What 1 Million ChatGPT Conversations Reveal about Users

A new report challenges key assumptions about AI use

Posted on September 24, 2025 in Blog Posts

The tenor of the online discourse around generative AI has shifted over the past month with the release of OpenAI’s newest model, ChatGPT 5, in early August. Though hyped for months as a game changer, it received immediate criticism from both techies and casual users alike for failing to deliver the intelligence, use value, and romantic charisma(?) it promised. In the intervening weeks, this disappointing launch has reinvigorated AI skeptics, who see in GPT-5 evidence of a ceiling to AI’s development and the likelihood of an AI market bubble (which even OpenAI’s CEO, Sam Altman, hidebound on playing prophet one way or another, has started to acknowledge). Put simply, opinion among the commentariat has soured on AI, even if companies continue to invest big. 

Though the skeptics are certainly in power at the moment (and perhaps deservedly), we know that the media ecosystem prizes bold and definitive statements on issues that aren’t even close to being resolved, and that opinion can change at the drop of a hat. But we are due for a reevaluation of our assumptions about AI use, development, and prospects.  

To that end, I wanted to cover a new economic report created by OpenAI, Harvard, and Duke researchers, which uses a large set of granular data to ask the critical question, How are people actually using ChatGPT? 

Drawing on a dataset of more than 1 million conversations between users and ChatGPT, this report gives us a panoramic view into the sorts of prompts people put forward to generative AI. Much of the discussion around AI literacy has hinged on possible or presumed usages of AI in a higher ed context, and though this report doesn’t examine college students as a specific sample group, it nonetheless offers some interesting insights into the actual deployment of this technology. 

---

🔥 Stay up-to-date with LibTech Insights by signing up for our free newsletter. Just one weekly email with our new blog posts, top tech news stories, and other bonus content. Check out some posts from our archive:

• Starting a GenAI Community of Practice
• GenAI in Information Literacy Sessions: Strategies for One-Shot Instruction
• Building the Imagination Lab: A Manifesto for Libraries in the Intelligence Age

🌟 Register now for our free email micro-course on AI literacy!

---

🏖️ Nonwork usage is outpacing work-related usage. 

One of the key takeaways from this report is that, though work-related usages of ChatGPT continue to grow, they are wildly outpaced by nonwork-related usages, which have grown from 53 to 73 percent of all ChatGPT messages (p. 2). This finding raises two important questions: Given its ostensible economic promises, why isn’t work-related usage growing faster, and why is nonwork usage growing so much? 

The researchers focus almost solely on the data and seldom speculate throughout the report. They also don’t give much attention to nonwork usage, given their economic interest. But we might be able to make an educated guess for the second question based on the finding that, of all the messages studied, 36.7 percent—the largest chunk—are nonwork instances where users ask ChatGPT a question with the purpose of getting information or advice (p. 20).

Put crudely, it looks like people are using ChatGPT as an alternative to Google. The researchers acknowledge that AI’s more personalized responses make it more flexible than a search engine (p. 16), and perhaps that’s the key. I imagine that people have turned to ChatGPT out of fatigue with Google’s endless, SEO-written results and increasingly enshittified, ad-ridden user interface. ChatGPT has a much cleaner interface and offers much more straightforward responses. 

✍️ Yes, people are using ChatGPT for writing, but not how you’d think. 

Given the prized role of writing in educational environments, many academics might assume that when people use ChatGPT “for writing,” they’re using it specifically to generate new text from scratch—hence, the return of blue books. What this report finds, however, is that about two-thirds of all writing tasks have ChatGPT modify existing text, e.g., editing it for errors, adjusting the tone, or offering critiques, rather than generating new text (pp. 14, 16). 

This finding may not hold if we could separate students from among all users. Imaginably, professionals with writing experience and developed writing styles and usages may well find that AI doesn’t measure up to their specific standards, whereas students, who have less writing experience, may believe that AI exceeds theirs. Hence, the former might be more inclined to use ChatGPT to modify their existing text, and the latter might be more likely to use it to generate new text. But we don’t know.  

As we await more data, I think it behooves us to keep in mind that “writing” encompasses a range of activities. While we might imagine that students are asking ChatGPT to “write a seven-page essay on the Civil War,” for instance, they might well be using it to “make this email sound more professional.” 

💸 The economic value of generative AI may not be in its ability to “do” things. 

With the commercial release of ChatGPT in November 2022, workers began to fear that AI would soon replace their jobs. A task that might take a human several hours to complete could now be done in a few minutes. One worker with an AI assistant might soon replace five employees. This would save companies massive amounts of labor costs while also accelerating productivity. 

But the data in this report suggests a different picture.  

Only 19.5 percent of all messages asked ChatGPT to “do” things in a work context—that is, create new outputs, such as text, spreadsheets, and multimedia objects (p. 20). Interestingly, the report finds that the overall share of messages asking ChatGPT to “do” things has fallen from roughly 45 percent to 35 percent between May 2024 and June 2025 (p. 19). This statistic, however, includes both work and nonwork usages, so it isn’t clear whether the share of work-related requests, specifically, has declined. But it is a fair possibility, and given OpenAI’s business commitments, it might well be telling that the report doesn’t parse out this data. 

Leaving my editorializing to the side, how are people using ChatGPT for work, then? To get a more granular picture, researchers ran all the work-related messages through a different taxonomy based on common work activities, e.g., communicating with supervisors, scheduling events, and training others. They found that 57.9 percent of work-related messages fell into two broad categories “1) obtaining, documenting, and interpreting information; and 2) making decisions, giving advice, solving problems, and thinking creatively” (p. 20). In other words, people are using ChatGPT less as a replacement worker and more as an advisor and research assistant. 

Undoubtedly, this finding will spark further research because it cuts into the heart of generative AI’s foundational economic narrative: that it will do our work for us. Rather, the report puts forward an alternative narrative based on these findings: “ChatGPT likely improves worker output by providing decision support, which is especially important in knowledge-intensive jobs where productivity is increasing in the quality of decision-making” (p. 36).  

While maybe not more reassuring (I certainly don’t want my bosses to make crucial decisions based on what ChatGPT tells them), this finding should adjust our expectations for how we imagine students will use AI in the workforce and the sort of education they need to manage this relationship. 

🧠 Librarians are right to be concerned about hallucinations, misinformation, and bias. 

It follows from the above discussion that librarians are right to be concerned about hallucinations, misinformation, and bias. About half of all usages fall into the category of “asking” ChatGPT for specific information (e.g., the GDP of China, presidential term limits in France) and advice (e.g., custom gym programs, insurance selections). If people continue to turn to ChatGPT as an alternative to search engine queries and begin to rely on it for decision-making, then information quality is a key issue. As we continue to rethink AI and information literacy, these findings demand that we keep information analysis at the center of library teaching and instruction. 

Related Posts