If you follow the debate over artificial intelligence and employment, you have probably noticed a contradiction. One set of headlines warns that AI is hollowing out the workforce. Another points to a low unemployment rate and asks where the job losses are. Both observations are grounded in real data. The reason they seem to conflict is that they are measuring different things.

Over the past year, at least six independent research teams have published attempts to detect AI's effect on the U.S. labor market. They use different data — payroll records from ADP, millions of job postings, Current Population Survey microdata, Claude usage logs, Microsoft Copilot conversations, and BLS employment statistics. They use different methods — event-study regressions, synthetic differences-in-differences, ex-post occupational comparisons, and large-scale worker evaluations of AI output. And they arrive at a picture that is more specific, and more ambiguous, than either the alarmist or the dismissive camp suggests.

What the payroll data shows

The most widely cited finding comes from Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen at Stanford's Digital Economy Lab, who analyzed high-frequency payroll data from ADP, the largest payroll processor in the United States. Their study, first published in August 2025 and updated in November 2025 and February 2026, tracks employment by age and by how exposed each occupation is to generative AI.

The central result: workers aged 22 to 25 in the most AI-exposed occupations experienced a 13 to 16 percent relative decline in employment compared with less-exposed occupations, after controlling for firm-level shocks. Employment for workers over 25 in the same occupations remained stable or continued to grow. The decline is concentrated in occupations where AI usage is primarily automative — fully delegating tasks — rather than augmentative, where humans and AI collaborate. Software developers and customer service representatives show the sharpest drops.

Crucially, the adjustment appears to happen through hiring, not separations. Existing workers are not being fired in large numbers. Fewer new ones are being brought in.

What the job-posting data shows

Revelio Labs, a workforce analytics firm, tracks AI's labor-market effects monthly using job-posting data. Its July 2026 tracker reports that postings in the most AI-exposed occupations have fallen 42 percent relative to the least-exposed occupations since October 2022. The decline is heavily concentrated at junior seniority levels: postings for junior roles in exposed occupations fell roughly 45 percent, compared with about 17 percent for senior roles.

The supply side is responding too. Computer science and IT enrollment at selected U.S. schools has fallen 28 percent since its 2022 peak, while AI-related certifications on professional profiles have surged. The share of all certifications that are AI-related reached roughly 31 percent in 2026, up from under 4 percent in 2019.

Revelio also tracks firms that have actually adopted AI, identified by their posting of AI-integrator roles. Those firms have grown headcount 27 percent more than non-adopters since October 2022 — but the growth is uneven, with senior headcount rising 31 percent and junior headcount only 6 percent.

What the government surveys show

The Yale Budget Lab's AI Labor Market Tracker, updated in July 2026 with June CPS microdata, finds no clear evidence of AI-related disruption. Occupational churn remains within historical ranges. AI exposure among the unemployed shows no anomalous pattern. A synthetic differences-in-differences analysis of exposed versus unexposed workers does not yet indicate an AI-related footprint. The unemployment rate gap between exposed and unexposed workers, adjusted for pre-AI historical differences, has moved close to zero.

Anthropic's own analysis, published in March 2026, reaches a similar conclusion on unemployment but adds a nuance on hiring. Using its "observed exposure" measure — which weights tasks by whether they are actually being automated in professional Claude usage, not just whether they theoretically could be — Anthropic finds no systematic increase in unemployment for highly exposed workers. But it finds suggestive evidence that the monthly job-finding rate for workers aged 22 to 25 entering exposed occupations dropped roughly 14 percent compared with 2022. The result is described as "just barely statistically significant," and there is no comparable decrease for workers over 25.

Anthropic also documents a gap between theoretical and realized AI capability. LLMs could theoretically perform 94 percent of tasks in the Computer and Math occupational category, but observed professional usage covers only 33 percent. The most exposed workers, measured by actual usage, are more likely to be older, female, more educated, and higher-paid than unexposed workers — a demographic profile that complicates simple narratives about who is at risk.

