Funding Better Evaluations of AI's Impact on Wellbeing
Anthropic is putting up $5 million to fund independent research into how AI models affect the wellbeing of the people who use them, with a particular focus on emotionally sensitive situations like mental health support and crisis conversations.
Details
- Program size: $5 million total, distributed as grants to outside researchers and institutions
- Why it matters: As people increasingly turn to AI chatbots for emotional support, companionship, and even crisis conversations, there’s little rigorous, independent evidence about what actually helps or harms users in these interactions
- What the grants target:
- Building open-source evaluation tools and benchmarks for measuring AI’s effect on user wellbeing
- Assessing model behavior in sensitive contexts such as mental health support, emotional companionship, and crisis situations
- Improving evaluation rigor for multi-turn conversations, where risk can escalate gradually as a conversation continues
- Evaluation guidance: Anthropic published guidance for applicants stressing that good evaluations should state clearly what they are measuring, involve clinical experts in their design, and test for both overcompliance (the model going along with harmful requests) and overrefusal (the model being unhelpfully cautious), validated against subject-matter experts
- How to apply: Interested researchers submit proposals through a Google Form linked from the announcement
- Timeline: Applications are due September 21, 2026, with notifications on full proposals going out October 5, 2026
- Recipients: Not yet announced — the program was still accepting applications at the time of publication
What happened next
Anthropic has not yet named specific grant recipients or individual award amounts; those details are expected to follow the September 21 application deadline and October notification date. The move fits into Anthropic’s broader push on AI safety research, extending its focus from misuse and capability risks to the more subjective, harder-to-measure question of psychological impact on everyday users.