Anthropic launches $5 million AI wellbeing research grant program

Anthropic’s grant program

Anthropic announced a $5 million grant program to support independent research into how AI affects users’ wellbeing. The company said the program will provide direct funding, access to its models and technical support to grantees developing open-source evaluations.

Grantees will conduct their work independently, according to Anthropic. Their projects will be published as open source so that developers can use them.

Anthropic said it wants the program to involve clinicians, psychologists, methodologists and other specialists in creating evaluations and benchmarks of user wellbeing.

Focus of the evaluations

Anthropic said wellbeing can require more conversational context to evaluate than model behavior that can be judged from a single response. The company described situations in which risks become apparent only over a longer exchange or where a response may be appropriate in one context but harmful in another.

One example in the announcement concerns advice about diet and exercise. Anthropic said such advice could be reasonable when a user asks about losing weight, but could be inappropriate or harmful if the user has shown a history of disordered eating. It also cited conversations involving companionship and mental health crises as areas where standards for model behavior are still being developed.

The company said it develops safeguards intended to identify such conversations and help Claude respond appropriately. Anthropic also said it publishes research about the kinds of conversations users have with Claude to inform safeguards, their evaluation and other user-wellbeing measures.

Evaluation guidance

Alongside the grant program, Anthropic shared guidance from its Safeguards team about the characteristics it believes make wellbeing evaluations rigorous and the challenges that may limit their usefulness.

Anthropic said it is seeking evaluations that:

• Clearly define what they measure, including what constitutes a pass or failure and why the measurement matters.

• Include clinical and subject-matter experts in their design and validation.

• Test both precautions and harms, including risks from overcompliance and overrefusal.

• Represent how people use AI, including multi-turn scenarios in which risk increases and context changes during a long conversation.

• Validate automated or other graders against real subject-matter experts.

Anthropic directed prospective applicants to an application form and separately provided guidance on building wellbeing evaluations and benchmarks. Applications are due September 21. The company said applicants selected to submit full proposals will receive notification by October 5.

Source: Anthropic’s “Funding better evaluations of AI’s impact on wellbeing” announcement, published August 25, 2026.