OpenAI's 1.5M-conversation study finds most people come to ChatGPT to ask, not to have it do things

What the study actually measured

OpenAI, working with Harvard economist David Deming, published an NBER working paper analyzing 1.5 million ChatGPT conversations against a population of 700 million weekly active users. The company describes it as the most comprehensive study of actual consumer AI use released to date.

Two demographic findings anchor the paper. Among users with names classifiable as masculine or feminine, the share with typically feminine names rose from 37% in January 2024 to 52% by July 2025. And by May 2025, adoption growth in the lowest-income countries was running at over 4x the rate of the highest-income countries.

The methodology note matters for how much weight to put on the categories that follow: researchers did not read messages. Automated tools classified usage patterns without human review of content.

Asking beats Doing — and that cuts against the agent trend

The study sorts messages into three buckets: Asking (49%), Doing (40%), and Expressing (11%). OpenAI characterizes Asking as the fastest-growing and highest-rated category, describing people who 'value ChatGPT most as an advisor rather than only for task completion.'

That is a notable frontend signal. Much of the current product energy — agents, tool-calling, autonomous workflows — sits squarely in the Doing category: drafting text, planning, programming. But the plurality of real consumer traffic is people asking for guidance and information.

For anyone building a chat surface, this suggests the highest-frequency interaction is closer to a consultation than a command. An interface optimized purely for executing tasks may be tuned for the smaller of the two dominant modes.

Writing is the workhorse; coding and self-expression are the tail

The paper reports that three-quarters of conversations cluster around practical guidance, seeking information, and writing. Writing is called out as the most common work task, while coding and self-expression 'remain niche activities.'

That ordering is worth sitting with, because developer-facing narratives tend to foreground coding. In this consumer dataset it is a minority activity. The bulk of value shows up in mundane, text-heavy help — the kind of interaction a plain input box and a good response handle well.

The work/non-work split reinforces the point: roughly 30% of usage is work-related and 70% is not, with both growing. The interface is serving a general-purpose everyday tool, not primarily a professional workbench.

The implication: design the default surface for advice, not automation

OpenAI frames access to AI as something that 'should be treated as a basic right,' and the study leans into that framing with its emphasis on shrinking gaps and value that GDP 'fails to capture.'

About half of messages (49%) are 'Asking,' a growing and highly rated category that shows people value ChatGPT most as an advisor rather than only for task completion.Montana Labs

For applied teams, the concrete takeaway is a prioritization one. If half of real usage is decision support and advice-seeking, then the front door of a product should make asking cheap, legible, and trustworthy — clear responses, easy follow-ups, visible reasoning — before it invests in the heavier machinery of agents that automate the Doing bucket.

The caveat is scope: these are consumer plans and automated classifications, not enterprise workflows or human-read transcripts. But as a read on how the broad public actually reaches for a chat interface, the study points builders toward the advisor pattern as the default, with task execution as the specialized case rather than the reverse.