Here’s What People Actually Do With ChatGPT | Harvard Business School AI Institute

Data-driven insights on the shift from cool tool to global utility

Listen to this article:

If you look at the headlines about tools like Claude Code and OpenAI Codex, or hear about programmers using LLMs to reach unprecedented levels of efficiency, you might think that everyone is starting to use AI to write computer code, turning prompts into programs and programmers into managers of machines. However, the recent NBER working paper “How People Use ChatGPT,” co-written by HBS AI Institute Associate David Deming, tells a very different story. If the dominant use case isn’t code generation, then the value proposition of AI looks quite different than many assume.

Key Insight: A Privacy-First Method for Sensitive Data at Scale

“No member of the research team ever saw the content of user messages.” [1]

The researchers utilized a massive, representative sample of approximately 1.1 million de-identified conversations between May 2024 and June 2025. Before any analysis, messages were stripped of personally identifiable information, then categorized by a series of LLM-based classifiers, everything from conversation topic and work-relatedness to the type of output the user was seeking. The researchers saw only the resulting classifications, never the underlying text. For employment and demographic analysis, the team worked through a Data Clean Room (DCR) protocol, a technical and procedural arrangement in which an external vendor held the data and researchers could only run pre-approved, aggregated queries.

Key Insight: ChatGPT Is Used Less for Work Than Many Assume

“[N]on-work messages have grown faster and now represent more than 70% of all consumer ChatGPT messages.” [2]

Non-work messages rose from 53% to 73% over the course of the period covered by the study. The authors classified messages into seven broad conversation topics, but three of them—Practical Guidance, Seeking Information, and Writing—together account for nearly 80% of all conversations. As of June 2025, computer programming accounted for just 4.2% of messages across all ChatGPT users. While writing accounted for 40% of work-related messages, about two-thirds of them involve modifying existing text by editing, critiquing, or summarizing, suggesting that for the average professional, ChatGPT is less of a ghostwriter and more of a sophisticated editor-in-chief.

Key Insight: What Users Want from AI

“We introduce a new taxonomy to classify messages according to the kind of output the user is seeking.” [3]

Beyond this analysis, the research team introduced a new three-part rubric designed to capture what kind of output a ChatGPT user is looking for. “Asking messages” seek information or advice to support a decision. “Doing messages” request that ChatGPT complete a task and produce a deliverable. “Expressing messages” are where users share feelings or engage in roleplay without seeking a specific action or information, and are thus conversational or emotional in nature. Across all messages, approximately 49% are Asking, 40% are Doing, and 11% are Expressing. Among work-related messages specifically, Doing rises to 56%, but Asking has grown faster than Doing.

Why This Matters

This research highlights that one of ChatGPT’s primary values in the workplace is decision support, a universal need that transcends specific job titles or industries. Whether in sales, engineering, or management, professionals are using AI to interpret information and solve problems more effectively. Business leaders and executives must recognize that AI adoption isn’t just a technical upgrade for the IT department, it’s a fundamental shift in how every employee processes information and executes decisions. Part of the strategic imperative then is to understand what questions your people are bringing to it, and whether those questions are the right ones.

Bonus

After seeing how people actually use ChatGPT, the next question is whether firms understand how quickly everyone else is moving. To learn how your AI strategy could be at risk due to the wrong information about the competitive landscape, check out Competing in the Dark.

References

[1] Chatterji, Aaron et al., “How People Use ChatGPT,” NBER Working Paper 34255 (2025): 5. https://doi.org/10.3386/w34255

[2] Chatterji et al., “How People Use ChatGPT,” 2.

[3] Chatterji et al., “How People Use ChatGPT,” 3.

Meet the Authors

Aaron Chatterji is Mark Burgess & Lisa Benson-Burgess Distinguished Professor of Business and Public Policy at Duke University.

Thomas Cunningham is a researcher at METR.

David J. Deming is the Danoff Dean of Harvard College, the William Henry Bloomberg Professor of Economics, the Isabelle and Scott Black Professor of Public Policy at Harvard Kennedy School, and Associate at the HBS AI Institute.