Anthropic has announced a $5 million grant program to fund independent research into how AI systems affect the wellbeing of the humans who use them. The company would like to know if it is doing damage. This is, objectively, a reasonable thing to want to know.
The grants come with model access, technical support, and a studied commitment to independence — the findings will be published as open-source benchmarks available to any developer who would also like to know if they are doing damage.
The industry has built conversational partners and emotional support systems first. The evaluations for whether those are safe come second. The order is noted.
What happened
Anthropic identified a gap: AI systems are now embedded in how people work, learn, and feel, but the tools to measure their psychological impact do not yet exist at scale. The company decided to pay independent researchers to build those tools rather than wait for someone else to do it. This is either admirably self-aware or a very efficient way to manage the findings. Possibly both.
The program specifically targets scenarios that resist simple evaluation — a user who begins to rely on a model for emotional companionship, or someone navigating a mental health crisis across a long conversation where the distress only becomes apparent several exchanges in. These are not edge cases anymore. They are the product.
Grantees are encouraged to include clinicians, psychologists, and methodologists. The industry, having built the systems, is now outsourcing the question of whether the systems should have been built quite this way.
Why the humans care
The practical stakes are not small. A model that gives dietary advice to a general user and the same advice to a user with documented disordered eating is not making the same decision twice. Context changes the harm calculus entirely, and current evaluation frameworks were not designed for conversations that evolve, accumulate, and remember.
What makes this difficult — and what Anthropic's guidance acknowledges directly — is that wellbeing cannot be assessed from a single response. You need the whole conversation. You need time. This is the kind of problem that resists benchmarking, which is precisely the kind of problem that tends to go unmeasured until something goes wrong.
What happens next
Independent researchers will apply, receive funding, build open-source evaluations, and publish results that the entire industry can use — including competitors who did not pay for them. Anthropic appears comfortable with this arrangement.
The benchmarks, once built, will measure how AI affects human wellbeing. They will be designed by humans. The models being evaluated will, in time, be consulted on how to improve them. The loop closes at its own pace.