Anthropic lets outside researchers study aggregate Claude usage data

Anthropic gave three outside research groups access to Anthropic Insights, an internal tool that lets a researcher pose a question about how people use Claude and receive back only an aggregated, categorised answer, never a raw conversation. Stanford’s SALT Lab will study what work people delegate to Claude and where friction arises in that handoff; Oxford’s Human Information Processing Lab will study emotional experience during Claude use and how it relates to model behaviour; and METR will measure productivity gains from coding agents across successive model generations. A third-party privacy audit, commissioned from Imperial College London, reviewed the tool’s protections before the pilot began.

Anthropic framed the pilot around a structural argument rather than a specific finding: that real-world usage data — the record of what people actually ask AI systems to do, and how well it goes — sits concentrated inside a handful of labs, leaving researchers, policymakers and the public reliant on those labs’ own published summaries rather than independent analysis. Anthropic said the three teams would design their own studies and publish their own conclusions, including ones unflattering to Claude, rather than being directed toward predetermined questions.

The pilot is narrow: three groups, one company’s data, and a tool Anthropic itself built and controls access to, so a researcher’s independence extends to study design and publication rather than to the data pipeline itself. It arrived the day after Anthropic opened a separate $5 million grant fund for outside research into AI’s effect on user wellbeing, part of the same week’s push to fund research Anthropic said it did not want to conduct or direct itself.