New ChatGPT work data is opening a window on a question that employers, workers and governments have struggled to answer: what are people actually doing with artificial intelligence during the working day? OpenAI’s release on 26 October 2026 organises work-related conversations into 165 kinds of activity and covers nearly 100 countries. It is a map of requests made to a tool, not a map of jobs that have disappeared.
The difference matters. Someone may ask ChatGPT to prepare teaching material, explain a spreadsheet or help draft advertising copy without surrendering the whole task. Yet seeing where such requests cluster could reveal how AI becomes part of ordinary work long before job titles change. The new figures offer a way to examine that shift across occupations and national borders without pretending every prompt amounts to automation.
What the ChatGPT work data actually measures
The release uses a classification system linked to O*NET, an established framework for describing the activities involved in jobs. Instead of sorting conversations only into broad labels such as writing or computing, OpenAI connects them to common activities shared by different professions. That makes it possible to compare an office worker analysing information with someone carrying out a similar activity in a different industry.
The published series runs quarterly from the first quarter of 2025 to the third quarter of 2026. A researcher can examine how the share of work-related messages associated with particular activities changes within a country, then compare those patterns with other economic information. The company says the most common activities are becoming less dominant in the mix, suggesting that use is spreading into a wider variety of work.
That is an interesting observation, not an explanation. One country might make more requests about instructional material because of its occupational structure, language, internet access or adoption of the product. Another might use ChatGPT differently because employees already have dedicated workplace systems. The figures cannot identify which factor is responsible without additional research.
Why the results are not a jobs-replaced scoreboard
The underlying Signals data comes from individual ChatGPT accounts, rather than the full population of employees using AI. Enterprise messages and Codex activity are excluded. That creates an important blind spot: a business could be using AI extensively inside a managed account while barely appearing in this particular dataset.
Nor does a work-related conversation reveal whether its answer was useful. An employee might request a formula, reject it and complete the job by hand. A teacher might use a suggested lesson plan as a starting point, spend an hour revising it and never put the original wording in front of pupils. Counting the conversation describes a request, not the quality of the eventual work.
That distinction echoes earlier questions about reported AI use at work. A person saying they use AI does not tell an employer how much of their working day is genuinely saved. Adoption, productivity and employment are connected questions, but they demand different evidence and different measurements.
A second sensible use of the release is to compare an activity’s visibility in ChatGPT messages with its prominence in employment statistics. Such a comparison might reveal which routine activities people are most inclined to bring to a chatbot, but it cannot show who is doing the work or whether the activity is performed to a professional standard. The starting population is people using an individual account, not a representative census of employees.
Privacy protections and the limits of the sample
OpenAI says researchers do not read individual conversations to assemble the public statistics. Its systems remove identifying information before classifying messages, publish aggregated results, suppress data for small groups and introduce statistical noise designed to protect privacy. These safeguards also mean users should not expect the resulting charts to reproduce every small fluctuation exactly.
The data has another limitation: the classification is automated. A complicated prompt can contain several activities, while a short request may be ambiguous even to a human reader. The categories enable consistent large-scale analysis, but they remain an interpretation of text. They do not provide an independent account of what happened later in the workplace.
What employers and workers can learn next
For an employer, the practical value is not copying the most popular activity in another country. It is asking which tasks in the organisation involve repeated preparation, searching, editing or checking, and whether AI helps without creating new mistakes. Staff should be able to explain what was checked independently and what remains a human decision.
Workers may find the changing mix more useful as a guide to emerging expectations than as a personal forecast of redundancy. If AI is moving beyond generic email writing, the valuable skill may be knowing how to frame a problem, verify an answer and integrate a tool into a real process. A polished result is not necessarily the same as a well-judged one.
That point also connects with research into how AI changes the content of jobs. The greatest effect may sometimes be a redesign of responsibilities, not a straightforward replacement of one worker by one machine. Measuring the effect requires outcomes such as error rates, turnaround time, quality and employee experience.
OpenAI has made the data available for researchers to explore rather than presenting a definitive verdict about the future of work. Its most valuable finding may emerge when independent analysts connect the activity patterns to surveys, employment statistics and rigorous workplace experiments. The public can now ask better questions. It should be cautious of anyone claiming this release already supplies all the answers.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.