Reinventing work and learning with ChatGPT: what 1,5 million conversations reveal

In less than three years, ChatGPT has gone from a technological curiosity to an everyday tool for nearly one in ten adults worldwide. But what do its users actually do? In September 2025, OpenAI, for the first time, opened its usage data to economists and published, with the National Bureau of Economic Research, a study based on 1,5 million conversations. The results challenge several preconceived notions: professional use is not the majority, coding and emotional "companionship" are marginal, and the tool's value lies primarily in its decision-making support. This article summarizes the findings of this study, verifies the figures, and extends them with data published by OpenAI in 2026.

Article published in September 2025 and completely rewritten in September 2026. The first version summarized, without citing it, the study "How People Use ChatGPT" by Aaron Chatterji, Thomas Cunningham, David Deming, Zoe Hitzig, Christopher Ong, Carl Yan Shan and Kevin Wadman (OpenAI and NBER, September 2025); this version explicitly cites it, verifies the figures and supplements it with the updates published by OpenAI in May and August 2026.

A groundbreaking study: 1,5 million conversations, none read by a human

Until 2025, all that was known about ChatGPT usage came from self-reported surveys or field observations. The OpenAI study takes a different approach: it examines a sample of 1,5 million conversations from consumer offerings (Free, Plus, and Pro) between May 2024 and July 2025. No messages were read by a human. Automatic classifiers categorized each message into a topic family, an intent, and, for professional uses, a work activity from the American O*NET classification. Demographic data (age, gender, country) is derived in aggregate and anonymized form.

Two points should be noted when reading this study. First, the study was conducted by the publisher itself, even though it involved renowned academics (David Deming at Harvard, Aaron Chatterji at Duke) and published its methodology in an NBER working paper. Second, it does not cover Enterprise, Team, or Edu accounts, nor the API, nor coding tools: the actual proportion of professional use within the entire ecosystem is therefore underestimated. These reservations aside, the picture it paints is the most accurate available.

An unprecedented adoption, and one that is becoming increasingly feminized.

By July 2025, ChatGPT had 700 million weekly active users, roughly 10% of the world's adult population, and processed 2,5 billion messages per day, nearly 29,000 per second. No other consumer technology had achieved such widespread adoption in such a short time. And the growth wasn't concentrated in wealthy countries: by spring 2025, the adoption rate in the lowest-income countries was more than four times higher than in the highest-income countries. Where access to expertise, education, or guidance is scarce, the assistant fills a gap.

The user profile has also changed. In January 2024, only 37% of users had a typically female first name; by July 2025, this figure had risen to 52%. The tool, initially adopted by a male and tech-savvy audience, has reached gender parity. Young adults, however, remain the most active: those under 26 send 46% of the messages. Older age groups are also present, but their usage is more work-related.

Seven out of ten messages have nothing to do with work

This is the most counterintuitive result of the study. While generative AI is presented primarily as a productivity tool, the share of messages unrelated to work increased from 53% in June 2024 to 73% in June 2025. In other words, fewer than three out of ten messages on consumer accounts are for professional use.

This shift is not due to a change in the user base. The authors show that the same user cohorts have broadened their uses: those who initially discovered ChatGPT for work now also use it to organize their daily lives, stay informed, learn, or create. Professional use remains correlated with education level and income: among graduates with advanced degrees, 48% of messages are work-related, compared to 37% among users without a license. The authors draw an important economic conclusion: the value created by ChatGPT is not measured solely in productivity gains within companies, but also in "consumer surplus," meaning time saved, better-informed decisions, and easier access to knowledge in their personal lives.

For an organization, the implications are twofold. On the one hand, its employees are already training themselves at home, at their own pace and without structure. On the other hand, they are bringing back to the office skills acquired using personal tools, with the confidentiality risks we described in our article on the Shadow AIThe question is no longer whether teams will use generative AI, but within what framework.

Practical advice, information, writing: three out of four uses

By July 2025, three categories accounted for nearly 80% of conversations: practical advice (29%), information retrieval (24%), and writing (24%). Practical advice encompasses tutoring, personalized guidance ("How do I organize my week?", "How do I structure a business plan?", "What exercises can I do to improve my English?"), creative ideas, and decision-making support. Information retrieval is akin to a conversational alternative to search engines, offering a capacity for synthesis and contextualization that search engines lack; this category is growing the fastest, to the point of catching up with writing, whose share is declining.

Writing remains the primary professional use, accounting for 40% of work-related messages across all sectors. One detail sheds light on the nature of this usage: approximately two-thirds of writing messages request modifications to a text provided by the user (correcting, rewriting, summarizing, critiquing, or translating it) rather than the creation of a text from scratch. ChatGPT acts more as a proofreader and editor than as a writer.

Conversely, two highly publicized uses turn out to be marginal. Computer programming accounts for only 4,2% of messages, and self-expression 2,4%, of which 1,9% is for relationships and personal reflection and 0,4% for games and role-playing. Technical support represents 5% and multimedia (image generation and analysis) 7%. The image of ChatGPT being used primarily for coding or companionship does not hold up to scrutiny, at least not for public accounts.

Asking, doing, expressing: the grammar of intentions

Beyond the topics themselves, the study categorizes each message according to the user's intention. "Asking": seeking information, advice, or an explanation to make a better decision. "Doing": obtaining a result, a text, a table, a program, or a completed task. "Expressing": sharing a thought or feeling without expecting a deliverable. Across the entire period, the first category represents 49% of messages, the second 40%, and the third 11%.

