OpenAI Signals maps how people use ChatGPT for work, with limits

What OpenAI released

OpenAI announced the dataset on October 26 [1]. It maps work-related consumer ChatGPT messages to 165 Intermediate Work Activities from O*NET and reports quarterly country-level patterns from Q1 2025 through Q3 2026.

The release covers nearly 100 countries. OpenAI presents the files as research data that people can combine with economic and occupational sources, and it publishes them under a Creative Commons Attribution 4.0 license.

The Signals individual-data page [2] says the underlying sample covers Free, Go, Plus and Pro accounts, which people generally manage for themselves. This population matters because workplace use through managed company accounts is outside the dataset.

What the 51.1% figure means

For the 74 countries observed in every quarter, OpenAI calculated the share of classified work messages represented by each country's ten most common activities. The average fell from 56.0% in Q1 2025 to 51.1% in Q3 2026, a decline of 4.9 percentage points.

A lower top-ten share means classified messages were spread across a broader set of activities. It does not show that workers became more specialized, that output increased or that ChatGPT caused a change in jobs. Country patterns could also reflect differences in occupations, product access, language coverage, sampling and classifier performance.

The unit is a message. One person can send many messages, and different tasks can require very different numbers of turns. Researchers should avoid converting message share directly into a share of users, work hours or economic value.

What the dataset excludes

Enterprise messages and Codex use are excluded. OpenAI says that omission probably underrepresents business and technical use cases. The sample also omits deleted accounts and people who opted out of allowing their data to be used for training, according to the published methodology.

The comparison with U.S. employment has another large boundary. OpenAI reports that about 43% of the employment-weighted task mix for mapped U.S. occupations falls into an Other category outside the individually reported activities. Physical work, vehicle operation and some record-maintenance activities are examples.

The dataset therefore gives a view of messages sent to one consumer service. It does not represent every task in an economy or every way workers use AI.

How to use the data without overclaiming

OpenAI's data methodology [3] documents the downloadable files, privacy approach and citation requirements. Automated classifiers assign work-relatedness and activity categories after identifying information is removed. Aggregate results include privacy protections.

O*NET defines Intermediate Work Activities [4] as a common layer between broad work activities and detailed occupation tasks. That common vocabulary helps researchers join the message data to occupational sources, but the join does not turn the ChatGPT sample into a labor-force survey.

A responsible analysis reports the population, unit, time range, effective sample rules and excluded products next to every chart. It also describes each classifier output as an estimate and checks whether the conclusion changes when categories or countries are removed.

A practical analysis workflow

• Write the question before downloading data. Decide whether you need activity mix, country comparison or change over time.
• Read the data dictionary and record the exact file version, extraction date, units and suppression rules.
• Choose a matched panel when comparing quarters so countries do not enter or leave the denominator unnoticed.
• Reproduce the published 56.0% and 51.1% top-ten calculation before creating a new measure.
• Plot sample support beside activity shares and flag estimates near publication thresholds.
• Run sensitivity checks with alternate category groupings and without the largest countries or activities.
• Describe associations in the result. Reserve causal claims for research designs that identify a cause.
• Publish code, transformed data and a citation to the versioned source when licensing and privacy rules allow it.

Limitations and open questions

• OpenAI operates ChatGPT and produced the dataset, so independent researchers should reproduce key calculations where possible.
• Automated labels can vary across languages, short prompts and messages that mix several work activities.
• Message counts do not measure answer quality, task completion, time saved or economic output.
• The public release cannot show how one person's use changes because the files contain privacy-preserving aggregates.
• Country differences do not explain whether occupation mix, product access, language, price or another factor produced the pattern.
• Future updates can revise classifications or coverage, so analyses should name the dataset version and date.

Frequently asked questions

• Does the dataset cover company ChatGPT accounts? No. It covers individual Free, Go, Plus and Pro accounts and excludes enterprise messages.
• Does it include Codex coding work? No. OpenAI explicitly excludes Codex use.
• Does 51.1% mean half of all work uses ChatGPT? No. It is the average share of classified work messages in each country's ten most common activities for a matched set of 74 countries.
• Can the data show which jobs will disappear? No. It records message categories and does not measure job loss, automation, productivity or wages.
• Can researchers reuse the files? Yes. OpenAI publishes Signals data under CC BY 4.0 and requires attribution.
• What is the safest first analysis? Reproduce one published measure with the documented population and denominator before testing a new question.