Digital Economy

This is how we really use ChatGPT, an OpenAI report explains

16 September 2025

4' min read

Approximately 1.5 million conversations ended up under the magnifying glass to understand how people interact with the world's most popular generative artificial intelligence chatbot (around 700 million active users on a weekly basis): this is the essence of the National Bureau of Economic Research (NBER) working paper produced by OpenAI's economic research team in collaboration with Duke University and Harvard University, a study aimed precisely at tracing (completely anonymously and by means of automatic classifiers applied to data without personal identifiers) the evolution of consumers' relationship with ChatGPT from its launch three years ago to today. In short, this is a more complete picture than ever of the spread and usage habits of the tool, which confirms its enormous popularity on a planetary scale, with the volume of daily messages having risen from 451 million to over 2.6 billion (about 18 billion every week), thanks to a particularly strong increase recorded in low- and middle-income countries.

The scenario data and the democratic nature of the instrument

Assuming that a digital 'safe room' was used to collect and process the data on employment and education of the surveyed users, where aggregate statistical results were generated, the report highlighted how demographic and gender gaps are narrowing, with adoption widening from its beginnings across different consumer categories, and how most conversations concern everyday tasks such as seeking information and practical guidance, both in the personal and professional spheres, for the benefit of productivity at work and the improvement of one's personal life. OpenAI, in this sense, emphasised a concept it holds dear (for well-known reasons), namely the democratic nature of access to artificial intelligence, a technology (reads the note accompanying the study) 'available to people to unlock their potential and shape their future'.

Who uses it: growing female users The 'democratisation' effect, the report confirms, is reflected in the numbers measuring the ChatGPT usage gaps, which are now significantly reduced compared to the recent past. In January 2024, of the users with names classifiable as male or female, 37% clearly traced back to members of the fairer sex; by July of this year, this share had risen to 52%. The horizontal accessibility of GenAI irrespective of people's economic status, on the other hand, is demonstrated by other indicators, namely the fact that, by the end of May 2025, the growth rates

ChatGPT adoption in lower-income countries was more than four times higher than in higher-income nations. As for the profile of users, while the under-26 class is the most represented in terms of messages exchanged with the chatbot, the percentage of those using the tool for work is also clearly growing, especially among university graduates and in high-income professional sectors.

How usage is evolving A figure reflecting a clear trend in the use of ChatGPT is certainly that which photographs the incidence of 'prompts' sent to the machine for personal purposes, which have distanced themselves in percentage terms from professional and work-related ones, rising from 53% in June last year to 73% last June. How can this progressive evolution be explained? OpenAI interprets these numbers as concrete proof of its creature's dual role as a productivity tool and as a generator of value for people's daily lives. Three out of four conversations, in general, focus on practical guidance, information seeking and writing, with the latter being the most common work task, while programming and personal expression remain niche activities. The authors of the study classified usage patterns into three categories - 'Asking', 'Doing' and 'Expressing' - and found that about half of all messages (49% to be precise) concerned 'Asking', a sign of a growing consideration of ChatGPT as an advisor and not just a tool for completing tasks. 'Doing' covers 40 per cent of usage (about a third is for professional tasks) and includes interactions geared towards word processing and planning or scheduling, as opposed to using Gen AI to generate output or complete practical tasks. 'Expressing', finally, captures the remaining 11 per cent and covers interactions that do not fall into the first two categories and involve personal reflection, exploration and play.

Conversation topics: less writing, more research

A further interesting classification made by the experts finally concerns the skills that users are requesting from ChatGPT, grouped into seven thematic clusters. The three most common conversation topics are 'Practical Guidance', 'Seeking Information' and 'Writing', which together account for about 77 per cent of all conversations with AI. The former includes requests for tutoring, personalised advice and creative ideas (e.g. practical advice on everyday activities, from physical training to cooking) and accounts for about 29% of all conversations. Writing, the most popular category in July, dropped from 36 per cent to 24 per cent, but remains a main focus in professional interactions, with more than 40 per cent of the work-related messages recorded in July 2025 related to text creation: the majority of ChatGPT conversations related to 'Writing' (around

two-thirds), however, consists of requests to edit documents provided by users, rather than for the creation of new content. The search for information, on the other hand, grew in the same period from 14% to 24%, gathering queries ranging from current affairs to products to purchase, and confirming the increasingly important role of Gen AI compared to traditional search engines. Lastly, a final piece of data illustrates the existing relationship between users and OpenAI's generative AI as far as programming is concerned: coding, which is often associated with artificial intelligence, accounts for only 4.2% of conversations, and according to the authors of the study, this figure has a logical explanation: thanks to autonomous programming agents, the use of large language models to write software has grown very rapidly via APIs, thus outside the scope of ChatGPT.