The News, Explained
OpenAI released new Signals data on October 26, 2026, examining the work people bring to ChatGPT. It allows country-by-quarter comparisons of work-related messages from Q1 2025 through Q3 2026. Workers considering where AI might help with their own tasks can use it as a reference. Source
The categories come from O*NET Intermediate Work Activities, or IWAs: groups of tasks shared across occupations. Preparing teaching materials or analyzing business data are examples. The focus is the requested activity, rather than a job title. Classifiers infer work-relatedness and activity from message context; rare activities are grouped into “Other IWA.” Source Source
OpenAI reports that, across 74 countries observed in every quarter, the ten most common activities accounted for an average of 56.0% of work messages in Q1 2025 and 51.1% in Q3 2026. Each country’s work-message mix is calculated first, then those country-level shares are averaged. In this sample, requests became less concentrated in the most common activities and spread toward other activities. Source
| Measure | What it counts | What to keep aligned |
|---|---|---|
| Activity share by country and quarter | The mix of that country’s work-related messages | Country and classification system |
| Average top-ten activity share | The average of shares calculated for each of 74 countries | The same countries observed in every quarter |
| Overall Signals sample | Messages sampled from individual accounts | Data version and included population |
These distinctions establish the denominator. A share is the portion of messages in the relevant group assigned to an activity. People can send different numbers of messages, so translating that mix into a headcount or an occupational distribution changes the object being measured. Source
The methodology describes an overall sample of one million individual ChatGPT messages per month from July 2024 through September 2026. Accounts must have a declared age of at least 18; users who deleted their accounts or opted out of training are excluded. The individual plans include Free, Go, Plus and Pro. Enterprise messages and Codex use are excluded, and OpenAI says business and technical uses are therefore likely underrepresented. Source Source
Version 3.0 also revises historical periods from 300,000 to one million sampled messages per month. Comparing earlier and later quarters within this current version provides a consistent basis. Joining a value saved from an older release to a new one would mix usage changes with sample revisions. Source
Researchers do not read raw conversations for this analysis. Automated classifiers work on messages scrubbed of identifying information. Published statistics use differential privacy, adding random noise to protect individual messages. The country-quarter IWA analysis also requires at least 1,000 effective messages per quarter after weighting and noise. The methodology notes that smaller groups can have greater sampling variation. Source
OYOPICK’s Take
We see the value in asking which specific tasks someone might share with AI. Common activity categories provide a starting point for comparing similar preparation or analysis requests across different job titles. Whether a wider range of requests produces better results needs further evidence, such as the accuracy of task outputs or users’ assessments.
The historical sample revision deserves attention too. Being able to compare a time series within one version lets readers revise their judgments as the evidence changes. If future updates continue to disclose sampling and classification changes, this could provide a firmer basis for examining how AI expands the work people can do.