The US labour market has not experienced discernible disruption in the 33 months since ChatGPT’s release, with measures of AI exposure, automation and augmentation showing no relation to changes in employment or unemployment, according to research from Yale University’s Budget Lab.
The analysis examined whether the pace of labour market change differs from past periods of early technological change and whether economy-wide employment effects are evident. Researchers compared the rate of change in the occupational mix since ChatGPT’s November 2022 launch with disruptions caused by computers and the internet.
The occupational mix changed by around seven percentage points during the internet era between 1996 and 2002, meaning only seven per cent of workers would need to switch occupations to match the 1996 labour market composition. Recent changes appear on a path only about one percentage point higher than the turn of the 21st century, with shifts in the occupational mix already underway during 2021 before generative AI’s release.
The information, financial activities, and professional and business services sectors saw larger shifts in job mix compared to aggregate labour market changes. However, trends within these industries began before ChatGPT’s release, with significant shifts in the information industry emerging as a characteristic of the sector itself, rather than a consequence of any single technological development.
Analysis of OpenAI’s exposure data revealed stable proportions of workers in the lowest, middle, and highest occupational exposure groups, at approximately 29%, 46%, and 18%, respectively. Amongst unemployed workers, 25 to 35 per cent of tasks on average could be performed by generative AI regardless of unemployment duration, with no clear upward trend.
Anthropic’s usage data showed employment proportions in occupations with high levels of task AI automation or augmentation remained stable at around 70 per cent or 11 per cent, respectively, when considering observed tasks only. Computer and mathematical occupations dominated Claude usage, with arts and media also considerably overrepresented.
The researchers noted that comprehensive usage data from all leading AI companies, at both the individual and enterprise levels, is needed to accurately measure AI’s impact. The analysis will be updated regularly moving forward to track how AI’s job impacts change over time.
The research was conducted by Martha Gimbel, Molly Kinder, Joshua Kendall and Maddie Lee.
