What the 2025 NBER working paper reveals about how people use ChatGPT.
ChatGPT launched in November 2022. By July 2025, according to the NBER working paper “How People Use ChatGPT” (Chatterji et al., 2025), it had about 700 million weekly users, around 10% of the world’s adults, sending 18 billion messages a week. That works out at roughly 29,000 messages a second.
The paper gives a detailed picture of how people use ChatGPT in real life: what they ask, what they create and how that is changing.
It draws on a sample of about 1.1 million conversations from May 2024 to June 2025, with one message taken from each and classified automatically (by GPT-5-mini) in a privacy-preserving pipeline, so no person read the messages. In the authors’ words, “this speed of global diffusion has no precedent”.
How fast it spread
Usage has grown faster in low- and middle-income countries than in high-income ones over the last year.
The gender gap has narrowed. In the first months after launch, about 80% of active users had typically masculine first names; by June 2025 that share was 48%. The researchers infer gender from first names, so these figures are estimates rather than a count of every user.
Among users who report their age, around 46% of messages come from people aged 18 to 25.
InvuAI’s take: adoption seems to follow usefulness in everyday life. When an assistant answers personal “how do I…” questions well, it stops feeling like a tool and starts feeling like a habit.
Work vs. life: the personal era of AI
Between June 2024 and June 2025, daily messages grew from about 451 million to 2.6 billion. Over the same year, the share of messages not related to work grew from 53% to 73%.
Work use is more common among educated users in highly paid professional occupations. The paper also cites economists Collis and Brynjolfsson, who estimate a consumer surplus of at least $97 billion in the US in 2024 from generative AI (WSJ).
Our reading: people aren’t only delegating work, they’re using ChatGPT to work through uncertainty.
What we talk about with ChatGPT
Nearly 80% of conversations fall into three topics. The shares below are for July 2025:
| Category | Share | Description |
|---|---|---|
| Practical Guidance | ~29% | How-to advice, tutoring, creative ideation |
| Seeking Information | ~24% | Specific information, products, recipes |
| Writing | ~24% | Drafting, editing, translating (down from 36% a year earlier) |
Technical help fell from 12% of messages in July 2024 to around 5% a year later; the authors suggest programming may have moved to the API and coding tools. Multimedia grew from 2% to just over 7%, with a spike in April 2025 when new image generation launched. Messages about relationships, personal reflection and role play are small, at about 2.4%.
These shifts suggest people use ChatGPT not just to produce content but to structure their thinking.
Asking, doing, expressing
The researchers classify messages as:
- Asking (49%): seeking information or advice,
- Doing (40%): asking for a task to be done, and
- Expressing (11%): sharing views or feelings.
By late June 2025, Asking had grown to 51.6%, while Doing fell to 34.6% and Expressing reached 13.8%.
At work, Doing still leads (about 56%), and nearly three-quarters of those Doing messages are Writing tasks. Across the board, Asking messages are rated higher quality, both by a satisfaction classifier and in direct user feedback.
Our reading: people are shifting from using AI as a worker to using it as a thinking partner.
When AI enters the job description
Mapped against O*NET’s Generalized Work Activities, nearly half of all messages (45%) fall under three activities about using information:
- Getting Information (19.3%)
- Interpreting the Meaning of Information for Others (13.1%)
- Documenting or Recording Information (12.8%)
Among work messages, the most common activities are Documenting or Recording Information (18.4%), Making Decisions and Solving Problems (14.9%) and Thinking Creatively (13.0%). Getting Information and Making Decisions and Solving Problems are in the top five in nearly every occupation.
Our reading: ChatGPT works mostly as a decision-support and information-handling tool, rather than a coder or content factory.
Who uses it for work
Share of messages that are work-related, by occupation (unadjusted):
| Occupation | Work share |
|---|---|
| Computer-related | 57% |
| Management and business | 50% |
| Engineering and science | 48% |
| Other professional | 44% |
In management and business, Writing accounts for 52% of work messages.
Quality and satisfaction
By July 2025, “good” interactions were more than four times as common as “bad” ones, up from about three times nine months earlier. Multimedia and Technical Help have the lowest good-to-bad ratios.
Interaction quality was judged automatically by GPT-5, from the next message the user sent.
What this all means
Our reading: ChatGPT has become a mainstream thinking companion, a place where decisions start, not just where text gets finished. InvuAI calls that cognitive leverage: people sharpening their own judgement through conversation rather than handing it over.
“Writing dominates work-related tasks,” the authors note, and they find that ChatGPT “provides economic value through decision support, which is especially important in knowledge-intensive jobs.”
InvuAI’s closing thought
We’re still only beginning to understand how people use ChatGPT, and how it may be changing how people think. In this data, use is moving from drafting and searching towards asking, reasoning and reflecting.
For brands, that matters: when people ask an assistant for advice before they decide, being part of that answer is worth tracking.
Sources and further reading
- NBER Working Paper No. 34255 (September 2025), Chatterji, Cunningham, Deming, Hitzig, Ong, Shan and Wadman: How People Use ChatGPT. All figures above are from this paper unless another source is linked.