The BaseResearchNBER
How People Use ChatGPT
An NBER working paper by mostly OpenAI-affiliated authors uses automated classifiers, with no person reading the text, on a random sample of consumer ChatGPT messages. The authors report that messages unrelated to work grew from 53% to 73% of daily volume between June 2024 and June 2025, and that writing is the most common work use.
Why it matters
The same thing, for different readers.
What it means for you
Most consumer ChatGPT use is advice, information and writing help, and over 70% of messages are not about work, so user counts alone say little about jobs.
For lawmakers
Adoption statistics should not be read as labor-market exposure: this data, which excludes business and enterprise plans, shows what people ask, not which tasks replace paid work.
For researchers
A privacy-preserving method for studying chatbot use at scale, validated on a public chat dataset; the work/non-work label is automated, and no pay or output outcomes are measured.
The question it raises
The authors note most economic analysis of AI looks at paid work, and argue its effect on activity outside work is of similar scale and possibly larger.
In their own words
Every quotation was found, word for word, in the text we saved from the source. That shows the words are there; it does not make them true, or prove the speaker meant them as quoted.
From the source’s own text.
“We find steady growth in work-related messages but even faster growth in non-work-related messages, which have grown from 53% to more than 70% of all usage.”
Found word for word in the source
“Writing is the most common use case at work, accounting for 40% of work-related messages on average in June 2025.”
Found word for word in the source
“Work usage is more common for educated users in highly-paid professional occupations.”
Found word for word in the source
“Our sample includes the three consumer plans (Free, Plus, or Pro). OpenAI also offers a variety of other ChatGPT plans (Business fka. Teams, Enterprise, Education), which we do not include in our sample.”
Found word for word in the source
“Our categories often apply to the user’s intention, rather than the text of the conversation, and as such we never directly observe the ground truth.”
Found word for word in the source
How the finding was reached
In our words, from the source. Weigh the finding by how it was made.
- What was done
- Language-model classifiers sorted a random sample of about 1.1 million consumer ChatGPT messages (May 2024 to June 2025) by work or non-work, topic, intent and O*NET work activity, with no person reading the messages. Daily message counts cover all consumer users since November 2022, and aggregated job and education categories for about 130,000 users came through a secure data clean room.
- What it found
- By July 2025 ChatGPT had more than 700 million weekly users sending 18 billion messages a week, about 10% of the world's adults. Non-work messages rose from 53% of daily volume in June 2024 (238 of 451 million messages) to 73% in June 2025 (1,911 of 2,627 million); work messages also grew, from 213 to 716 million. Practical guidance, seeking information and writing make up about 77% of conversations, and writing is about 40% of work messages. The work share is 37% for users without a bachelor's degree against 48% for those with graduate study, and 57% for computer-related occupations against 40% for non-professional ones (unadjusted).
- What it does not show
- Written partly by OpenAI employees from OpenAI's own data, so a company report even though NBER circulates it as an unreviewed working paper. Covers consumer plans only: Business, Enterprise and Education plans, logged-out users, under-18s and users who opted out of training are excluded. Whether a message is work-related is a language model's judgment from the message and its context, checked against 149 conversations from a public dataset (agreement with the human majority 0.83 on Cohen's kappa; humans among themselves 0.66). The education and occupation results cover a subset of users (about 130,000, and about 40,000 in the regressions) that may not be representative. The paper measures no productivity, pay, hours or job outcomes; its claim that ChatGPT raises worker output by supporting decisions is the authors' argument.
- Status
- The company’s own report about its own work
The record
- Questions
- What happens to work, income and the economy? · How is AI changing childhood, health and daily life?
- People
- Aaron Chatterji (Wrote it, Duke University Fuqua School of Business and OpenAI) · Thomas Cunningham (Wrote it, OpenAI) · David J. Deming (Wrote it, Harvard Kennedy School and NBER) · Zoe Hitzig (Wrote it, OpenAI and Harvard Society of Fellows) · Christopher Ong (Wrote it, Harvard University and OpenAI) · Carl Yan Shan (Wrote it, OpenAI) · Kevin Wadman (Wrote it, OpenAI)
- What we read
- We read the whole source.
- Found by
- A research pass: work, income and the economy, gathered October 13, 2026, via
research-work(tier 1: the source’s own channel or an official record) - Cite this
- NBER, “How People Use ChatGPT”, September 2025. Entry in the Ctrl AI Base, https://ctrlai.com/base/nber-2025-09-how-people-use-chatgpt
Our notesText: the NBER PDF (w34255.pdf), word for word the copy the research pass saved; read in full. The PDF says NBER working papers are not peer reviewed. Status company-report: the data are OpenAI's, six of the seven authors list an OpenAI affiliation, and the NBER page discloses that Chatterji (Duke and OpenAI) and Hitzig (OpenAI) are OpenAI employees. Quotations use the first author's id though the paper has seven authors. The paper is loose with itself in two places: the abstract and introduction put the three top topics at nearly 80% of use, Section 5.2 at about 77% and the conclusion at nearly 78%; the conclusion gives Expressing as 1% of messages where Section 5.3 says 11%. The entry uses the body numbers. User counts are accounts and may exceed people (the paper's footnote 20). Appendix A reproduces the study's classifier prompts; they are part of the paper and were not acted on. The IMF note and the World Bank report in this pass both cite this paper (the IMF for the speed of AI adoption, the World Bank for people using AI for practical advice, information and everyday tasks); the editor may want to relate them.