The OpenAI and Harvard economists found that only 4.2 percent of ChatGPT messages are related to computer programming.

“ChatGPT launched in November 2022. By July 2025, 18 billion messages were being sent each week by 700 million users, representing around 10 percent of the global adult population. For a new technology, this speed of global diffusion has no precedent.”

So begins a new paper by Harvard economist David Demin, OpenAI chief economist Ronnie Chatterji, and their coauthors that studies how consumers are using ChatGPT, the first mass-market chatbot. This National Bureau of Economic Research paper is timely and important: generative AI tools are seemingly everywhere, but there is little systematic evidence of how they are being used. (I highlighted a much-discussed New York Times article with anecdotes about AI usage here on the Corner last month.)

I was surprised by their finding that 73 percent of ChatGPT messages are used for non-work purposes. The authors of the study: “While most economic analysis of AI has focused on its impact on productivity in paid work, the impact on activity outside of work (home production) is on a similar scale and possibly larger.”

More:

Nearly 80% of all ChatGPT usage falls into three broad categories, which we call Practical Guidance, Seeking Information, and Writing. Practical Guidance is the most common use case and includes activities like tutoring and teaching, how-to advice about a variety of topics, and creative ideation. Seeking Information includes searching for information about people, current events, products, and recipes, and appears to be a very close substitute for web search. Writing includes the automated production of emails, documents and other communications, but also editing, critiquing, summarizing, and translating text provided by the user. Writing is the most common use case at work, accounting for 40% of work-related messages on average in June 2025. About two-thirds of all Writing messages ask ChatGPT to modify user text (editing, critiquing, translating, etc.) rather than creating new text from scratch. About 10% of all messages are requests for tutoring or teaching, suggesting that education is a key use case for ChatGPT.

Despite all the talk about ChatGPT replacing human coders, the OpenAI and Harvard economists found that only 4.2 percent of ChatGPT messages are related to computer programming.

The authors introduce a new taxonomy to get at how users are using ChatGPT:

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.

And:

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.

The authors argue that ChatGPT is providing the most economic value by providing decision support, and not by performing tasks.

Despite the fact that the minority share of ChatGPT usage is for work, I strongly suspect that, over time, OpenAI (with which I am doing some work) and other generative AI companies will have a substantial impact on the labor market. This will happen in part because new occupations will be created that are designed around the capability of these new technologies, and in part because the wealth these new technologies create will in turn lead consumers to demand new goods and services.

From my recent National Affairs article:

The AI revolution will create many opportunities that we cannot conceive today. Standing in the year 2024 and trying to predict the jobs of the future is no easier than standing in the year 1944 and trying to predict that the labor market of the future would contain systems analysts, circuit-layout designers, fiber scientists, and social-media managers. In fact, about 60% of jobs held by workers in 2018 had not been invented as of 1940. New occupations emerge in large part because technology advances, creating new goods and services that in turn require human workers to engage in new occupational tasks. Technological advances also make society wealthier, increasing the demand for goods and services — especially new goods and services — which in turn raises the demand for workers’ skills, talents, and efforts.

To illustrate, imagine trying to explain to the 19th-century classical economist David Ricardo the jobs of all the people who support Bruce Springsteen’s records and tours: sound engineers, digital editors, graphic designers, photographers, videographers, art directors, instrument technicians, social-media directors, marketing professionals, bookers, stage hands, sound directors, lighting engineers, body men, commercial-vehicle drivers, and, of course, the jobs of the members of the mighty E Street Band and Mr. Springsteen himself. These occupations and the tasks workers perform for them did not exist in Ricardo’s time because the technology that enables them had not been invented. They also did not exist because the wealth created by today’s technology had not been generated: Society in Ricardo’s day could not have afforded rock bands.

But the new paper by OpenAI and Harvard economists highlights the role in our non-work lives that AI tools can play — and, apparently, are already playing.