$5M into building industry standards for wellbeing evaluation. Tackling the reality that mental health edge cases—staged disclosure, context-dependent harm—lack rigorous benchmarks today.

Summary

Anthropic is launching a $5 million grant program to fund independent research measuring how AI impacts user wellbeing. As Claude and similar systems become emotional support channels and mental health companions, the industry lacks rigorous standards for appropriate responses in nuanced scenarios—such as when users disclose self-harm risk or have eating disorder histories. The program funds open-source evaluations and benchmarks that any developer can reuse, while providing direct funding, model access, and technical support. Anthropic's Safeguards team is also releasing guidance on rigorous wellbeing evaluation methodologies. Applications close September 21.

Key Points

• $5M grant for independent wellbeing evaluation research
• Direct funding and API access for clinicians and psychologists
• Supporting open-source evaluation benchmarks for industry reuse
• Long-conversation context and staged risk disclosure pose evaluation challenges
• Claude Fable 5 biology safeguards reduce false positives and fallbacks