Anthropic is paying outside researchers to grade AI's effect on user wellbeing, and says it won't direct the work

Anthropic announced a $5 million research-grant fund on August 25, 2026, for independent researchers building open-source evaluations and benchmarks of how AI affects the people who use it. Grantees get direct funding, access to Claude models and technical support. Per Anthropic, they work "fully independently," and every output must publish as an open-source project regardless of what it finds.

Who it's for

The program explicitly targets clinicians, psychologists and methodologists alongside ML engineers. Measuring whether a model's output is good for a person's psychological state is a different problem from checking math or code correctness. Anthropic notes that AI systems "can serve as sources of emotional support during difficult times," and existing evaluation suites mostly aren't built to catch when that support is inappropriate for the specific person receiving it. Initial applications are due September 21; applicants selected from that pool submit full proposals by October 5.

The independence clause

Wellbeing is a hard quality problem: there's no compiler to check against and no test suite that passes or fails. Anthropic is funding outside researchers to build the evaluation, requiring publication regardless of outcome, and disclaiming any right to steer the findings. That structure produces the kind of evidence a vendor can't credibly produce about itself. Whether the resulting evaluations hold up can be judged when they publish. The funding structure is the part other labs should copy.