Anthropic opens Claude usage data to external researchers in privacy-preserving pilot

Jack Clark, a co-founder of Anthropic, announced a pilot that will let researchers outside the company analyse real-world usage data from the Claude language model without compromising user privacy. Until now, studies of model impact in the field have been possible only within the organisations that operate the models, and this effort is intended to break that monopoly on insight.

The pilot rests on the premise that AI systems are “massive socio-technical things”, a view Clark articulated. Rather than expecting labs to map every possible measurement and evaluation themselves, Anthropic proposes to release not only model characteristics but also platform telemetry – data on how the systems interact with people in practice.

The initiative is hosted by The Anthropic Institute, a broader project that measures and externalises details of Anthropic’s systems and platforms. It complements the work of the company’s economics team, which publishes the Anthropic Economic Index, a series of reports that attempt to quantify the models’ impact on employment and productivity. Both efforts share the assumption that without field data, any assessment remains speculative.

Anthropic issued a call for research proposals for the next phase of the project. Researchers whose proposals are accepted will receive access to analysis tools built by Anthropic, designed to work on data that have undergone privacy-preserving processing: personally identifying information has been removed and the data have been aggregated so that activity cannot be linked to any individual user. This is the first time an external organization will be granted such live usage data from an advanced model.

Opening the tools and data to the wider research community could reduce organisational bias, product-pressure influences and the lack of genuine peer review that arise when measurements are confined to internal teams. The main caveat is that the program remains a limited-scope pilot; the outcomes will depend on the questions researchers choose to ask and on what the data can actually reveal.