Auto-generated description: A chart displays various granular conversation topic shares represented by colored horizontal bars, each with labels indicating different topics and their percentage contributions to the total population.A few months ago, when I shared the work of Marc Zao-Sanders about how people use generative AI, I noted that his “a rigorous, expert-driven curation of public discourse, sourced primarily from Reddit forums” didn’t actually include a methodology.

In the last few days, OpenAI has released a paper in collaboration with scholars at Duke University and Harvard University which would suggest that people’s actual use of ChatGPT is… quite difference to the picture that Zao-Sanders gave. To me, that’s unsurprising given that he was sourcing his insights from Reddit, which skews young and male.Auto-generated description: A bar graph illustrates the difference in the share of topic prevalence in messages between users with typically masculine and feminine names across various categories, with feminine names favoring topics like social interaction and self-expression, while masculine names lean towards technical writing and persuading others. Interestingly, the demographics of ChatGPT users have changed markedly in the years since it was released. Apparently, in the first few months after it was made available, four-fifths of active users had “typically masculine first names” with that number dropping to less than half as of June 2025. Now, active users are slightly more likely to have typically feminine first names. It seems like different genders use generative AI differently, too (Fig.19).Finally, it’s telling that, as the use of ChatGPT grew five-fold from June 2024 to June 2025 the share of “non-work” uses rose from 53% to 73%. It’s definitely interesting times. Don’t mind me, I’m off to re-watch Her (2013).

First, we show evidence that the gender gap in ChatGPT usage has likely narrowed considerably over time, and may have closed completely. In the few months after ChatGPT was released about 80% of active users had typically masculine first names. However, that number declined to 48% as of June 2025, with active users slightly more likely to have typically feminine first names. Second, we find that nearly half of all messages sent by adults were sent by users under the age of 26, although age gaps have narrowed somewhat in recent months. Third, we find that ChatGPT usage has grown relatively faster in low- and middle-income countries over the last year. Fourth, we find that educated users and users in highly-paid professional occupations are substantially more likely to use ChatGPT for work.

We introduce a new taxonomy to classify messages according to the kind of output the user is seeking, using a simple rubric that we call Asking, Doing, or Expressing. Asking is when the user is seeking information or clarification to inform a decision, corresponding to problem-solving models of knowledge work… Doing is when the user wants to produce some output or perform a particular task, corresponding to classic task-based models of work… Expressing is when the user is expressing views or feelings but not seeking any information or action. We estimate that about 49% of messages are Asking, 40% are Doing, and 11% are Expressing. However, as of July 2025 about 56% of work-related messages are classified as Doing (e.g., performing job tasks), and nearly three-quarters of those are Writing tasks. The relative frequency of writing-related conversations is notable for two reasons. First, writing is a task that is common to nearly all white-collar jobs, and good written communication skills are among the top “soft” skills demanded by employers (National Association of Colleges and Employers, 2024). Second, one distinctive feature of generative AI, relative to other information technologies, is its ability to produce long-form outputs such as writing and software code.

We also map message content to work activities using the Occupational Information Network (O*NET), a survey of job characteristics supported by the U.S. Department of Labor. We find that about 58% of work-related messages are associated with two broad work activities: 1) obtaining, documenting, and interpreting information; and 2) making decisions, giving advice, solving problems, and thinking creatively. Additionally, we find that the work activities associated with ChatGPT usage are highly similar across very different kinds of occupations. For example, the work activities Getting Information and Making Decisions and Solving Problems are in the top five of message frequency in nearly all occupations, ranging from management and business to STEM to administrative and sales occupations.

Overall, we find that information-seeking and decision support are the most common ChatGPT use cases in most jobs. This is consistent with the fact that almost half of all ChatGPT usage is either Practical Guidance or Seeking Information. We also show that Asking is growing faster than Doing, and that Asking messages are consistently rated as having higher quality both by a classifier that measures user satisfaction and from direct user feedback.

…We argue that ChatGPT likely improves worker output by providing decision support, which is especially important in knowledge-intensive jobs where better decision-making increases productivity (Deming, 2021; Caplin et al., 2023). This explains why Asking is relatively more common for educated users who are employed in highly-paid, professional occupations. Our findings are most consistent with Ide and Talamas (2025), who develop a model where AI agents can serve either as co-workers that produce output or as co-pilots that give advice and improve the productivity of human problem-solving.

Source & image: OpenAI