Consumer ChatGPT use skews toward personal tasks, a paper finds

The setup

In September 2025, OpenAI researchers and a group of outside economists, including Harvard's David Deming, released a National Bureau of Economic Research working paper, titled 'How People Use ChatGPT,' alongside an OpenAI summary post. The paper studies ChatGPT's consumer product specifically, tracking message volume and content from November 2022 through July 2025. It is explicitly a working paper, meaning it has not gone through peer review and its numbers could change on revision, a status the document states on its own cover page.

What the documents show

The authors classify a representative sample of consumer conversations with what they describe as a privacy-preserving automated pipeline, without a person reading the underlying chats. The headline finding is a shift in mix: personal, non-work conversations grew from 53% to more than 70% of total usage across the study period, even as work-related messages also grew in absolute terms. Practical guidance, seeking information, and writing together make up about 80% of all usage, and professional use concentrates among more educated workers in specialized roles. The paper frames this as evidence of consumer-side decision support rather than a claim about workplace deployment.

The friction

The paper's own scope note matters here: it studies ChatGPT's consumer product, not the separate enterprise or API products through which many companies actually deploy the technology at work, and it says nothing about how conversations translate into finished output or how often a suggestion was checked before use. Because the classification pipeline is automated, its category boundaries reflect the researchers' scheme rather than a worker's own description of the task, and a working paper's numbers carry less certainty than a published, peer-reviewed result.

What changed in the work

What the paper adds, compared with a marketing claim or a small user survey, is scale and a task taxonomy applied consistently across years of traffic, which lets a reader see the personal-versus-work mix shift over time rather than at a single snapshot. Treating that shift as proof of a broader trend in how all knowledge work gets done would go beyond what one product's consumer-side traffic can support, and the paper does not make that claim itself.

• Does the source study a consumer product, an enterprise product, or both, and does the quoted figure specify which?
• Is the finding from a peer-reviewed paper or a working paper still open to revision?
• Does 'usage' in this context mean messages sent, unique users, or something else entirely?

The paper's authors describe further analysis as ongoing, and OpenAI has said more releases on usage patterns are planned. Until a peer-reviewed version replaces it, this working paper is best read as a detailed but provisional description of one product's consumer side, not a settled account of how people use AI at work.