OpenAI publishes ChatGPT work-use data
On 26 October 2026 OpenAI's economic research team published the data set OpenAI Signals v3.0: shares of work-related messages in individual ChatGPT plans across 165 types of work activity (the O*NET classification) in almost 100 countries, quarterly from the first quarter of 2025 to the third quarter of 2026. The data are aggregated with statistical noise added and licensed CC BY 4.0.
Why it matters
Researchers get open data that can be set beside occupational statistics: which work tasks people bring to ChatGPT in different countries and how that changes over time. OpenAI's own first finding: in the 74 countries observed every quarter, the ten most common activities were on average 51.1% of classified work messages in the third quarter of 2026, against 56.0% in the first quarter of 2025. The data cover only individual plans and were collected by OpenAI itself.
What the pages say. The post of 26 October 2026 (section Research) says the data cover 165 Intermediate Work Activities (IWA) from the O*NET database in almost 100 countries, quarterly from the first quarter of 2025 to the third quarter of 2026; an interactive chart shows the distribution by country. Work use has become less concentrated on the most common activities: in the 74 countries observed every quarter the ten most common made up on average 51.1% of classified work messages in the third quarter of 2026 and 56.0% in the first quarter of 2025. A comparison with US occupations shows that the two distributions differ markedly, largely because about 43% of the tasks of the matched occupations (weighted by employment) fall into "Other IWAs", which the sample does not report separately. The data cover work-related messages in individual ChatGPT plans, without enterprise messages and without Codex. For privacy, researchers do not read single conversations, automated systems classify messages after identifying information is removed, and only aggregates are published, with no small groups and with noise added. The download page: CSV files and a data dictionary under the CC BY 4.0 licence, cited as "OpenAI Signals v3.0" with nine authors. What the record does not claim. All numbers are OpenAI's own and were not checked independently, and they describe only consumers in individual ChatGPT plans, not the use of AI in the economy at large. The CSV files and the data dictionary were not opened; the classification method, the sample and the noise are judged only from the description in the post. The post calls Signals the continuation of ongoing work; the atlas holds no earlier record of the series.