NBER: ChatGPT hits 700M weekly, 70% of use non-work

A working paper out of the National Bureau of Economic Research this month puts a number on something the industry has argued about since ChatGPT launched, which is how much of it is actually work. The NBER paper, by OpenAI's Economic Research team and Harvard economist David Deming, reports that non-work-related messages have grown from 53% to more than 70% of all usage between November 2022 and July 2025.

The scale sits underneath that share. By July 2025 the authors estimate ChatGPT was reaching about 10% of the world's adult population, with 700 million weekly active users sending 18 billion messages per week. Early adopters skewed male, but the gender gap has reportedly narrowed, and growth rates are higher in lower-income countries than in the ones where the product first landed.

The categories inside that traffic are the part worth chewing on. Three buckets, Practical Guidance, Seeking Information, and Writing, collectively account for nearly 80% of all conversations. Writing dominates the work half of the pie at roughly 40% of work messages, and about two-thirds of those writing tasks involve modifying existing user-provided text rather than generating something from scratch. That is a smaller claim than the ghostwriter framing you tend to hear. It looks more like a very fast editor than a very fast author.

Why this matters if you are selling into enterprises: the vendor pitch that ChatGPT is a productivity tool bolted onto office work is now arguing against a dataset the authors describe as a representative sample of conversations. Consumer habit is the larger surface, and it is what pulls workers back into the tool during the workday.

The honest caveat is that this is a working paper drawing on OpenAI's own logs, so the classifier choices behind 'work' and 'non-work' are ones outside auditors cannot easily replicate. What the reporting doesn't give you is a country-by-country breakdown or a paid-versus-free split, both of which would change the strategic read. Still, the direction is what to watch, because the app people actually use is closer to a general-purpose assistant than a workplace copilot, and product roadmaps that ignore that are betting against their own users.