AI companies like Anthropic and OpenAI publish selective reports on how people use their products, leaving researchers without independent verification. Anka Reuel, a PhD candidate at Stanford's STAIR Lab, co-leads the AI Observatory—a public platform aggregating real AI conversations from 85,633 conversational turns across 24,521 conversations from five major models between 2023 and 2025.

The AI Observatory found that 48% of conversations would be filtered out by Anthropic's work-focused Economic Index, revealing significantly more sensitive behaviors including health discussions (44.2% vs 31.2%), adult topics (7.9% vs 2.1%), and harassment (27.5% vs 5.66%). The platform also discovered that usage patterns vary dramatically between models—Grok leads in news and misinformation, while users turn to Anthropic for coding and ChatGPT for homework.

Researchers observed evolving behavior over time, with longer conversations and increased small talk suggesting growing AI companionship. However, the voluntary nature of the datasets likely underrepresents sensitive uses. The team hopes their work will pressure companies to share data with independent researchers while protecting user privacy.

As Reuel warns, consequential AI policy decisions are currently being made without knowing what's actually happening beyond company narratives.