人们如何使用ChatGPT

Aaron Chatterji1,2Tom Cunningham1David Deming3Zo¨e Hitzig1,3Christopher Ong1,3Carl Shan1Kevin Wadman1 1OpenAI2Duke University3Harvard University September 15, 2025 Abstract Despite the rapid adoption of LLM chatbots, little is known about how they are used.Wedocument the growth of ChatGPT’s consumer product from its launch in November 2022 throughJuly 2025, when it had been adopted by around 10% of the world’s adult population.Earlyadopters were disproportionately male but the gender gap has narrowed dramatically, and we findhigher growth rates in lower-income countries. Using a privacy-preserving automated pipeline, weclassify usage patterns within a representative sample of ChatGPT conversations. We find steadygrowth in work-related messages but even faster growth in non-work-related messages, which havegrown from 53% to more than 70% of all usage. Work usage is more common for educated usersin highly-paid professional occupations. We classify messages by conversation topic and find that“Practical Guidance,” “Seeking Information,” and “Writing” are the three most common topicsand collectively account for nearly 80% of all conversations.Writing dominates work-relatedtasks, highlighting chatbots’ unique ability to generate digital outputs compared to traditionalsearch engines. Computer programming and self-expression both represent relatively small sharesof use. Overall, we find that ChatGPT provides economic value through decision support, whichis especially important in knowledge-intensive jobs. 1Introduction ChatGPT launched in November 2022. By July 2025, 18 billion messages were being sent each weekby 700 million users, representing around 10% of the global adult population.1For a new technology,this speed of global diffusion has no precedent (Bick et al., 2024). This paper studies consumer usage of ChatGPT, the first mass-market chatbot and likely thelargest.2ChatGPT is based on a Large Language Model (LLM), a type of Artificial Intelligence (AI)developed over the last decade and generally considered to represent an acceleration in AI capabilities.3 The sudden growth in LLM abilities and adoption has intensified interest in the effects of artificialintelligence on economic growth (Acemoglu, 2024; Korinek and Suh, 2024); employment (Eloundouet al., 2025); and society (Kulveit et al., 2025). However, despite the rapid adoption of LLMs, thereis limited public information on how they are used. A number of surveys have measured self-reportedadoption of LLMs (Bick et al., 2024; Pew Research Center, 2025); however there are reasons to expectbias in self-reports (Ling and Imas, 2025), and none of these papers have been able to directly trackthe quantity or nature of chatbot conversations. Two recent papers do report statistics on chatbot conversations, classified in a variety of ways(Handa et al., 2025; Tomlinson et al., 2025). We build on this work in several respects. First, the poolof users on ChatGPT is far larger, meaning we expect our data to be a closer approximation to theaverage chatbot user.4Second, we use automated classifiers to report on the types of messages thatusers send using new classification taxonomies relative to the existing literature. Third, we report thediffusion of chatbot use across populations and the growth of different types of usage within cohorts.Fourth, we use a secure data clean room protocol to analyze aggregated employment and educationcategories for a sample of our users, lending new insights about differences in the types of messagessent by different groups while protecting user privacy. Our primary sample is a random selection of messages sent to ChatGPT on consumer plans (Free,Plus, Pro) between May 2024 and June 2025.5Messages from the user to chatbot are classifiedautomatically using a number of different taxonomies: whether the message is used for paid work,the topic of conversation, and the type of interaction (asking, doing, or expressing), and the O*NETtask the user is performing. Each taxonomy is defined in a prompt passed to an LLM, allowing us toclassify messages without any human seeing them. We give the text of most prompts in Appendix Aalong with details about how the prompts were validated in Appendix B.6The classification pipeline isprotected by a series of privacy measures, detailed below, to ensure no leakage of sensitive informationduring the automated analysis.In a secure data clean room, we relate taxonomies of messages toaggregated employment and education categories. Table 1 shows the growth in total message volume for work and non-work usage. Both types of messages have grown continuously, but non-work messages have grown faster and now represent morethan 70% of all consumer ChatGPT messages. While most economic analysis of AI has focused on itsimpact on productivity in paid work, the impact on activity outside of work (home production) is on asimilar scale and possibly larger. The decrease in the share of work-related messages is prima