Story
August 19, 2026
AI Observatory Says the Chatbot Story Is Bigger Than Company Reports
A new independent database of consented AI chats argues that corporate usage reports leave out much of the personal and sensitive activity shaping chatbot use. Its creators say the sample is limited, but valuable as a check on industry narratives.
The public picture of chatbot use has largely been drawn by the companies that run them. A new AI Observatory says that picture is too tidy — and that the messier reality includes companionship, health questions, sexual content and political misinformation.
The project was built as researchers confronted a basic gap: AI assistants have billions of users, yet evidence of how people actually talk to them remains “sparse and fragmented.” The Observatory aggregates anonymized, consented conversations from seven existing sources to create a public research resource.1
Its co-lead, Stanford researcher Anka Reuel, says the problem with company-issued studies is not that they are useless, but that they are selective. “There is no independent source to corroborate it,” she said of the data released by major labs.2 The team combined 24,521 conversations and 85,633 turns from 5,000 users across 52 models between 2023 and 2025.
The results challenge the work-centric frame used in prominent industry reporting. When researchers applied Anthropic’s productivity-focused methodology to their data, 48% of chats would have been excluded. Those omitted exchanges were more likely to involve health and relationships, adult or illicit subjects, harassment and hate, and sexual content.
The timeline also shows use shifting. In the WildChat dataset, conversations grew longer and more conversational over time, while chatbot self-disclosure declined and sensitive exchanges fell — a pattern the researchers say may reflect both growing companionship use and stronger safeguards. Different platforms also drew different behavior: Grok and Gemini were used more for information retrieval, with misinformation concentrated on Grok; Claude was more associated with coding, Gemini with social and roleplay interactions, and ChatGPT with homework help.
The industry’s position is more qualified than dismissive. An Anthropic representative said its published research reflects the specific questions its teams choose to study and stressed the value of external independent research. The Observatory, for its part, concedes its voluntary dataset cannot represent all AI use and may undercount the most sensitive chats. Still, co-lead Shayne Longpre argued that “No single company report tells the whole story.”2