OpenAI Signals work activity data
OpenAI’s country-by-quarter dataset classifying work-related messages from personal ChatGPT plans by activity
What changed
On October 26, OpenAI released Signals work activity data that maps work-related messages from personal ChatGPT plans to a shared task classification across countries and quarters. Reading the dataset accurately means distinguishing the mix of activities in messages from actual work outcomes.
Key readingAI-assisted analysis
A shared task classification turns AI adoption into activity distributions that can be compared across countries and quarters. Linking these distributions with independent evidence on task outcomes and employment is necessary to explain changes in productivity and skills.
In focusTracking AI at work across 165 activities
This breakdown starts with a story focused on this topic. It is not a complete assessment.
The facts
01Mapping work messages to common activity categories
The published results cover nearly 100 countries and 165 intermediate work activities, or IWAs, organized by country and quarter from Q1 2025 through Q3 2026. Because the O*NET framework groups similar tasks from different occupations into shared categories, the data can be compared more directly with occupational statistics.
Read the full story02Average share of the ten most common activities declined
For each of the 74 countries observed in every quarter, the study calculated the share of classified work messages accounted for by its ten most common activities, then averaged those country shares. That average fell from 56.0% in Q1 2025 to 51.1% in Q3 2026.
Read the full story03Message uses and employment tasks differ in coverage
These work messages come from personal plans and exclude enterprise messages and Codex. Across mapped US occupations, about 43% of the employment-weighted task mix falls into Other IWAs that are not reported individually, such as moving objects, driving, and maintaining medical records. Those categories had not met the threshold for sustained representation in the US ChatGPT sample.
Read the full story04Classified aggregate statistics and privacy controls
OpenAI notes that researchers do not read individual conversations; instead, an automated classifier infers activity labels from de-identified messages. The published results apply minimum thresholds for small groups and add statistical noise. The v3.0 data dictionary update also revised historical sample sizes.
Read the full storyH/ACC insights
01Tracing how tasks enter workflows
H/ACC editorial interpretation: this dataset breaks AI adoption down into shifts across specific activities. Using a shared classification lets researchers identify which tasks enter conversations and compare them against occupational structures, education, or infrastructure data, grounding discussion in verifiable task categories.
Read the full storyWhat to watchAI-assisted analysis
- For the same countries and quarters within a single data version, how much would activity shares and cross-country differences change if the classifier were adjusted or Other IWA categories were included?
- When activity shares are aligned with independently measured completion rates, time spent, and employment structures, which changes persist and which are explained by sample composition or tool access?
Limits of this assessment
- The data covers only sampled messages from personal ChatGPT plans and excludes enterprise messages and Codex. It does not represent all workers, occupations, or global AI use, and the classification itself may contain errors.
- The 56.0% and 51.1% figures are averages across countries of the activity shares within each country's classified work messages. The 43% figure is a share of the employment-weighted task mix for mapped US occupations. None represents a share of users, a productivity gain or an unemployment rate. The source does not state whether the top ten activities remained a fixed set across periods.
- Version v3.0 recalculated historical data. Comparisons across periods should use the same version and account for small-group thresholds and statistical noise. Missing entries must not be treated as 0.
- H/ACC has not downloaded the CSV files to independently recalculate the statistics or evaluated classifier errors.