OpenAI opens up ChatGPT work data covering 165 activities in nearly 100 countries

OpenAI released a new batch of Signals data on Monday that maps how people use ChatGPT for work across roughly 100 countries, classified into 165 activities and reported quarterly from the first quarter of 2025 through the third quarter of 2026.

The classification is built on O*NET, a standardized framework that groups occupation-specific tasks into broader work activities shared across occupations. OpenAI maps work-related messages to 165 “Intermediate Work Activities,” or IWAs, so that researchers can compare ChatGPT usage with existing occupational and employment data on a common footing rather than relying on categories the company invented.

One finding is already visible: work use is becoming less concentrated. Across 74 countries observed in every quarter, the ten most common activities accounted for an average of 51.1% of each country’s classified work messages in the third quarter of 2026, down from 56.0% in the first quarter of 2025. OpenAI reads that as countries specializing over time, with different countries specializing in different kinds of work.

The data also shows wide variation between countries. OpenAI’s interactive charts let users drill down by country and activity group to see whether a given task is growing or shrinking. Quarterly series of this kind are rare in debates about AI’s economic impact, where research often has to infer effects from surveys or hiring data.

There are clear limits. The dataset covers work-related messages on individual ChatGPT plans only; it excludes enterprise messages and Codex use. In other words, it describes consumers and individual professional users, not what happens after a company buys and deploys the product.

OpenAI also flags a structural gap in comparisons with the U.S. labor market. About 43% of the employment-weighted task mix for mapped U.S. occupations falls into “Other IWAs,” including activities such as moving materials, operating vehicles and maintaining medical records. Those categories did not meet the threshold for consistent representation in the original U.S. ChatGPT sample, which helps explain why the two distributions look so different.

OpenAI frames the release as research infrastructure rather than a product update: the data sits on its Signals site, and researchers are invited to connect it with other economic indicators. For policymakers, labor economists and employers, the value is not a conclusion but a reproducible baseline: when AI use spikes in a given occupation, there is at least a public quarterly series to check it against.

The next question is coverage. Enterprise and Codex usage remain outside the dataset, even though those are the places where AI is most deeply embedded in production workflows. Including them in later releases would make the picture of AI’s effect on jobs considerably more complete.