Anthropic gives researchers $5 million to study AI and well-being
Anthropic is funding $5 million in research to understand how AI can affect users’ mental and emotional well-being.
AI isn't just a mere chatbot today. People are also turning to it for advice, support and, sometimes, a place to talk. Anthropic is putting $5 million behind research into what that means for people’s wellbeing.
Announced on 25 August 2026, the grant programme will fund independent researchers working on better ways to evaluate how AI models affect users in sensitive, personal and emotionally charged situations.
The selected researchers will receive funding, access to Anthropic’s models and technical support. Their work will also be published as open-source projects, allowing others to use the evaluations.
AI gets harder to measure when conversations get personal
It is relatively easy to test whether a chatbot gets a maths question right or summarises a document accurately. Wellbeing is a different challenge. People may use AI when they are stressed, lonely or uncertain, and those conversations can develop over time.
A response that seems harmless on its own could become problematic when combined with what a user has said earlier. Anthropic points to weight-loss advice as an example. General guidance on diet and exercise may be reasonable, but the same advice could be inappropriate if a user has shown signs of disordered eating.
That makes context important. AI wellbeing evaluations need to look at longer conversations and changing situations, rather than judging one response at a time.
What Anthropic wants researchers to test
The programme will focus on evaluations and benchmarks, essentially tests that show how AI behaves in specific situations. Claude's maker wants researchers to develop tests with clear measures and pass-or-fail criteria, backed by clinical or subject-matter experts.
The evaluations should look at both potential harm and excessive caution. That last part is important. An AI can be too willing to respond to a risky request, but it can also become so cautious that it refuses harmless ones. Researchers will therefore need to examine both sides.
Applications for the programme are due by 21 September 2026, with applicants selected to submit full proposals expected to be notified by 5 October 2026.
Why open-source research matters
Anthropic says the research will be independent and publicly available. That means the evaluations will not be limited to Claude or kept within the company. This could help bring clinicians, psychologists, AI evaluation experts and other specialists into a conversation that is becoming harder for technology companies to handle on their own.
As AI becomes more conversational and personal, simply measuring whether an answer is correct is no longer enough. Companies also need to understand what happens when users turn to AI during difficult moments.