How is ChatGPT actually being used by real people?

Chatterji, Cunningham, Deming, Hitzig and co-authors — researchers from OpenAI and Harvard — classify a representative sample of ChatGPT conversations with a "privacy-preserving automated pipeline" to document how usage has shifted since launch. By July 2025, ChatGPT had been adopted by roughly 10% of the world's adult population. The central finding: "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." Classified by topic, "Practical Guidance," "Seeking Information," and "Writing" are the three most common, together accounting for "nearly 80% of all conversations." Within the work-related share, "Writing dominates work-related tasks," which the authors attribute to "chatbots' unique ability to generate digital outputs compared to traditional search engines." Work use itself skews toward "educated users in highly-paid professional occupations." This is measured behavior — actual conversations run through a classifier — not a self-report survey of users describing their own habits.

The authors frame the economic story as "decision support": ChatGPT's value to knowledge-intensive work comes less from retrieving facts than from producing a draft, memo, or analysis a person can act on, which is why writing — not programming (a "relatively small" share) or self-expression — carries the work-related load. The faster growth of non-work use suggests the tool is diffusing into everyday life (practical guidance, personal information-seeking) faster than it is becoming embedded in paid work, even as professional use stays concentrated among already-advantaged, educated, highly-paid users. Worth flagging: OpenAI co-authors a study of OpenAI's own product's usage, so the "decision support" framing and the economic-value claim come partly from the company with the clearest commercial interest in that conclusion.

This complicates Did ChatGPT cause Stack Overflow posting to decline?: that study shows chatbots displacing a Q&A retrieval venue, while this one finds work-related displacement running mostly through generation (writing) rather than lookup — two different substitution mechanisms operating at once. It also gives an internal-usage counterpart to How fast did LLM writing adoption actually spread?, whose external, document-level writing surge matches what this paper finds inside ChatGPT's own conversation logs: writing is the task chatbots are actually used for. Set against Does AI chatbot adoption change worker pay and hours?, the "decision support" claim here is the kind of value story a reorganizing-tasks-before-pay finding would need to cash out as earnings eventually, and hasn't yet, at least in Denmark. And against the single-organization case in How are national lab staff actually using generative AI?, this paper's population-scale lens shows the same copilot-like pattern — the model drafts, a professional decides — generalizing well beyond one lab's internal tool.

The excerpt gives no sample size, no detail on how the classifier was validated, and no causal test linking usage composition to the "economic value" it claims — "decision support" is an interpretive label the authors apply to a usage pattern, not a measured productivity outcome. It also doesn't say whether non-work use substitutes for or adds to work use, or whether the reported demographic narrowing (the gender gap, growth in lower-income countries) describes adoption rates or usage intensity. The better-supported finding is narrower but still useful: at global scale, and inside the company's own logs, ChatGPT functions more as a general-purpose writing and decision-support tool for a minority of professional users than as a search-replacement for the majority of its conversations, which skew personal.