OpenAI has released a new set of aggregated data about how people use ChatGPT for work. The release, part of OpenAI Signals, organizes work-related messages into 165 activity categories across nearly 100 countries. Its quarterly observations run from the first quarter of 2025 through the third quarter of 2026.
The dataset is designed to help researchers investigate how ChatGPT use differs across types of work, places, and time. It is a data and research release—not a new ChatGPT feature or a change to the product. OpenAI announced it on October 26, 2026, and provides downloadable data and a data dictionary for further analysis. OpenAI’s announcement describes the scope and purpose of the release.
What OpenAI released
The data connects work-related ChatGPT messages to 165 categories called Intermediate Work Activities, or IWAs. OpenAI says these categories draw on O*NET, a standardized framework that groups tasks associated with occupations into broader work activities. Examples mentioned in the announcement include preparing instructional materials, developing marketing content, programming, and analyzing business data.
That activity-based structure matters. Rather than sorting usage only by job title, the dataset gives researchers a way to look at the kinds of work people bring to ChatGPT. Similar activities can occur in different occupations, so a shared classification can help support comparisons across jobs and with other occupational or employment data.
The release reports quarterly observations from Q1 2025 through Q3 2026. OpenAI says researchers can explore the mix of activities in a country, follow how it changes over time, and connect the data with other economic indicators. The figures are aggregated; the release is not a collection of individual conversations.
A finding about concentration—not proof of a labor-market shift
OpenAI highlights a change in how concentrated work-related use was among the most common activities. Across 74 countries with data in every quarter, the ten most common activities made up an average 51.1% of each country’s classified work messages in Q3 2026, down from 56.0% in Q1 2025.
In other words, in this balanced group of countries, the top ten activities accounted for a smaller share of classified work messages at the end of the period than at the start. OpenAI describes the pattern as work-related use becoming less concentrated, with countries tending to specialize in different activity mixes over time.
That result is a useful starting point for investigation, but it should not be treated as evidence that national economies or employment patterns changed in the same way. OpenAI notes that differences might reflect the kinds of work people do in each country or differences in access to AI tools. The dataset can help researchers examine those possibilities; the headline statistic alone does not explain them.
How researchers can use the Signals data
The release can support descriptive comparisons and help develop research questions. For example, a researcher could examine whether the share of messages associated with a particular activity rises or falls across quarters, or compare activity mixes among countries represented in the data. They could then compare those patterns with relevant economic or occupational indicators, while checking that the definitions and time periods align.
OpenAI also presents a U.S. comparison between ChatGPT activity categories and activities performed across U.S. jobs. The company says the distributions differ substantially. One important qualification: about 43% of the employment-weighted task mix for mapped U.S. occupations falls into “Other IWAs,” outside the individually reported categories. OpenAI lists examples such as moving materials, operating vehicles, and maintaining health or medical records, and says these categories did not meet its threshold for consistent representation in the original U.S. ChatGPT sample.
That gap is a reminder that a classification built from observed ChatGPT use is not a complete inventory of all work. When making comparisons, researchers should inspect how activities are grouped and whether categories are missing or combined. In particular, a low or absent ChatGPT share for an activity should not be mistaken for evidence that the activity is rare in the economy.
Practical steps for working with the data
- Start with the data dictionary. OpenAI provides CSV downloads and a dictionary explaining the data. Review the category definitions and fields before calculating shares or combining files. The Signals download and methodology page also provides OpenAI’s suggested citation.
- Keep the denominator clear. A share of classified work messages describes the composition of the messages represented in the dataset. It is not automatically a share of workers, jobs, hours worked, company activity, or national output.
- Use the balanced-country comparison carefully. The top-ten statistic is reported for 74 countries observed in every quarter. That is a specific comparison group, not necessarily the full set of nearly 100 countries in every period.
- Align external data before joining it. If comparing Signals with labor-market or economic measures, check that country coverage, time periods, and activity definitions are comparable. A quarter of ChatGPT messages and a national employment statistic may measure different things.
- Separate description from explanation. A difference between countries or quarters can identify a pattern to study, but the data by itself may not establish what caused it or what it means for productivity, hiring, wages, or job design.
What the dataset does not cover
The scope is a central limitation. OpenAI says this release covers work-related messages on individual ChatGPT plans and excludes enterprise messages and Codex use. As a result, it should not be read as a complete picture of AI use in workplaces, organizations, or technical workflows. OpenAI’s separate Signals materials likewise caution that individual-plan data can underrepresent business and technical use cases.
It also measures ChatGPT messages, not all uses of AI or all work people perform. People may use other tools, may not bring a task to ChatGPT, or may use ChatGPT in ways that are difficult to classify. And a message associated with a work activity does not establish that the output was used, was correct, saved time, or changed the worker’s responsibilities.
Privacy protections shape what is available. OpenAI says researchers do not read individual conversations for this release. According to the company, automated systems classify messages after identifying information is removed; published results are aggregated, small groups are withheld, and statistical noise is added to protect individual messages. Those safeguards are important context for interpreting the figures, and the aggregate data is not a substitute for examining individual-level records.
Why this release is notable
The main contribution is a structured, downloadable view of work-related ChatGPT use across activities, countries, and quarters. A common activity framework can give researchers a more detailed starting point than broad labels such as “writing” or “coding,” while the time series makes it possible to track changes within the period covered.
For people following AI products, the announcement is less about a new capability and more about evidence. It offers researchers, journalists, and policy analysts a source to examine how consumer ChatGPT use intersects with work. Its value will depend on careful analysis: understanding the sample, using the activity definitions accurately, and resisting conclusions that the dataset was not designed to prove.
Frequently asked questions
What does “OpenAI Signals ChatGPT work data” refer to?
It refers to OpenAI’s aggregated Signals dataset that classifies work-related messages on individual ChatGPT plans into 165 Intermediate Work Activities across nearly 100 countries, with quarterly data from Q1 2025 through Q3 2026.
Is this a new ChatGPT feature?
No. It is a research-data release. OpenAI provides an interactive exploration experience and downloadable CSV files, but the announcement does not describe a new ChatGPT capability for users.
Does the data include ChatGPT Enterprise or Codex?
No. OpenAI says the release covers work-related messages on individual ChatGPT plans and excludes enterprise messages and Codex use.
Can the data show whether ChatGPT is replacing jobs?
Not by itself. The dataset describes patterns in classified ChatGPT messages. It does not establish job losses, hiring effects, productivity changes, or causation. Those questions require additional evidence and appropriate research methods.
Where can researchers get the files?
OpenAI provides CSV downloads, a data dictionary, and a suggested citation on its Signals data and methodology page. The data is licensed under Creative Commons CC BY 4.0, which permits use, sharing, and adaptation with attribution.
Bottom line
OpenAI’s Signals release makes global, activity-level data on individual-plan ChatGPT work use available for study. Its coverage—165 work activities, nearly 100 countries, and quarterly observations through Q3 2026—offers a practical resource for examining patterns and forming new research questions. The most responsible reading is also the simplest: these are aggregated patterns in ChatGPT messages, not a direct census of work or a standalone measure of AI’s effects on the economy.
