Anthropic Launches $5M Fund for AI Wellbeing Evaluations
Anthropic launches a $5 million grant program for independent, open-source tests of AI's impact on wellbeing, with applications due September 21, 2026.
Summary
Anthropic announced on August 25, 2026, a $5 million grant program for independent research measuring how AI affects user wellbeing. Grantees receive direct funding, access to Anthropic models and technical support while retaining full independence. Their evaluations will be published as open-source projects available to any developer. Anthropic wants clinicians, psychologists, methodologists and other specialists to help establish standards for Claude conversations involving companionship, emotional support and mental health crises.
Wellbeing is difficult to judge from one response because risks can emerge across long, shifting conversations, including delayed disclosure of self-harm or weight-loss advice given after a history of disordered eating. Anthropic already develops safeguards and studies Claude conversations, but says methods must evolve with models and usage. Proposed evaluations should define pass and fail criteria, involve clinical and subject-matter experts, test harms from overcompliance and overrefusal, represent escalating multi-turn conversations and validate automated graders against experts. Applications close September 21, and applicants selected to submit full proposals will be notified by October 5.
Positives
- $5 million will fund independent research into AI's effects on user wellbeing.
- Grantees receive direct funding, Anthropic model access and technical support without surrendering research independence.
- All funded evaluations will be open-source, allowing any developer to reuse them.
- Clinical experts, psychologists and methodologists are being invited to shape evaluation standards.
- Required tests will examine both overcompliance and overrefusal in realistic multi-turn conversations.
Risks & concerns
- The AI industry still lacks clear standards for companionship, emotional-support and mental-health conversations.
- Self-harm risk may become visible only after a long conversation, making single-response evaluations inadequate.
- Claude's ordinary diet and exercise advice could harm someone with a demonstrated history of disordered eating.
- Wellbeing judgments change with context, while model capabilities and user behavior continue evolving.
- Poorly designed benchmarks may mislead unless their graders are validated against real subject-matter experts.