Yale Budget Lab finds no labor-market disruption 33 months after ChatGPT

Thirty-three months after ChatGPT’s public debut, the U.S. labor market shows no measurable disruption from generative AI. That is the central finding of a new analysis from the Yale Budget Lab, published in October 2025 by Martha Gimbel, Molly Kinder, Joshua Kendall, and Maddie Lee. The paper directly undercuts the most alarming headlines about AI-driven job losses.

The Budget Lab team compared the occupational mix — the distribution of workers across all jobs — in the 33 months since ChatGPT’s November 2022 launch against three earlier periods of technological change: the computerization wave of the 1990s, the internet adoption of the early 2000s, and the period leading up to and during the pandemic. Their finding: the occupational mix has changed only about 1 percentage point faster than it did during the internet adoption era. “The broader labor market has not experienced a discernible disruption since ChatGPT’s release 33 months ago,” the authors write, “undercutting fears that AI automation is currently eroding the demand for cognitive labor across the economy.”

This is not a contrarian outlier. It is consistent with historical precedent. Computers did not become commonplace in offices until nearly a decade after their release to the public, and it took even longer for them to transform workflows. The Budget Lab’s data shows that the occupational mix shifts currently attributed to AI were “well on their way during 2021, before the release of generative AI.” The changes predate the technology.

The paper examines two specific domains where AI’s impact should be most visible: highly exposed industries and recent college graduates. The Information, Financial Activities, and Professional and Business Services sectors — industries with the highest exposure to generative AI — have seen larger shifts in job mix than the aggregate labor market. But again, the data shows those trends started before ChatGPT. The Information sector’s volatility, the authors note, “seems to be a feature of the industry itself rather than a consequence of any one technological development.”

For recent college graduates (ages 20–24), the occupational mix has diverged slightly from that of older workers (ages 25–34) in recent months. This could be consistent with research from Erik Brynjolfsson and colleagues showing a possible impact of AI on early-career workers. But the Budget Lab team cautions that the divergence “rarely deviates outside of the 30-33% range” since January 2021 and may simply reflect a slowing labor market hitting younger workers harder. The sample sizes are small enough that the authors flag them explicitly.

The paper also uses the best available exposure data — from OpenAI and Anthropic — to ask whether workers in occupations most exposed to generative AI are seeing employment changes. OpenAI’s exposure metric measures whether GPT-4 technology can reduce task completion time by at least 50% for a given occupation. Anthropic’s data tracks actual usage of Claude. Neither shows a relationship with employment shifts.

The share of workers in the lowest, middle, and highest occupational exposure groups has stayed stable at roughly 29%, 46%, and 18% respectively since ChatGPT’s launch. Even among the unemployed, there is no clear growth in exposure to generative AI. The Budget Lab team examined unemployment duration by exposure level and found no pattern consistent with AI-driven displacement.

None of this means AI will not reshape the labor market. The paper is explicit about its limitations. “Our analysis is not predictive of the future,” the authors write. They plan to update the analysis monthly. The historical pattern is clear: widespread technological disruption in workplaces tends to occur over decades, not months or years. The 33-month window since ChatGPT is, in the context of economic history, a blink.

The paper’s real value is in providing a factual baseline against which future claims can be measured. Every new study claiming AI is destroying jobs or creating a new class of displaced workers now has a clear benchmark: the occupational mix changed about 7 percentage points between 1996 and 2002 during the internet adoption period. The AI-era change so far is roughly 1 point faster. That is the number to beat.

For AI builders, the implication is not that their work is irrelevant to labor markets. It is that the time horizon for labor-market effects is measured in years and decades, not months. The technologies that will actually transform work are likely those that are embedded in workflows, not just used as standalone tools. The Budget Lab data suggests that the current generation of generative AI tools, however impressive in demos, have not yet reached the level of integration into business processes that computers and the internet achieved over the course of the 1990s and early 2000s.

The most honest conclusion from the Budget Lab’s analysis is also the most boring: we do not know yet. The data cannot distinguish between a future where AI transforms the labor market and a future where it does not. What it can do is establish that 33 months in, the fear of mass displacement is not supported by the evidence. That is worth knowing.