AI Hasn’t Dented US Jobs, Yale Budget Lab Study Finds

Nearly three years into the generative AI boom, the overall U.S. job market looks more stable than the headlines suggest, according to new research from The Budget Lab at Yale.

In an analysis — “Evaluating the Impact of AI on the Labor Market: Current State of Affairs” — released Oct. 1, researchers Martha Gimbel, Molly Kinder, Joshua Kendall, and Maddie Lee report no economy-wide employment disruption tied to AI since ChatGPT’s November 2022 debut. Measures of AI exposure, automation, and augmentation show no clear relationship with changes in employment or unemployment, the team said.

“While the occupational mix is changing more quickly than it has in the past, it is not a large difference and predates the widespread introduction of AI in the workforce,” the authors write, adding that better usage data is needed to judge AI’s true labor impact.

What they found

Occupational churn is modest. A dissimilarity index comparing today’s job mix with earlier tech shifts (PCs in the 1980s, the internet in the late 1990s) shows slightly faster but not atypical change since late 2022.

Sector hot spots aren’t new. Information, financial activities, and professional & business services have seen bigger mix shifts than the economy overall, with Information leading. But the trends began before ChatGPT and mirror longer-running volatility in those industries.

Early-career signal is faint. Differences in job mix between recent college grads (20–24) and older grads (25–34) have edged up, which could fit emerging evidence that AI affects early-career workers — but the pattern also matches a cooling labor market and small samples caution against firm conclusions.

Exposure ≠ impact. Using OpenAI’s occupation-level “exposure” scores, the shares of workers in low, middle, and high exposure buckets have been stable since early 2023. Among the unemployed, the average share of tasks exposed (roughly 25–35%) shows no upward trend by duration of unemployment.

Usage is concentrated. Anthropic’s task-level data on its Claude model indicates that actual AI usage skews heavily toward coding and quantitative roles, with arts/media also overrepresented. Depending on how missing tasks are treated, the share of workers in occupations with “automation-heavy” usage ranges from about 70% (observed tasks only) to low single digits (if missing tasks are coded as zero), underscoring how measurement choices swing results.

Why it matters

The findings push back on fears of a swift, economy-wide collapse in demand for cognitive labor. Historically, workplace tech shocks — from PCs to the internet — have unfolded over decades, not months. The Yale team argues it’s too early to draw firm conclusions and pledges regular updates as better data arrives.

Mind the data gaps

OpenAI exposure is theoretical as it doesn’t capture whether firms actually deploy AI.

Anthropic usage reflects one tool and one user base; it likely understates or misstates usage patterns in sectors where other models dominate (e.g., Copilot, Gemini, ChatGPT Enterprise).

The authors call for comprehensive, privacy-safe usage data across leading AI providers, including enterprise and API activity.

Bottom line

For now, AI looks transformative in potential but incremental in labor-market impact. The occupational mix is shifting, but largely along pre-AI trajectories, with no clear jobs shock visible in national data since late 2022.