People mostly use the ChatGPT chatbot to ask for advice, look up information, and get help with writing, and roughly three in four consumer messages have nothing to do with a job. A National Bureau of Economic Research working paper issued in September 2025 reports that ChatGPT reached 700 million weekly users and 18 billion messages a week by July 2025, and that practical guidance, information seeking, and writing made up about 77 percent of conversations.
Key takeaways
- The paper counts about 700 million weekly users by the end of July 2025, close to 10 percent of the world's adults.
- Non-work messages rose from 53 percent of consumer messages in June 2024 to 73 percent in June 2025.
- Computer programming was 4.2 percent of messages, and relationships and personal reflection was 1.9 percent.
- Writing was the largest work-related use at about 40 percent, and about two-thirds of writing requests asked for edits to text the user supplied.
- The data cover consumer plans only, and two of the seven authors work at OpenAI.
What the study measured
"How People Use ChatGPT" is NBER Working Paper 34255, by Aaron Chatterji, Thomas Cunningham, David J. Deming, Zoe Hitzig, Christopher Ong, Carl Yan Shan, and Kevin Wadman. It is a working paper, so it has not been through journal peer review. The main sample is about 1.1 million conversations from logged-in users on the Free, Plus, and Pro plans, sent between May 2024 and June 2025. Users who reported being under 18, logged-out users, people who opted out of sharing messages for model training, and deleted conversations were left out.
The authors say no member of the research team read message content. Messages were stripped of personal information by an automated filter and then sorted into categories by language model classifiers. Chatterji and Hitzig state that they are OpenAI employees, so the work is a company-affiliated analysis, and it describes usage rather than the quality of answers.
Growth, from launch to 700 million weekly users
ChatGPT was released as a research preview on November 30, 2022, and the paper reports more than one million registered users by December 5. Weekly active users on consumer plans passed 100 million after one year and reached almost 350 million after two. By the end of July 2025 the figure was above 700 million. The authors also report faster growth in lower-income countries and a gender gap among early adopters that has narrowed.
Daily message counts show where the growth came from. The table uses the paper's figures for the seven-day averages ending June 26 of each year; the multiples are simple arithmetic on those figures.
| Messages per day | June 2024 | June 2025 | Growth multiple |
|---|---|---|---|
| Work-related | 213 million | 716 million | about 3.4 |
| Non-work | 238 million | 1,911 million | about 8.0 |
| All consumer messages | 451 million | 2,627 million | about 5.8 |
The work and non-work split is itself an estimate, produced by classifying a random sample of each day's conversations. The paper notes that the shift came mainly from changing behavior within each group of users who signed up at the same time, not from newcomers who used it differently.
What people ask for
The three largest topics are Practical Guidance, Seeking Information, and Writing. Practical Guidance, which covers tutoring, how-to advice, and creative ideation, held steady at about 29 percent of usage. Seeking Information grew from 14 percent to 24 percent between July 2024 and July 2025, and Writing fell from 36 percent to 24 percent. Technical Help fell from 12 percent to about 5 percent, and Multimedia grew from 2 percent to just over 7 percent after image generation was added in 2025.
Two shares run against common assumptions. Computer programming accounted for 4.2 percent of messages, and the authors suggest a lot of coding moved to the API, code editors, and autonomous agents outside the ChatGPT app. Relationships and personal reflection accounted for 1.9 percent, and games and role play for 0.4 percent. About 10 percent of all messages were requests for tutoring or teaching.
Asking, doing, and expressing
The paper also sorts messages by intent. Across all messages, 49 percent were Asking (seeking information or advice), 40 percent were Doing (requests to produce output such as an email, a table, or code), and 11 percent were Expressing. Work-related messages tilt toward Doing, at nearly 56 percent, compared with 35 percent for Asking and 9 percent for Expressing. Writing at work is mostly editing: about two-thirds of Writing messages ask the chatbot to modify text the user supplied, such as critiquing, editing, or translating it, rather than to create text from scratch.
Limits of the numbers
Four limits apply when quoting the paper. The data cover consumer plans, so Business, Enterprise, Edu, and API traffic are outside the counts. Every category comes from an automated classifier, which the authors validated against human labels on a public dataset but which can still misfile messages. Frequency of a request says nothing about whether ChatGPT answered it correctly. And the study does not compare ChatGPT with other assistants: it contrasts its coding share with a published Claude study only to point out that the two samples differ.
How to read a chatbot usage statistic
- Check the population: which plans, which countries, which age groups, and which dates.
- Check the unit: users, weekly users, conversations, or messages.
- Check who classified the data and whether any human reviewed the labels.
- Check who wrote the report and whether the company that runs the product is involved.
- Treat usage as a description of what people do, not evidence that the output was right.
FAQ
Is ChatGPT mostly used for work?
No. The paper estimates that 73 percent of consumer messages in June 2025 were not work-related, up from 53 percent a year earlier.
How many people use ChatGPT each week?
The paper reports more than 700 million weekly active users on consumer plans by the end of July 2025. That is the authors' count and does not include every product tier.
Did anyone read users' conversations for the study?
According to the paper, no. Messages were scrubbed of personal information and classified by automated tools, and the authors say no one looked at message content.
Is the paper peer reviewed?
It is an NBER working paper from September 2025. Working papers circulate before journal review, so the findings are best treated as preliminary.
Related reading
- ChatGPT in teaching and learning
- What counts as the most advanced AI
- Building an AI voice assistant
- AI chat companions such as Chai
More in AI Tools and Industry.
Last verified: October 2026
Sources
- Chatterji A, Cunningham T, Deming DJ, Hitzig Z, Ong C, Shan CY, Wadman K. How People Use ChatGPT, NBER Working Paper 34255, National Bureau of Economic Research, September 2025