How Are We Really Using AI?

For the first time, major AI labs have released large-scale data on how their platforms are being used. Forget speculation; this is our first real look into the daily habits of millions of AI users. The findings from OpenAI (ChatGPT) and Anthropic (Claude) paint a picture of a technology becoming a general-purpose utility, but with vastly different applications depending on the platform.

Across both services, writing is a universal power-use [1, 2]. This includes everything from drafting emails and reports to editing and summarizing existing text. But here's what's surprising: 67% of ChatGPT writing tasks involve modifying existing text rather than generating from scratch [1]. Users aren't asking AI to replace their voice — they're asking it to refine it.

For ChatGPT, the most common use cases are Practical Guidance (28.8%), Seeking Information (24.4%), and Writing (23.9%) [1]. These three categories account for nearly 80% of all usage. Practical Guidance spans everything from workout routines to tax advice, while Seeking Information often mirrors traditional web search but with conversational follow-ups. Notably, ChatGPT processes 18 billion messages weekly from 700 million users — roughly 10% of the global adult population [1].

Claude's usage, on the other hand, leans more heavily into professional tasks. Computer and mathematical occupations account for 40.2% of usage, with coding dominating the conversation [2, 3]. Since Claude 3.7 Sonnet's launch, there's been a measurable increase in usage for education (+1.8%) and life/physical sciences (+0.5%), suggesting the model's improvements are opening new use cases [3]. The "extended thinking" mode introduced with Claude 3.7 is used primarily by computer science researchers (9.7% of their tasks) and software developers (8.4%), indicating highly technical users are leveraging advanced capabilities [3].

Figure 1: Nearly 80% of ChatGPT usage falls into three broad categories.

Figure 2: Coding and mathematical tasks represent the largest share of Claude.ai usage.

Co-pilot vs. Autopilot: The Universal Desire for an AI Partner

While the platforms have different strengths, the most powerful insight is how their usage patterns align. Both reports, using different language, find that users overwhelmingly prefer to collaborate with AI rather than simply offload tasks.

ChatGPT users "Ask" for advice and decision support more than they ask the AI to "Do" a task for them. Specifically, "Asking" makes up 49% of use, "Doing" comprises 40%, and "Expressing" (neither asking nor doing) accounts for 11% [1]. Crucially, Asking messages receive higher satisfaction ratings — a good-to-bad ratio of 4.45 compared to 3.67 for Doing messages [1]. This 21% higher satisfaction for collaborative queries suggests users derive more value from AI as a thinking partner than as a task executor.

Similarly, 57% of Claude's usage is "augmentative," with users and the model co-creating together, particularly in fields like copywriting where task iteration rates reach 58% [3]. The highest augmentation rates appear in community and social service occupations (75%), while even technical fields like computer and mathematical occupations show balanced augmentation (50%) versus automation (50%) [3].

This shared trend is a powerful statement. The dominant behavior is not replacement, but partnership. When ChatGPT usage is mapped to O*NET work activities, the top categories are "Getting Information" (19.3%), "Making Decisions and Solving Problems" (14.9%), and "Thinking Creatively" (13.0%) — all cognitive tasks that benefit from dialogue rather than delegation [1]. Users are seeking a co-pilot to enhance their own judgment, not an autopilot to take the wheel completely.

Figure 3: The majority of ChatGPT's usage is for personal, non-work-related tasks.

A Global Phenomenon: How AI Usage Varies Across Cultures

The analysis reveals that AI is not a monolith; its application reflects the distinct cultural and economic priorities of its users around the globe.

The Clio report found fascinating cultural nuances in how Claude is used. For example, conversations in Japanese are 5.9x more likely to discuss elder care compared to the baseline, reflecting Japan's aging society challenges [2]. Chinese users show a 4.4x higher rate of crime and mystery fiction writing, while Spanish-language conversations are 3.5x more likely to involve economic theory discussions [2]. These aren't minor variations — they represent fundamentally different use cases shaped by local contexts.