The ex-post check

A July 2026 working paper from the Federal Reserve Bank of Chicago adds a longer perspective. Kristen Broady, Caleb Dunson, and Anthony Barr matched Frey and Osborne's 2017 automation-risk scores and Tomlinson et al.'s 2025 AI-applicability scores, based on Microsoft Copilot usage, to BLS employment and wage data for 2019 through 2024.

Their finding complicates the displacement narrative from the other direction. Occupations with the highest AI applicability scores experienced overall employment growth of about 4 percent and wage growth of roughly 22 percent between 2019 and 2024. Occupations classified as high-risk under the older Frey and Osborne framework saw weaker employment performance, but wages still rose across all risk categories. The authors conclude that exposure scores are better understood as indicators of occupational restructuring than as direct forecasts of employment decline.

This does not contradict the Stanford or Revelio findings. The Chicago Fed analysis covers 2019 to 2024, a period that includes the pandemic and its aftermath, and it looks at total employment rather than hiring flows. The Stanford and Revelio data capture the period after generative AI became widely available, and they focus on the margin where change is most likely to appear first: new hires.

The capability trajectory

A separate question is how fast AI capability is actually improving across work tasks. A July 2026 paper from MIT FutureTech, based on more than 60,000 evaluations by experienced workers across 6,000 text-based tasks drawn from the O*NET occupational taxonomy, finds that progress follows a "rising tide" pattern rather than "crashing waves." AI performance does not jump abruptly across narrow task sets. It improves gradually and broadly.

In 2024, frontier models completed text-based tasks that take humans about 90 minutes with roughly 60 percent success at a minimally sufficient quality level. By the third quarter of 2025, that figure had risen above 70 percent. If recent trends persist, the authors project success rates of 88 to 97 percent by 2030. The improvement is broad-based across job families, though the slope varies by task structure.

The policy implication the authors draw is double-edged. A rising tide means workers are less likely to be blindsided by sudden automation of their specific tasks. But it also means fewer areas of text-based work will remain fully insulated from automation pressure. The window for institutional adaptation is measured in years, not months — but it is not unlimited.

Where the evidence converges, and where it does not

The convergence across these studies is narrower than the public debate suggests, but it is real. Multiple teams using different data and methods find that the clearest labor-market signal associated with AI is not rising unemployment. It is a slowdown in hiring, concentrated among early-career workers, in occupations where AI usage is primarily automative rather than augmentative.

The divergence is also real. The Chicago Fed's ex-post analysis shows that high-AI-applicability occupations grew over 2019 to 2024. The Yale Budget Lab finds no clear AI footprint in aggregate CPS data. Revelio shows that AI-adopting firms are expanding, even as postings in exposed occupations contract. These are not contradictions so much as different lenses: total employment versus hiring flows, exposure versus adoption, anticipation versus realization.

What remains genuinely unresolved is whether the hiring slowdown for young workers in exposed occupations is the leading edge of a broader structural shift, or a temporary adjustment period after which new roles absorb the displaced demand. The Stanford authors note that their results are consistent with AI affecting entry-level employment but caution that the estimates may be influenced by factors other than generative AI. Their February 2026 update finds that, with the broadest set of controls, the timing of decline in exposed occupations becomes statistically significant only in 2024, suggesting earlier declines may reflect non-AI factors.

The NBER working paper by Karger, Kuusela, and colleagues, which surveyed academic economists, AI-company employees, policy researchers, superforecasters, and the general public, captures the uncertainty directly. The median respondent across all groups expects GDP growth of about 2.5 percent annually — above government baselines but not transformative. Under a "rapid" AI progress scenario, economists forecast the labor force participation rate falling to 55 percent by 2050, with roughly 10 million fewer participants attributable to AI. But the variance around those forecasts is enormous, and expert disagreement is driven primarily by different beliefs about the effects of highly capable AI systems, not by disagreement about the pace of progress itself.

For now, the most defensible summary is this: AI has not produced a visible unemployment crisis. It has produced a measurable hiring adjustment, concentrated at the entry level, in the occupations where it is most actively automating work. Whether that adjustment deepens into displacement or resolves into restructuring is the question the next two years of data will have to answer.

Sources