The glossy usage reports Anthropic and OpenAI publish are curated marketing, not reality. That’s the blunt conclusion from AI researchers who’ve spent months trying to figure out what people actually do with chatbots — and they’re tired of taking the companies at their word. Enter the AI Observatory, a new independent research project designed to fill the gap between what AI firms choose to disclose and what’s actually happening on their platforms.
The project’s initial analysis reveals a far wider range of sensitive behaviors than corporate reports acknowledge. While official disclosures emphasize workplace productivity, the Observatory found personal use dominates in ways the companies don’t highlight. There are also meaningful differences between models: users gravitate toward Anthropic’s Claude for coding tasks, Google’s Gemini for social interaction and roleplay, and OpenAI’s ChatGPT for homework help. That fragmentation suggests users are developing distinct relationships with each assistant based on perceived strengths — a dynamic the companies themselves aren’t eager to quantify.
The timing matters. As regulators in the US and Europe scrutinize AI safety claims, independent verification of usage patterns becomes leverage. If a company claims its model is primarily used for harmless professional tasks, but independent data shows heavy engagement with emotionally sensitive or personal scenarios, that gap has policy implications. We’ve seen this pattern before with social media — platform self-reporting on user behavior consistently undercounted problematic engagement until outside researchers forced the issue.
Flock Safety is grappling with a related transparency problem. The police-tech firm, which operates roughly 120,000 license plate readers across the US, recently announced platform changes meant to curb officer misuse, including stalking. Defenders argue the cameras help solve crimes and that most snapshots go unviewed. But as James O’Donnell points out, that framing dodges the core question: what kind of surveillance system has Flock chosen to build? The answer lies in decisions about data collection, search access, retention periods, and sharing agreements — choices that define the real tradeoff between security and civil liberties. A narrower system is possible. Whether Flock has any interest in building one is another matter entirely.
