Evidence at a glance
From “Work-Related” to Specific Activities
On October 26, 2026, OpenAI Economic Research released OpenAI Signals data that maps work-related messages from consumer ChatGPT to 165 work activities and aggregates them by country and quarter. It covers nearly 100 countries and uses O*NET’s Intermediate Work Activities (IWA) framework, with quarterly data from Q1 2025 through Q3 2026. The research question is what kinds of work people bring to ChatGPT, and how those uses vary by place and over time.
The change is not a new census of the labor market. It is a product-use signal organized as a sequence of categories that can be compared with occupational and employment data. For researchers, activity-level categories make it easier to pose testable questions than a broad measure of “work use.” For technical leaders, they offer a view into how users bring a tool into their work, but the view captures the composition of ChatGPT messages, not the composition of all work.
A Broader Mix Does Not Mean More Use
The release’s most quotable figure is that, across 74 countries observed every quarter, the ten most common activities accounted for an average of 56.0% of each country’s classified work messages in Q1 2025, falling to 51.1% in Q3 2026. That means the leading activities made up a smaller share of classified messages, so the mix became less concentrated. OpenAI describes the pattern as countries developing different areas of specialization in their work-related use.
But this is a share, not a count of work-related messages or a measure of ChatGPT adoption among workers. A lower share for the top ten could simply mean that other categories took up more of the classified mix. The figure alone cannot tell us whether people are bringing work to ChatGPT more often. Any claim about adoption scale or growth needs to keep composition measures separate from totals and user coverage.
The Classification Framework Is Part of the Measurement
Users do not fill in a form saying which work activity they are performing. According to the data dictionary, automated systems first identify work-related messages, then use prompt-driven model labels to assign activities. The rules require the message itself to indicate what the user is trying to do; ambiguous cases can be marked unclear. Results are then aggregated by country and quarter. O*NET provides a shared activity framework across occupations, allowing ChatGPT messages to be compared with occupational and employment data in a common vocabulary.
Researchers are therefore looking at a measurement produced through model judgment, classification, and aggregation, not an uninterpreted record of behavior. The announcement describes 165 activities, while the data dictionary lists 332 candidate O*NET IWAs and says rare categories may be combined into “Other IWA.” The available materials do not explain how the 165 reported activities correspond to the 332 candidates. That difference should not be papered over: when replicating an analysis or comparing datasets, category coverage is itself a variable to check.
The Missing Categories in the U.S. Comparison Matter
OpenAI also compares the activity mix of ChatGPT work messages with the activities performed across U.S. occupations. The distributions differ substantially, but the announcement notes that about 43% of the employment-weighted task mix falls into “Other IWAs,” which are not shown individually. Examples include moving materials, operating vehicles, and maintaining medical records. These categories did not meet the threshold for consistent representation in the original U.S. ChatGPT sample.
The gap should not be read as proof that ChatGPT favors certain kinds of work or that some tasks are more likely to be replaced by AI. One side is classified messages from consumer ChatGPT; the other is an employment-weighted mix of occupational activities, with a large share folded into a broad category. The comparison can help researchers identify mismatches worth investigating, but it cannot rank tasks by their degree of automation or their impact on occupations.
Useful for Questions, Not Standalone Conclusions
The broader Signals dataset samples one million consumer ChatGPT messages per month and spans July 2024 through September 2026. The work-activity data excludes enterprise messages and Codex use, so it cannot represent all usage inside organizations or through coding tools. Account eligibility rules, classification thresholds, and the bundling of rare activities also shape which uses appear in the published results.
OpenAI says researchers do not read individual conversations. Messages are stripped of identifying information before automated classification, and published statistics suppress small groups and include statistical noise to protect privacy. This design allows analysis of aggregate patterns without exposing raw conversations, but privacy protections do not guarantee accurate labels or remove sample boundaries. For technical teams, the sound use is to treat the data as a signal for forming hypotheses and locating changes, then compare it with independent occupational and employment measures. Message shares alone cannot establish rising productivity, disappearing jobs, or a reorganization of work.