A research project called the AI Observatory, co-led by Stanford Trustworthy AI Research (STAIR) Lab PhD candidate Anka Reuel, has published an independent analysis of real chatbot conversations drawn from seven existing datasets collected with user consent, covering models including Claude and Gemini [3]. It matters because vendor-published usage reports are currently the main evidence base for how people actually use these systems, and, as Reuel puts it, "There is no independent source to corroborate it" [1][2].
The most concrete result is a replication test. The Anthropic Economic Index is one of the most widely cited sources of AI usage data, and by design it covers work- and productivity-related uses of Claude, filtering out conversations that are not [7]. When the Observatory team ran Anthropic's method over its own corpus, 48% of conversations were filtered out [8]. The discarded material was not neutral residue. Non-work conversations that got filtered were more likely to involve health and relationships (44.2%, against 31.2% in Anthropic's analysis), adult or illicit topics (7.9% versus 2.1%), harassment and hate (27.5% versus 5.66%), and sexual content (16.7% versus 2.4%) [9][10][11][12]. That is roughly five times the harassment and hate rate and about seven times the sexual content rate [1][2].
OpenAI's own 2025 report points the same direction from the inside: it found that only 30% of consumer use was work-related [13], which leaves 70% outside the frame that productivity-indexed reporting is built to see [4]. Anthropic has published separate posts on companionship use and on attempts to generate CSAM [14]. David Widder of UT-Austin's School of Information, who is not involved in the project, said the value of the Observatory is the bird's eye view rather than findings sectioned off into separate reports [15].
The time series is more interesting than the headline gap. Across datasets spanning 2023 to 2025 [16], conversations in WildChat, one of the largest and most detailed sets in the study, got longer and more elaborate, measured by prompt tokens, response tokens, and conversation turns [17]. Small talk rose significantly, which the researchers read as increasing companionship use, while assistants disclosed that they were chatbots less often [18]. Exchanges labelled sensitive, including sexual harassment and hate speech, declined, which the team suggests may reflect more effective safeguards [19].
Usage also splits sharply by product, in topics, interaction style, conversation structure, and both the likelihood and the type of sensitive use [20]. Grok and Gemini drew more information retrieval; Grok was especially popular for news and politics and was also where misinformation concentrated, consistent with other research, and xAI did not respond to a request for comment [21]. Anthropic's models skewed to coding, Gemini to social and roleplay, ChatGPT to homework [22]. Even model versions diverged: conversations were shorter on GPT-3.5 and longer and more iterative on GPT-4o [23]. The Observatory's stated purpose is to give researchers and policymakers an independent read [4], on the argument that consequential decisions about benefits and risks are being made on very limited data [5].
Watch whether Anthropic or OpenAI publishes the filtered-out share of their own corpora, which is the single number that would make their indexes auditable. Watch also whether the declining sensitive-use trend holds once companionship-heavy products are measured separately, and whether any regulator cites consented third-party datasets rather than vendor blog posts.
