OpenAI opens ChatGPT work data: 165 activities, nearly 100 countries

OpenAI’s Economic Research team has released a dataset showing how consumers use ChatGPT for work. According to the company’s announcement, OpenAI Signals data covers 165 work activities across nearly 100 countries, quarter by quarter from Q1 2025 to Q3 2026.

What the dataset measures and leaves out

Each work-related message is mapped to one of 165 Intermediate Work Activities (IWAs). These are broad task groups that recur across many occupations, drawn from the O*NET framework. OpenAI’s examples include creating teaching materials, writing marketing copy, coding and business data analysis. Because O*NET is a standard reference, researchers can line up ChatGPT usage against other occupational and employment statistics.

The scope is narrower than the headline suggests. Only messages sent on individual ChatGPT plans are included, so enterprise traffic and Codex use are out. OpenAI says the release extends OpenAI Signals, its running effort to publish evidence about how people use AI. A download page is linked from the announcement, and the page also carries an interactive chart for browsing activities by country and over time.

Countries are specializing, not converging

OpenAI reports that work use spread out across more activities rather than piling into the most common ones. For the 74 countries present in every quarter, the top ten activities accounted on average for 51.1% of a country’s classified work messages in Q3 2026. In Q1 2025 that share was 56.0%. OpenAI reads this as countries specializing over time, with different countries leaning toward different kinds of work.

The company offers two possible explanations: differences in the work done in each country, or differences in access to AI tools. It does not say which is stronger, and presents the question as one researchers can pursue with the data.

A second comparison sets ChatGPT usage beside the activities that make up U.S. jobs. The two distributions differ substantially. According to OpenAI, roughly four in ten units of employment-weighted task time for mapped U.S. occupations sit in an “Other IWAs” bucket outside the separately reported categories. That bucket holds work such as hauling goods, driving, and keeping medical records. Those activities were too sparse in the original U.S. ChatGPT sample to be tracked consistently, so the gap reflects partly how the data was built, not only what people do.

A research tool with privacy limits and open questions

The practical change is access. Earlier public analysis of ChatGPT usage was a snapshot. In September 2025, OpenAI and Harvard economist David Deming published an NBER working paper on how people use ChatGPT. PYMNTS, summarizing it, reported that non-work messages had risen from 53% to more than 70% of usage, and that practical guidance, seeking information and writing together accounted for nearly 80% of conversations. That study offered one aggregate view. The new release adds country-level detail, a time series and a link to a standard occupational taxonomy.

The citation attached to the release lists nine authors, Deming among them, alongside Aaron Chatterji, Zoe Hitzig and others. That makes it a product of OpenAI’s own economists, which matters for how to read it: the classification of messages, the thresholds and the choice of which categories to report are OpenAI’s decisions.

On privacy, OpenAI describes a process in which no human reviewer reads individual conversations. Software assigns each message to an activity only after personal identifiers have been stripped. The company publishes only aggregated figures, suppresses results for small groups and injects statistical noise. The noise protects users but also limits precision for small countries and niche activities, a point the announcement does not quantify.

Several things remain unestablished. The page does not give message counts, the share of messages that could be classified, or how many countries lack data in some quarters (only 74 are observed in every quarter). Nothing here independently verifies that the classification matches what users actually do, and the findings have not been reviewed outside OpenAI. The data also describes consumer plans, so it says little about how companies use ChatGPT at scale.

For labor economists, the value is a common yardstick. Whether the data reveals real differences in AI adoption between countries, or simply different job mixes, will depend on what researchers do when they join it to employment figures.