48% of Uses Filtered Out by a Pro Filter, According to AI Observatory

A consortium of researchers from Stanford, MIT, and other institutions has published an observatory on the real-world uses of generative AI. Their analyses reveal significantly different practices depending on the models used, with a large portion of personal or sensitive exchanges that are not visible in corporate reports. The corpus used, which was collected on a voluntary basis, remains partial, and public decisions are still based on limited data.

Decisions Based on Limited Data, Researchers Warn

Anka Reuel believes that significant decisions are currently at risk of being made blindly, beyond the narratives provided by companies. She also asserts that stakeholders are making judgments about the risks and benefits of AI with very limited data. David Widder explains that there is no way to answer certain questions about overall usage because usage information is proprietary. Independent researchers like Reuel and Widder argue that AI companies rarely share their chat data for analysis and that their reports tend to highlight favorable outcomes. According to AI researchers, publishers only release the figures they wish to make public, and Anka Reuel emphasizes the lack of independent sources to corroborate this information. OpenAI did not respond to requests for comment.

Evolving Uses and Declining Sensitive Content

The analyses show that the use of AI has evolved over time and that responses from platforms vary. In the WildChat dataset, conversations have become longer and more complex, with more tokens on both the request and response sides and an increase in conversation turns. At the same time, the frequency of brief exchanges has risen, suggesting a greater search for companionship, while the explicit revelation by assistants of their chatbot status has decreased. Exchanges classified as sensitive uses, including sexual harassment and hate speech, have declined, a trend that may reflect more effective protections implemented by the platforms.

Grok, Gemini, ChatGPT, Anthropic: Distinct Usage Preferences

According to the observatory, behaviors differ markedly depending on the models, whether in terms of topics addressed, interaction style, or the likelihood of sensitive uses. Grok and Gemini are more frequently used for information retrieval; Grok stands out for news and politics, and it is also where misinformation tends to concentrate, a finding corroborated by other studies. xAI did not respond to a request for comment. Users are turning more to Anthropic for coding, to Gemini for social and role-playing uses, and to ChatGPT for homework assistance. There are also discrepancies between versions of the same model: conversations are shorter with ChatGPT in GPT-3.5 and longer and more iterative with GPT-4.

A Public Corpus: 24,521 Conversations from 5,000 Users

The aggregated corpus includes 24,521 conversations, representing 85,633 dialogue turns, derived from seven datasets collected through previous research. It comes from 5,000 users who interacted with 52 models, including ChatGPT, Gemini, Claude, and Grok, between 2023 and 2025. WildChat is among the largest and most detailed datasets. The voluntarily provided data is likely under-representative of sensitive uses and does not claim to reflect all usages. The scale remains incomparable to the internal resources of laboratories: the Anthropic Economic AI Index is based on 1 million conversations with Claude, and the latest report from OpenAI covers 1.5 million exchanges.

Applying the "Work" Filter Would Eliminate 48% of Exchanges

The often-cited Anthropic Economic Index targets Claude's uses related to work and productivity, excluding non-professional conversations. According to researchers, applying the same criteria to their own dataset would have led to the exclusion of 48% of the conversations. Within these discussions unrelated to work, themes related to health and relationships appear more frequently (44.2% compared to 31.2% according to Anthropic's analysis), as do adult or illegal topics (7.9% compared to 2.1%), harassment and hate (27.5% compared to 5.66%), as well as sexual content (16.7% compared to 2.4%). An OpenAI report published in 2025 also noted that among individuals, only 30% of ChatGPT uses were work-related. Several companies also state that they prioritize the analysis of professional uses.

Company Reactions and Next Steps for the Observatory

A representative from Anthropic indicated that the published research reflects the questions and interests of its teams and emphasized the importance of supporting independent research. The company has released separate posts on the use of Claude for support or companionship, as well as on the generation of CSAM. David Widder, an assistant professor at UT-Austin and not involved in the observatory, believes that a coherent global analysis is useful and reminds us, like Shayne Longpre who co-led the study after a PhD at the MIT Media Lab, that corporate reports do not cover everything. The data from the AI Observatory should be made available to researchers, with the team aiming to gradually expand its corpus. Anka Reuel hopes that AI companies will share their data with independent researchers, with confidentiality guarantees.