Recently, a research report jointly conducted by OpenAI, Harvard University, and Duke University was released. Its title is unremarkable: "How People Use ChatGPT." But don't let the title fool you—it's arguably the most comprehensive "AI usage report" to date. Rather than relying on questionnaires or conjecture, the researchers directly sampled over one million real conversations, thoroughly analyzing the daily habits of ChatGPT users.
What makes this report unique is its real-world data and research: the sample covers the consumer version of ChatGPT from May 2024 to June 2025, with over 1.5 million conversations, processed using LLM automatic annotation and a privacy-preserving pipeline. In other words, the researchers never saw any of the original user messages, yet were able to unravel and compile a realistic picture of global AI conversation usage.
Image source: OpenAI
The report reveals that by July 2025, ChatGPT will have over 700 million weekly active users, representing approximately one-tenth of the world's adult population, and will send a staggering 18 billion messages per week. This scale is undoubtedly the largest AI application currently available and represents, to a large extent, the progress of AI application exploration.
This also leads to an intriguing question: when so many people use AI in their lives and work, what are they using it for? The answer may completely overturn your subjective impression.
From "automatic typewriter" to "decision-making plug-in", five key points of ChatGPT users
Many media outlets reporting on this research have summarized it with a single sentence: "Non-work usage has surged to over 70%," demonstrating the increasing prevalence of ChatGPT. However, it would be a shame to simply stop at this conclusion . The true value of this paper lies not in revealing what ChatGPT users are actually doing, but in revealing the emerging patterns of AI conversational use and how these patterns differ from our expectations.
The first stereotype that is overturned is that users are more likely to use AI to process rather than generate from scratch.
Research shows that programming-related conversations account for only 4.2%, while writing tasks account for 40% of workplace-related conversations. This is certainly because more AI programming work is concentrated in practical work scenarios such as IDEs and code editors, but it also demonstrates the importance of writing tasks.
Crucially, two-thirds of these writing tasks aren't generated from scratch, but rather involve "processing"—rewriting, polishing, translation, or optimizing logic. In other words, most users use ChatGPT not as an "automatic writer" but to help them polish existing content. This "rewriting" use case perfectly addresses the pain points of writing.
The second highlight is the distribution of intentions.
Image source: OpenAI
The paper also breaks down user motivations into three categories: asking, doing, and expressing. Overall, asking accounts for the highest percentage (51.6%), followed by doing (over 30%), and expressing (only 10%). However, in work-related scenarios, the situation reverses: doing jumps to 56%, with writing being the primary "doing" factor.
This also reveals that in daily life, people tend to view AI as a "pedagogy plus consultant," while at work, it's more like "productivity outsourcing." This division of labor actually hits the crux of product design: AI applications must simultaneously fulfill the dual roles of "decision support" and "direct output," rather than simply one or the other.
The third noteworthy detail is the change in the crowd portrait.
The paper points out that although the users of ChatGPT in the early stage of its launch were mainly male (accounting for about 80%), by mid-2025, the proportion of female users (52%) had equaled or even slightly surpassed that of male users, and there were obvious differences in usage needs.
Image source: OpenAI
Unsurprisingly, young people are more receptive to AI, with users under 26 contributing nearly half of all messages. However, surprisingly, the fastest growth in ChatGPT users is in low- and middle-income countries. This suggests that ChatGPT's user profile is increasingly approaching the average global population. For AI products, this presents challenges not only in terms of scale but also in terms of functionality and interaction design.
The fourth highlight is the correspondence with work activities: serving as a "decision-making plug-in."
The research team mapped the conversation content to the U.S. Department of Labor's ONET work activity classification and found that the three most common uses of ChatGPT were "decision-making and problem-solving," "recording information," and "creative thinking." This, to some extent, dispels the anxiety that "AI will take over our jobs": it's more like a "decision-making plug-in" for the human brain, helping you consider problems faster and more comprehensively.
Replacing humans is not the main storyline; enhancing human decision-making and creative realization is the truer story.
