Anthropic funds independent AI wellbeing evals with $5M

Anthropic launched a $5 million grant program on August 25, 2026 to fund independent research into how AI affects users' wellbeing. Funded teams receive direct funding, model access and technical support — and publish their evaluations as open source so any developer can use them.

Why this step? AI systems have long stopped being mere work tools; they are also conversational partners, up to and including emotional support in difficult moments. "As an industry, we are still working towards developing clear standards for how models should behave in these conversations," Anthropic writes — for instance when users seek companionship from a model or try to navigate a mental health crisis with AI.

The measurement problem is real: for most model behaviors, a single answer can be judged right or wrong. Wellbeing instead requires context across long, multi-turn conversations — risks often only surface over time. Anthropic's example: diet advice can be appropriate, but actively harmful for a user with a history of disordered eating.

Alongside the program, Anthropic is publishing guidance from its Safeguards team: it wants evaluations that state clearly what they measure, involve clinical experts in design and validation, test both overcaution and harm, reflect realistic multi-turn conversations, and validate their automated graders against real subject-matter experts.

Grantees will work fully independently, Anthropic says. The program joins a series of initiatives externalizing how model behavior gets measured — after economics and safety research, now explicitly for the psychological dimension of AI use.