OpenAI Economic Research has released a dataset that maps ChatGPT usage to 165 Intermediate Work Activities (IWAs) across nearly 100 countries. The data are provided at quarterly frequency and cover the period from Q1 2025 through Q3 2026.
What the release contains
- The dataset connects ChatGPT messages to 165 IWAs, task categories defined in the O*NET framework that group occupation-specific tasks into work activities shared across jobs.
- Examples of work activities represented in the data include preparing instructional materials, developing marketing content, programming, and analyzing business data.
- Data can be explored by country and over time, allowing researchers to compare how ChatGPT is used for work across locations, track changes, and link usage to other economic indicators.
Key observations disclosed by OpenAI
- Work-related ChatGPT usage became less concentrated in the most common activities over the covered period. For 74 countries observed in every quarter, the ten most common activities accounted for an average of 51.1% of a country’s classified work messages in Q3 2026, down from 56.0% in Q1 2025.
- The release finds that countries tended to specialize in different types of work over time; these cross-country differences may reflect variation in the kinds of work performed locally or differing access to AI tools.
- OpenAI provides paired visualizations comparing the distribution of ChatGPT activity usage with the activity mix performed across U.S. jobs. The two distributions differ notably, in part because roughly 43% of the employment-weighted task mix for mapped U.S. occupations falls into Other IWAs (for example, moving materials, operating vehicles, maintaining health records), categories that did not meet the threshold for consistent representation in the original U.S. ChatGPT sample and therefore were not reported individually.
Which messages are included and excluded
- The data cover work-related messages sent from individual ChatGPT plans. They do not include Enterprise messages or usage of Codex.
Research uses and limitations
- Using the standardized O*NET framework enables comparisons between ChatGPT message content and other occupational or employment datasets.
- The dataset supports a range of research questions, such as whether observed national specializations reflect real differences in economic structure or differences in access to AI tools.
Privacy and methodology
- OpenAI states that researchers do not read individual conversations. Automated systems classify messages after removing identifying information.
- The published output consists of aggregated statistics, withholdings for small groups, and added statistical noise to protect individual messages.
Purpose and next steps
- This release builds on OpenAI Signals, the company’s ongoing effort to share evidence about how people use AI. OpenAI invites researchers to explore the new dataset, combine it with other sources, and investigate what the observed patterns imply for workers and businesses.
