Economists trying to measure what AI does to work have mostly relied on surveys, which ask people what they remember doing. OpenAI has now released something closer to a record of behaviour: quarterly figures on the kinds of work people bring to ChatGPT, broken down by country, from the first quarter of 2025 to the third quarter of 2026. It is part of OpenAI Signals, the company's programme for publishing usage evidence.
The data maps work-related messages to 165 Intermediate Work Activities, categories taken from O*NET, the US government's standard description of occupations. Because the categories are shared across jobs, researchers can line ChatGPT use up against employment and occupational statistics. Nearly 100 countries are covered, and an interactive chart lets readers move from broad groups such as communicating or analysing down to tasks like preparing teaching materials, writing marketing copy or programming.
The first finding OpenAI highlights is that use is spreading out. Across the 74 countries present in every quarter, the ten most common activities made up an average of 56.0% of classified work messages in early 2025 and 51.1% by the third quarter of 2026. OpenAI reads this as countries specialising, each in different kinds of work, and suggests it may reflect differences in local economies or in access to AI tools. A comparison with the US economy shows how partial ChatGPT's footprint still is: about 43% of the employment-weighted task mix for US occupations falls into activities too rare in ChatGPT to report individually, such as moving materials, driving vehicles and keeping medical records.
For readers who use ChatGPT at work, the data is a mirror rather than advice, but it gives some grounding to arguments about which jobs AI is reaching first. It also lets researchers test claims made by vendors, including OpenAI, against observed use.
The limits are significant. The release covers only individual ChatGPT plans; enterprise accounts and Codex are excluded, which leaves out much of the heaviest professional use, including most programming. Every message is labelled by OpenAI's own automated classifier, so any bias in that model flows straight into the figures. To protect privacy, OpenAI removes identifying details, withholds results for small groups and adds statistical noise. All of that is sensible, but it means outsiders cannot check the classification. Treat it as one company's view of its own consumer users, not a measure of the labour market.