OpenAI's own study found most ChatGPT use isn't work

Start here

A person who assumes ChatGPT is mainly a workplace tool, because that is how it is marketed to businesses, is working from an incomplete picture. The largest analysis yet of actual consumer messages says otherwise. OpenAI reported a study, conducted as a National Bureau of Economic Research working paper with OpenAI's economic research team and Harvard economist David Deming, built on 1.5 million ChatGPT conversations.

What the documents say

This is a usage analysis of message content, not a survey of stated opinions: researchers state they did not read the underlying messages, instead using automated classifiers to sort conversation topics and preserve privacy. The NBER working paper lists Aaron Chatterji, Thomas Cunningham, David Deming, Zoe Hitzig, Christopher Ong, Carl Yan Shan and Kevin Wadman as authors and covers adoption from ChatGPT's November 2022 launch through July 2025. Both documents agree on the central split: about 70 percent of consumer use is non-work, up from 53 percent earlier in the period, with the remainder work-related. Three categories, practical guidance, seeking information, and writing, account for close to three-quarters of conversations, and the authors frame usage in three modes: 'Asking' for advice, 49 percent of messages; 'Doing,' drafting or task completion, 40 percent; and 'Expressing,' reflection with no clear task, 11 percent. The gender gap also narrowed sharply, with users holding typically feminine names rising from 37 percent in January 2024 to 52 percent by July 2025, and adoption grew fastest in lower-income countries.

Check this

A reader can check a claim about 'how people use AI' by asking whether it rests on message content at scale or on self-report; this study is the former, at a scale a survey could not match, though it covers consumer plans only, not the business, education or API traffic where use patterns likely differ.

What holds and what fails

The work and non-work split and topic breakdown hold as a description of ChatGPT's consumer plan traffic over the period studied, and the paper's stated privacy method is a real methodological safeguard worth naming when a company reports on user content. It fails as a claim about AI chatbots broadly: the study covers one vendor's consumer product, not competitors, and not the Enterprise, Team or API usage OpenAI itself excludes from the analysis. It is also, plainly, a vendor's own account of its own product, co-authored with an outside economist but published and framed by OpenAI, which is worth stating directly rather than treating as neutral.

• Note that a 'how people use AI' number from a vendor almost always describes that vendor's product, not AI use in general.
• Separate a study's stated scope, here consumer plans only, from the fuller picture across business and API use.
• Look for what a paper says about its own privacy method when a company reports on private message content.

The finding that most chatbot use is not about work is worth taking seriously precisely because it comes from message-level data rather than self-report, even as the source and its scope both belong in how the finding gets used.