Two observations warrant attention. First, "ask" messages increased faster than "do" messages over the year studied: ChatGPT initially established itself as an advisor and conversational search engine before being perceived as a production automation tool. Second, these are also the highest-rated messages: user ratings (thumbs up or down) are significantly more favorable for requests for advice or explanations than for requests for execution. Perceived quality is higher when the tool informs a decision than when it produces a deliverable. In the workplace, however, the proportion is reversed, and "do" messages prevail.

At work: a decision-making aid rather than an automaton

For work-related messages, the authors used the O*NET nomenclature of work activities. Four activities predominate: documenting and recording information (18,4% of work-related messages), making decisions and solving problems (14,9%), thinking creatively (13,0%), and working with computers (10,8%). When these categories are grouped together, the core of work-related use consists of obtaining information, interpreting it, and using it to make decisions. This is the central thesis of the paper: ChatGPT creates economic value primarily as a decision support tool, particularly in knowledge-intensive professions.

Another key takeaway is the universality of these applications. Whether you're a manager, salesperson, IT specialist, teacher, or healthcare professional, you ask the tool to do the same things: synthesize information, document it, compare options, and formulate a recommendation. Generative AI doesn't create new professional silos; it disseminates a common practice across all roles, each applying it to its own specific problems. This explains why training in the fundamentals (defining a problem, providing context, iterating, and verifying) is transferable from one profession to another, while the use cases themselves must remain specific.

What has changed since then: the 2026 data

The study captures the summer of 2025. Since then, OpenAI has published two quarterly updates that extend the series and refine some of its conclusions. Regarding the user base, ChatGPT surpassed 900 million weekly active users in February 2026, and OpenAI now speaks of "more than one billion people." The Q1 2026 update, published in May, confirms the increasing number of female users (female first names are now the majority among users whose gender can be inferred) and notes that users over 35 are gaining ground in terms of message volume. As for work, the production of written and visual content remains the most common activity but is declining in favor of more specialized tasks.

The report for the second quarter of 2026, published in August under the title "From Asking to Doing," marks a more pronounced shift. While the 2025 report observed that "asking" was growing faster than "doing," OpenAI notes a year later that users are more than twice as likely to delegate a task to ChatGPT at work than outside of it, and that "doing" now prevails over "asking" in the office. Outside of work, seeking information and advice remains dominant. Multimedia is becoming the fastest-growing use case: 7,8% of messages worldwide, and over 10% in Brazil and Colombia. Finally, and this is the most interesting figure for a French reader, the share of messages sent by those over 35 has increased by 5 percentage points globally in one year, and by more than 10 points in France and the Czech Republic. The tool of young professionals is becoming the tool of choice for all generations in the workplace.

What these figures change for organizations

This data suggests a need to re-evaluate several assumptions about generative AI adoption programs. First, it's no longer about introducing newcomers to the field. Since seven out of ten messages are personal, most employees already use AI, often with a free account and on topics unrelated to their job. The challenge is to transfer these skills to a professional context: which tool, which data, which rules, and what verification methods are needed? The second assumption to reconsider concerns writing. Because two-thirds of requests involve existing text, the skill to develop is not to "make the AI ​​write," but to entrust it with a critical review, a reformulation for a specific audience, a faithful summary, and then to assess the result.

The third lesson concerns the nature of value. While ChatGPT primarily serves to facilitate decision-making, the core skill is problem formulation: describing the situation, clarifying constraints, asking for options rather than a single answer, and demanding opposing arguments. This is as much a management skill as a digital one, and it is developed using real-world cases rather than generic exercises. The fourth lesson comes from the 2026 data: the rise of the over-35 age group, particularly pronounced in France, means that experienced professionals are now the ones demanding it. An adoption program can no longer implicitly target "young people who already know"; it must start from the professional expertise of senior employees, which is precisely what makes decision support useful.

The final lesson concerns the trajectory itself: from requesting tasks, then from doing them, to delegating them. The shift observed in 2026 toward task execution at work foreshadows the next stage: agents entrusted with entire processes rather than isolated messages. Organizations that have structured conversational usage (identified use cases, authorized data, verification rules) are the ones that will be able to take this step without losing control. Others will discover agents as they discovered ChatGPT: through their employees, on personal accounts.

ChatGPT is no longer a niche innovation but an everyday tool, adopted by one in ten adults worldwide, primarily for personal use, and increasingly at work, for doing rather than asking. What the OpenAI study reveals, fundamentally, is an aspiration: to make better decisions, write better, understand better, with an interlocutor available at any time. It's up to organizations to make it a collective skill rather than a clandestine practice.

Sources. Aaron Chatterji, Thomas Cunningham, David Deming, Zoe Hitzig, Christopher Ong, Carl Yan Shan and Kevin Wadman, How People Use ChatGPT, NBER Working Paper 34255, September 2025, and its presentation by OpenAI, How people are using ChatGPT, for all 2024-2025 figures; OpenAI, How ChatGPT adoption broadened in early 2026May 11, 2026; OpenAI, From asking to doing: how the world is putting ChatGPT to work, August 6, 2026, for data from the second quarter of 2026.

To learn more: our analysis of the adoption of AI by Generation Z and our article on the chain-of-thought prompting.