How People Use ChatGPT

Key Findings

• — By July 2025, 18 billion messages a week were sent by 700 million users. Daily consumer messages (7-day averages) rose from 451 million in June 2024 to 2,627 million in June 2025, and non-work messages grew from 53% to 73% of the total.
• — Practical Guidance, Seeking Information and Writing together account for about 77% of conversations; Practical Guidance, which includes the Health, Fitness, Beauty, or Self-Care category, held at roughly 29% of usage.
• — 1.9% of messages are about Relationships and Personal Reflection and 0.4% about Games and Role Play; the authors contrast this with estimates that therapy or companionship is the leading use of generative AI.
• — About 49% of messages are Asking, 40% Doing and 11% Expressing (users expressing views or feelings without seeking information or action).
• — Users who self-report an age under 18, opted-out users, logged-out users and deleted or banned accounts are excluded from the sample.

Methodology Notes

Automated classification of a sample of consumer ChatGPT conversations; topic shares are computed from approximately 1.1 million sampled conversations from 2024-05-15 to 2025-06-26, reweighted to daily message volume (OpenAI's companion post describes the analysis as covering 1.5 million conversations). Messages are categorized by five LLM-based classifiers (prompts reproduced in an appendix; the topic classifier is modified from one used by internal OpenAI research teams), and the paper states that the datasets were built 'without any human ever reading the contents of a message'. Harvard IRB IRB25-0983. NBER working paper, not peer reviewed; six of the seven authors list an OpenAI affiliation, and code is available on request. The sample excludes self-reported minors, so the figures describe adult consumer use only. Date: NBER issue date is month-only (September 2025); 2025-09-15 is the date of OpenAI's companion post.

Sources