Americans use ChatGPT at work most often to make decisions

On October 26, 2026, OpenAI released Signals work-activity data mapping work-related messages on individual ChatGPT plans to 165 activities across nearly 100 countries, with quarterly observations from Q1 2025 through Q3 2026. The release includes a U.S. view built around O*NET, a framework for describing activities shared across occupations. It shows what users ask ChatGPT about at work, rather than whether their work gets done.

Making decisions and solving problems was the largest classified activity in the U.S. view for Q1 2026: it accounted for 21.6% of work-related requests, according to PPC Land’s analysis of the activity chart. Thinking creatively followed at 15.6%, and documenting or recording information accounted for 11.8%. The ranking describes the share of messages assigned to each activity, not the share of Americans who performed it.

What leads the U.S. work-request mix

The leading category covers requests associated with evaluating information, choosing among options and solving problems. A person could ask ChatGPT to compare approaches to a client brief, then send several follow-up messages about the same choice. Each message contributes to the activity mix; the dataset does not establish whether the person accepted an answer or made the decision afterward.

The mix changed over the year covered by the U.S. comparison. Decision making remained the largest named activity even though its share edged down from Q1 2025 to Q1 2026. Creative thinking and requests classified as providing consultation or advice gained share, while interacting with computers lost share. A change in share shows how requests were distributed within this sample; it does not reveal whether the total amount of computer-related work rose or fell.

O*NET activity labels cross job boundaries. A request classified as creative thinking could come from a writer developing an angle, a marketer exploring a campaign or someone in another occupation considering alternatives. The label identifies the activity inferred from a message, not the sender’s job title. That distinction matters when a category sounds closely associated with one profession but is used across many kinds of work.

How the categories map to a creative assignment

A marketing content assignment moves from alternative concepts toward a selected composition, illustrating distinct work-activity categories.

For writers, marketers and independent professionals, the categories describe different kinds of help that can arise during one assignment. Consider a hypothetical campaign brief:

A request for several possible angles could be classified as thinking creatively.
A request to weigh those angles against an audience or constraint could fall under making decisions and solving problems.
A request to record the chosen plan could fall under documenting or recording information.

This is a workflow illustration, not a sequence traced in the data. The released shares do not connect messages to a particular campaign, show that the requests came from the same person or identify a finished deliverable. They do show why decision support can lead the request ranking even in work commonly associated with producing copy: choosing an approach and producing wording are different activities.

OpenAI’s examples of work brought to ChatGPT include developing marketing content, preparing instructional materials, programming and analyzing business data. Those examples span planning, production and analysis. A marketer might ask for ideas, seek advice on a choice and then request language for the campaign, but the activity assigned to each message depends on what the classifier detects in that request. Documenting information is therefore narrower than all writing work, while creative thinking reaches beyond copy production.

Why the U.S. economy comparison looks different

Vehicle operation represents work included in the U.S. economy comparison but outside the individually reported ChatGPT activity categories.

The U.S. release also sets classified ChatGPT requests beside an employment-weighted picture of activities across the economy. A substantial part of the economy-wide task mix sits in an “Other IWA” group outside the individually reported activities. Examples include moving materials, operating vehicles and maintaining health or medical records; these activities did not meet the release’s threshold for consistent representation in the original U.S. ChatGPT sample.

The two sides of that comparison have different units. One describes the distribution of messages from people using individual ChatGPT plans for work. The other describes the distribution of activities across mapped U.S. occupations. Their shared O*NET framework makes the categories comparable, but a larger ChatGPT message share for planning or decisions does not mean ChatGPT carries out that share of work in the U.S. economy.

The comparison also helps explain what a text-based request chart makes visible. Choosing an approach, asking for information and drafting language can all be expressed in a message. Operating a vehicle or moving materials is represented differently in the economy-wide task mix. The gap between the distributions is useful context for the U.S. view, rather than a measure of which occupations have become more productive.

Methodology: what the message shares measure

The OpenAI Signals v3.0 data dictionary describes a global sample of one million consumer ChatGPT messages per month from accounts with a declared age of at least 18. Automated classifiers assign work-relatedness and activity labels after identifying information is removed; researchers do not read the raw messages for this analysis. The activity figures are weighted shares of classified messages, and the labels are inferred categories rather than users’ own descriptions of their work.

Included: Work-related messages in the covered individual ChatGPT accounts. A person using an individual account for a job-related request can appear in the sample.
Excluded: Enterprise messages, Codex use, deleted accounts and accounts whose users opted out of training. Those exclusions limit what the data can say about business and coding work across OpenAI’s products.
Unmeasured: Whether an answer helped, a task was completed, time was saved, output improved or staffing changed. The shares count messages, so frequent users can contribute more requests than occasional users.

Privacy protections add statistical noise and withhold results for small groups. Narrower slices of the sample can also vary more, and an automated classifier can assign a message to an activity that a person would describe differently. The U.S. ranking is consequently an estimate of the requests visible within the covered accounts. For anyone interpreting changes in work, it supplies a starting measure of what people seek help with; effects on output, wages or employment require evidence beyond these messages.