How People Actually Use ChatGPT: What the Data Shows

A summary of findings from the NBER working paper "How People Use ChatGPT" (Chatterji et al., NBER 34255, September 2025).

ChatGPT launched in November 2022. By July 2025, it had been adopted by around 10% of the world's adult population—roughly 700 million weekly active users sending more than 18 billion messages per week. For a new technology, that speed of global adoption has no real precedent.

Yet until recently, there was little solid evidence on how people use it: for work, for learning, for fun, or for something else. A new NBER working paper by researchers from Duke, Harvard, and OpenAI uses internal ChatGPT data and a privacy-preserving classification pipeline to answer exactly that. Here's what they found.

1. Most use is not for work—and non-work use is growing faster

Both work-related and non-work-related messages have grown over time, but non-work use has grown faster.

In June 2024, about 53% of messages were non-work and 47% work-related.
By June 2025, 73% were non-work and 27% work-related.

So the share of work use has fallen, and that's mostly because within each cohort of users, people are shifting toward more non-work use—not just because new users are more "personal use" heavy. The paper suggests the welfare gains from generative AI outside of work could be large; other work (e.g. Collis & Brynjolfsson, 2025) has estimated U.S. consumer surplus from generative AI at on the order of $97 billion per year.

2. Three topics dominate: Practical Guidance, Seeking Information, and Writing

The authors classify messages into conversation topics. Nearly 80% of all usage falls into three buckets:

Practical Guidance (~29% of messages)
Tutoring, how-to advice, creative ideation, health/fitness/beauty, and similar. This is highly customized to the user and can adapt with follow-up (e.g. "give me a workout plan for my level").
Seeking Information (~14% → ~24% over the period)
Factual lookups: people, events, products, recipes, etc. Functionally close to web search, but with answers that can be tailored and refined in conversation.
Writing (~36% → ~24% over the period)
Emails, documents, editing, critiquing, summarizing, translating. Writing is the single most common work use, accounting for about 40% of work-related messages in mid-2025.

Within Writing, about two-thirds of messages ask ChatGPT to modify or work with text the user provided (edit, critique, translate, summarize) rather than to create something entirely from scratch.

3. Coding and "companionship" are smaller than you might think

Two patterns that differ from some prior narratives:

Computer programming is only about 4.2% of all ChatGPT messages (compared to much higher shares in some other chatbot datasets, e.g. work-focused Claude usage).
Relationships / personal reflection and games / role play together are only about 2.3% of messages (1.9% + 0.4%). So "therapy" or "companionship" as the dominant use case is not what this consumer ChatGPT data shows.

4. Asking vs. Doing vs. Expressing

The authors introduce a simple intent taxonomy:

Asking (~49%): User wants information or advice to make better decisions—no direct "deliverable" from the model.
Doing (~40%): User wants the model to produce an output (e.g. draft email, code, table) that can be used in a process.
Expressing (~11%): User is sharing views or feelings, not clearly asking for information or a task.

Among work-related messages, Doing is higher (~56%), and a large share of that is Writing. At the same time, Asking has been growing faster than Doing over time, and Asking tends to get higher satisfaction (both from automated sentiment and from user feedback). The paper argues that a big part of ChatGPT's value at work is decision support—helping people think and choose—not only producing finished outputs.

5. What work are people doing with ChatGPT? (O*NET)

When the authors map messages to the U.S. Occupational Information Network (O*NET), about 81% of work-related messages line up with two broad kinds of activities:

Obtaining, documenting, and interpreting information
Making decisions, giving advice, solving problems, and thinking creatively

The same high-level activities (e.g. "Getting Information," "Making Decisions and Solving Problems") show up in the top five across many different occupations—from management and business to STEM to administrative and sales. So ChatGPT use at work looks broadly similar across jobs: lots of information-seeking and decision support, with writing and documentation as central.

6. Who uses ChatGPT?

Gender
Early on, a large majority of active users had typically masculine first names (~80%). By June 2025, the split was roughly even, with a slight tilt toward typically feminine first names. So the gender gap in who uses ChatGPT appears to have closed in this consumer data.

Age
Among users who report age, about half of all messages come from people 18–25. Work-related share increases with age (except for the oldest group). Usage has become less work-focused over time for every age group.

Geography
Adoption has grown especially fast in low- and middle-income countries over the last year (e.g. in the $10k–$40k GDP-per-capita range).

Education and occupation

More educated users (bachelor's and especially graduate) send a larger share of work-related messages and more Asking (advice/guidance) relative to Doing.
Professional and technical occupations (e.g. computer, management, engineering, science) have the highest work share and the highest share of Asking in work use. For example, Writing dominates in management and business (~52% of work messages); Technical Help is highest in computer-related jobs (~37% of work messages).

So ChatGPT at work is used most heavily by knowledge workers for information and decision support, plus writing and documentation.

7. How the study was done (and how privacy was protected)

The paper uses:

Growth data: Total daily message volumes and basic, self-reported demographics (e.g. age buckets, country).
Classified messages: A large random sample of ~1.1 million conversations (one message per conversation) from May 2024–June 2025, classified by automated LLM-based classifiers—no human ever reads the raw messages.
Employment/education: For a subset of users, aggregated employment and education categories analyzed in a secure Data Clean Room with strict aggregation rules (e.g. no cell with fewer than 100 users).

Messages are first stripped of personally identifiable information, then mapped to fixed labels (work vs. non-work, topic, intent, O*NET activities, etc.). Classifiers were validated on public conversation data (e.g. WildChat) to align with human judgment where possible.

Bottom line

Scale: ChatGPT has reached ~10% of global adults, with message volume growing by a factor of more than five in one year.
Use: Most consumer use is non-work, and non-work is growing faster. When it is work, it's mostly Practical Guidance, Seeking Information, and Writing, with Writing as the top work use.
Value at work: Use aligns with decision support and information handling—getting information, documenting it, and making decisions—especially for educated users in professional jobs. "Asking" is growing and is rated higher quality than "Doing" in many cases.
Demographics: Gender gaps in usage have narrowed to near parity; young adults account for a large share of messages; growth has been strong in lower- and middle-income countries; and work use is concentrated among the more educated and in professional/technical roles.

Overall, the paper suggests that ChatGPT's economic impact comes not only from automating tasks but from improving how people get information and make decisions—especially in knowledge-intensive work—while a large and growing share of value is in non-work use, from learning and advice to writing and search.

Source: Chatterji, A., Cunningham, T., Deming, D. J., Hitzig, Z., Ong, C., Shan, C. Y., & Wadman, K. (2025). "How People Use ChatGPT." NBER Working Paper 34255. http://www.nber.org/papers/w34255