How people use ChatGPT: NBER 2025 key figures
How people use ChatGPT: economists from OpenAI, Harvard and Duke analyzed about 1.1 million conversations with ChatGPT and counted who uses it and why. By July 2025 about 10% of the world’s adults used ChatGPT every week, and more than 70% of messages were not related to work. Below are the main figures with page references, the method, the limits and the conclusions for a company.
Key figures
10%
of the world’s adults, about 700 million people, used ChatGPT every week by July 2025; they sent 18 billion messages a week
more than 70%
of messages are not related to work; in June 2024 the share was 53%, in June 2025 it was 73%
almost 80%
of conversations fall into three topics: practical guidance, seeking information and writing
40%
of work messages in June 2025 were about writing; about two thirds of them ask to edit the user’s text, not to write a new one
4,2%
of all ChatGPT messages are related to programming
56%
of work messages by July 2025 were requests to do a task; across all messages, 49% were questions, 40% requests to do something and 11% expressions
81%
of work messages relate to two kinds of work activity: getting and interpreting information; and making decisions, giving advice and solving problems
48%
of messages are work-related for users with a master’s degree or higher; 46% for those with a bachelor’s degree and 37% for those without one
The report in brief
- Published by
- NBER (US National Bureau of Economic Research); authors from OpenAI, Harvard and Duke
- Published on
- 15 September 2025
- Original title
- How People Use ChatGPT (NBER Working Paper 34255)
- Length and language
- 64 pages, in English
- Checked
- 16 October 2026
- Next edition
- not periodic; this is an NBER working paper, with no new versions as of 16 October 2026
“adopted by around 10% of the world’s adult population”NBER, abstract (PDF 2)
How the report was made
The authors took a random sample of about 1.1 million conversations of Free, Plus and Pro users from May 2024 to June 2025 (sec. 3, p. 4). Automatic classifiers based on language modelslarge language model. A large language model (LLM) is a neural network trained on a huge amount of text. It writes, summarizes and analyzes text on request. labeled the messages; no person read the texts. Occupation and education were studied on a subsample of about 130,000 users in a secure environment and only in aggregate form (sec. 3, pp. 4–9, PDF 6–11).
The authors give the share of the three main topics in different ways: almost 80% in the abstract, about 77% inside the paper (p. 14, PDF 16) and almost 78% (p. 36, PDF 38). This page uses the abstract’s wording.
What the report does not measure
Only personal plans were in the sample. Business, Enterprise and Education were not counted, nor were users under 18, those who opted out of data sharing and those working without signing in (footnote 5 on p. 1, table 1 on p. 2). The paper has not been peer reviewed by NBER.
Countries are shown only in groups by GDP per capita, with no figures for single countries. The study does not measure the effect on productivity or income.
What this means for a company in Russia
Russia in the report: the report has no data on Russia. The authors excluded from the country analysis the countries where ChatGPT is blocked (sec. 6.3, p. 27, PDF 29); Russia is not named in the text.
- Writing makes up about 40% of work messages, and people more often ask to edit their own text than to write a new one (p. 2). Start with editing emails, contracts and reports, where the effect shows in the first week. Ready scenarios are collected in the section on generative AI.
- 81% of work messages are about finding information and preparing decisions (p. 3), and programming takes only 4.2%. So the main demand from employees is for answers from the company’s documents. An internal AI assistant connected to your knowledge baseAI knowledge base. An AI knowledge base is a set of selected and prepared company documents that an AI assistant searches for answers: rules, instructions, a product catalog and answers to common questions. meets it.
- The study sees only personal plans, and 56% of work messages there are already requests to do a task (p. 3). If employees move work documents into personal accounts, the company does not control where the data goes. Set the allowed services and data types in an AI use policy.
What to compare it with
- Anthropic Economic Index 2026: an analysis of conversations with another service, also from the company’s internal data
- Microsoft: AI diffusiondiffusion model. A diffusion model is a neural network that creates an image, a video or a sound by removing noise from a random picture, step by step. 2026: the share of people using generative AI by country, estimated from telemetry, a different method
- ILO: generative AI and jobs 2025: which occupations are exposed to generative AI, a view from the labor market