafting and refining loop rather than handing over whole tasks, and in which those with deeper domain expertise both succeed more often and recover better when the model stumbles.

Where AI is showing up, on his account, is productivity rather than headcount: outlets summarising the essay report him citing annual output-per-hour growth of roughly 2% from early 2022 to early 2026, against about 1.6% before the pandemic. Independent work is more cautious. A February 2026 Federal Reserve Bank of Kansas City bulletin found US labour productivity running above its pre-pandemic trend since late 2022, but concentrated in a small set of industries, and reported that while higher AI adoption tracks faster productivity growth across industries, it explains little of the aggregate shift.

How solid is the research behind it?

Anthropic published the underlying research itself. “Labor market impacts of AI: A new measure and early evidence”, released on 5 March 2026 by Maxim Massenkoff and Peter McCrory, introduces a metric called observed exposure, which combines O*NET’s roughly 800 occupations, task-level capability estimates from Eloundou et al. (2023), and Anthropic’s own record of how Claude is used, weighting automated use above augmentative use.

The resulting picture is specific. Claude covers 33% of tasks in the Computer and Math category. Computer programmers are the most exposed occupation at 75% coverage, data entry keyers 67%. Thirty per cent of workers have zero coverage. Workers in the most exposed quartile earn 47% more on average than the unexposed group, and 17.4% hold graduate degrees against 4.5% of the unexposed. Against BLS projections to 2034, every 10-percentage-point rise in coverage corresponds to a 0.6-percentage-point fall in projected employment growth.

Three limits matter. The measure reads AI adoption through one vendor’s traffic, so occupations served mainly by other models are underweighted by construction. The authors are candid that observed use falls well short of theoretical capability, which means the finding describes deployment as it stands rather than as it may stand in a year. And the entry-level result is thin: the paper puts the drop in the monthly job-finding rate for 22 to 25-year-olds entering highly exposed occupations at about 14% against 2022, then notes the estimate is “just barely statistically significant”. Figure 7, which carries that finding, was corrected on 8 March after its labels were reversed.

Why do marketers expect losses the data does not show?

Marketers’ expectations have moved far faster than the employment data. SmarterX’s 2026 State of AI for Business report, based on 2,109 professionals surveyed between February and April 2026, found 71% expect AI to eliminate more jobs than it creates against 13% expecting net creation, while only 20% are concerned about their own role. Among marketers the share expecting disruption rose from 53% to 70% in a single year. The sample is not neutral: it was recruited through SmarterX and Marketing AI Institute channels, so respondents skew more AI-aware than the wider workforce, and 84% work at B2B organisations with marketing the largest function at 32%.

That gap between measured unemployment and expected disruption is not irrational, because the two measure different things. Unemployment counts people out of work and looking; it does not count a vacancy quietly left unfilled, a contractor not renewed, or a team held flat while its remit grows. The Helm reported in June that marketing budgets have been flat for a third year while AI funds headcount cuts rather than growth — a departmental pattern the aggregate unemployment rate is not built to detect.

The entry-level signal is the one that lands closest to marketing practice. Erik Brynjolfsson and colleagues, whose “canaries in the coal mine” work Anthropic cites, report a 6% to 16% fall in employment among 22 to 25-year-olds in exposed occupations, driven by slower hiring rather than by separations; the range is wide because it is measured against several counterfactuals. Marketing teams have long built senior capability by hiring juniors and letting them learn on low-stakes work. That is precisely the work most exposed to automation, and precisely where both Anthropic’s own data and the independent research show the earliest movement.

McCrory has said Anthropic intends to revisit the analysis as new usage and employment data arrive. The Bureau of Labor Statistics publishes its preliminary second-quarter 2026 productivity figures on 6 August.