Additionally, the ChatGPT paper notes remarkable demographic shifts. Early adopters were 80% male (based on name analysis), but by June 2025, the gender split had reversed to majority female users [1]. Usage has grown relatively faster in low- and middle-income countries, with these regions showing higher month-over-month growth rates than wealthy nations [1]. This suggests that as access broadens, AI is being adapted to solve a much wider and more diverse set of problems than those prioritized in high-income tech hubs.

Age patterns are equally striking: 46% of ChatGPT messages come from users aged 18-25, but work-related usage increases with age — from 23% for users under 26 to peak around 35% for middle-aged users, before dropping to 16% for users over 66 [1]. This lifecycle pattern suggests AI serves different purposes at different life stages.

Reading Between the Lines: What This Data Doesn't Tell Us

While groundbreaking, this research provides an incomplete picture. First, this data is primarily from consumer tiers only and excludes the potentially massive amount of usage from business clients and API calls. ChatGPT's API likely processes billions more messages for enterprise customers, and Claude's enterprise usage remains entirely unmeasured [1, 2].

Second, these are snapshots in time in a field that is evolving incredibly fast. The data captures usage patterns for current model capabilities, but as models improve, usage patterns shift dramatically. For instance, ChatGPT's multimedia usage jumped from 2% to over 7% after new image generation capabilities were released in April 2025 [1].

Third, there's a critical selection bias: we only observe users who choose to use AI. The non-users — still the majority of the global population — remain invisible in these datasets. Their reasons for non-adoption, whether due to access, awareness, or aversion, could reshape our understanding of AI's societal impact.

Finally, the classification methods differ between studies. ChatGPT uses purpose-based categories while Claude maps to occupational tasks, making direct comparisons challenging. These reports are best seen not as final conclusions, but as the first important data points in an ongoing story.

Figure 4: A majority of interactions on Claude.ai are "augmentative," showing a collaborative pattern.

The Next Frontier: An "AI Trends" for Society?

This brings us to a compelling future possibility. Just as Google Trends offers a public window into what the world is searching for, a similar, aggregated system showing AI trends from all major chatbot providers could be transformative.

Anthropic's Clio system demonstrates the technical feasibility. Using hierarchical clustering and multiple privacy barriers, Clio can identify patterns across millions of conversations without any human seeing individual messages [2]. The system successfully identified coordinated abuse campaigns (like networks generating SEO spam) that would be invisible at the individual conversation level [2]. It also revealed safety classifier failures, finding that translations of explicit content were being under-flagged while Dungeons & Dragons combat descriptions were over-flagged [2].

Such a platform would provide unprecedented transparency for researchers, identify underserved needs in the market, and spark new startup opportunities focused on the most common or fastest-growing AI use cases. It could track the real-world impact of AI capabilities as they emerge — for instance, monitoring usage during elections (as Anthropic did for the 2024 US election) or after launching new features like computer use [2].

Creating such a platform would be a complex undertaking, requiring industry-wide coordination and robust privacy standards. But these initial reports demonstrate that the methodology is not only possible but essential for understanding AI's true societal impact. By providing a clear, data-driven view into how humanity is partnering with AI, it could help guide the technology's development in a more beneficial and responsive direction.

References

  1. Chatterji, A., Cunningham, T., Deming, D. J., Hitzig, Z., Ong, C., Shan, C. Y., & Wadman, K. (2025). How People Use ChatGPT (Working Paper No. 34255). National Bureau of Economic Research. https://www.nber.org/papers/w34255
  2. Tamkin, A., McCain, M., Handa, K., Durmus, E., Lovitt, L., Rathi, A., ... & Ganguli, D. (2024). Clio: Privacy-Preserving Insights into Real-World AI Use. arXiv preprint arXiv:2412.13678. https://arxiv.org/abs/2412.13678
  3. Anthropic. (2025, March 27). Anthropic Economic Index: Insights from Claude 3.7 Sonnet. https://www.anthropic.com/news/anthropic-economic-index-insights-from-claude-sonnet-3-7