Nearly three years after ChatGPT launched, artificial intelligence has not caused widespread job displacement across the economy, according to a new analysis from Yale’s Budget Lab and the Brookings Institution.
The study, released Oct. 1, tracked U.S. employment data over 33 months from November 2022 to July 2025 and found no measurable disruption in how workers are distributed across occupations — a key indicator of technological impact on jobs.
“Our metrics indicate that the broader labor market has not experienced a discernible disruption since ChatGPT’s release 33 months ago, undercutting fears that AI automation is currently eroding the demand for cognitive labor across the economy,” the researchers wrote.
Occupational shifts mirror historical patterns
The research team compared current labour market changes to previous periods of technological disruption, including the introduction of personal computers in the 1980s and mass internet adoption in the 1990s.
The occupational mix — how workers are distributed among all jobs in the economy — has changed about seven percentage points since ChatGPT’s launch, according to the study. That pace is only one percentage point higher than changes seen during internet adoption from 1996 to 2002.
The researchers noted these shifts began before ChatGPT’s November 2022 release, suggesting the changes are not necessarily caused by AI.
High-exposure sectors show pre-existing trends
Industries considered most exposed to AI showed larger occupational shifts, according to the analysis. The information sector, which includes newspapers, movies and data processing, experienced the most change. Financial activities and professional and business services also showed above-average shifts.
However, data indicates these trends started before ChatGPT became available, the researchers found.
No evidence of AI-driven unemployment
The study examined whether workers in AI-exposed occupations faced higher unemployment rates. The analysis found no clear patterns.
Among unemployed workers, about 25 to 35 per cent had previously worked in jobs where tasks could theoretically be performed by generative AI, regardless of how long they had been unemployed, according to the report. That proportion remained stable with no upward trend.
College graduates show limited impact
The research examined recent college graduates aged 20 to 24, comparing their occupational distribution to workers aged 25 to 34. While some recent data suggested slightly faster divergence between these groups, the pattern appeared to predate ChatGPT’s release.
The researchers cautioned that small sample sizes limit definitive conclusions about this demographic.
Gap between exposure and actual usage
The study revealed a disconnect between which jobs theoretically could be affected by AI and which jobs actually use the technology.
OpenAI’s “exposure” data measures whether AI could reduce task completion time by at least 50 per cent. Anthropic’s usage data tracks actual interactions with its Claude chatbot. The two measures showed limited correlation.
Computer and mathematical occupations, including coders, dominated actual AI usage in the Anthropic data. Arts and media workers were also overrepresented. Meanwhile, clerical sectors with high theoretical exposure showed much lower actual adoption.
Data limitations require caution
The researchers acknowledged significant limitations in available data. OpenAI’s exposure metrics are theoretical rather than based on actual workplace adoption. Anthropic’s usage data comes from only one AI tool and skews heavily toward coding and writing tasks.
“To accurately measure AI’s impact on the labor force, the most important data needed is comprehensive usage data from all the leading AI companies at the individual and enterprise level, including APIs,” the researchers wrote.
Historical context suggests slow adoption
The lack of immediate disruption aligns with historical patterns of workplace technology adoption, according to the study. Computers took nearly a decade after public release to become commonplace in offices, and even longer to transform workflows.
The researchers plan to update their analysis monthly as more data becomes available. They emphasized that current stability does not predict future impacts.
The study analyzed U.S. Current Population Survey data and used dissimilarity indexes to measure how occupational composition changed over time. Canadian labour markets typically follow similar patterns to U.S. trends, though specific impacts may vary.