Image source: Lei Technology
Finally, there is a trend that is easily overlooked: user satisfaction.
The study used automated methods to label "good" and "bad" interactions and found that "good" interactions grew much faster than "bad" ones. By mid-2025, positive interactions were expected to be four times the number of negative ones. This demonstrates that the model's progress isn't limited to laboratory benchmarks; it's also being directly perceived by users in real-world conversations.
Taken together, these details paint a clear picture: AI is effectively becoming a "writing assistant," "life advisor," and "decision-making co-pilot" for users around the world. It's not replacing humans, but rather helping them to more smoothly and confidently process existing content, ideas, and decisions. This may be the true value of this research.
Behind user reports lies the real challenge of AI product design
On the surface, this report on ChatGPT provides a snapshot of user habits: non-work scenarios are growing faster, writing is a core workplace use case, and penetration is accelerating among young people and emerging markets. However, the truly noteworthy figures aren't the numbers themselves, but the larger question they reveal: how manufacturers and developers should rethink the form of AI applications.
The report clearly reveals that most writing tasks require AI to process and optimize existing texts rather than starting from scratch. This aligns with the usage habits of many developers and researchers: they don't expect AI to produce the ultimate answer, but rather to help them avoid the inefficient steps of revision, polishing, and patching.
A survey by Stack Overflow, a developer Q&A community, also pointed out that although most developers are using AI, the most common applications are not complex system development, but code snippet generation, error interpretation, and document writing.
For products, this means that the entrance design should be more in line with real needs . Instead of placing a "blank input box" in the middle of the interface, it is better to prioritize functions such as pasting, annotation, and difference comparison, so that AI can truly become an "enhancer" to fill fragmented needs rather than completely replace professional software.
Image source: Doubao
The most direct example may be Google's Nano Banana (Gemini 2.5 Flash image model). Many designers have expressed their opinions to professional design software manufacturers such as Adobe. In fact, Adobe and Figma also quickly announced the introduction of Nano Banana into software such as Photoshop.
Of course, different groups have vastly different expectations of AI: Novice users demand structured templates, voice choices, and step-by-step guidance. Experienced users, on the other hand, demand quick commands, customizable toolchains, and even deeply integrated APIs. In short, with the diversification of user profiles, applications that fail to provide a layered experience are likely to face a dilemma: either too complex or too superficial.
From a broader perspective, "process embedding," a principle repeatedly emphasized by many companies, is also a key element of large-scale AI adoption. The ChatGPT report also points out that frequent AI use does not necessarily equate to high trust. Users' willingness to rely on AI requires that its results can be verified, traced, and integrated into the organization's existing compliance systems. This requires that AI applications incorporate mechanisms such as data source annotation, version comparison, and result auditing from the outset, rather than as after-the-fact patches.
More importantly, don't just focus on showing off. Today, AI applications have moved from "showing off" to "refining." For developers and manufacturers, the key now isn't proving how complex the model can be written into code or articles, but rather how to enable users to use it faster, more securely, and more widely in real-world scenarios. The real competition has shifted from the model itself to user experience, trust, and universal access.
Final Thoughts
OpenAI's research provides a glimpse into the real-world applications of ChatGPT: using AI to write, acquire information, and make decisions. However, these applications aren't simply about numbers; they point to a more fundamental question: what exactly should AI products look like?
As the report reveals, at least at this stage, users view AI more as an enhancement than a replacement, demanding low-friction experiences rather than flashy models. Without continuous refinement of accessibility, interaction, trust, and accessibility, applications will struggle to truly integrate into users' daily lives. For everyone, AI has shifted from a technological competition to a test of design and user experience.
What's even more worth considering is that as hundreds of millions of people around the world engage in daily conversations with AI, we're witnessing a new kind of "human-computer interaction" gradually becoming a habit. How will this change everything? This is an open question facing everyone. The future direction of AI may not be determined by the technology itself, but by how people choose to use it.
This article comes from "Lei Technology" and is authorized to be published by 36Kr.
