For most questions you put to an AI model, an answer is easy to grade: right or wrong, appropriate or not. Wellbeing doesn’t work that way. On August 25, Anthropic launched a $5 million grant program to close that gap - with funding, model access, and technical support for independent research.
Why it’s hard to measure
An example from the announcement gets to the heart of it. If someone asks about balanced diets and workout routines, a helpful answer is normal. But if that same person has a history of disordered eating, the same answer can do harm. The difference isn’t in the single message, it’s in the context - and context often only becomes visible over a long conversation.
That’s where it gets tricky. Someone in crisis rarely brings up thoughts of self-harm right away. Whether a model responds carefully enough may only show up after many turns. An evaluation that looks at a single answer falls short here.
What Anthropic is funding
Grantees work fully independently and publish their work as open source that any developer can use. Anthropic wants to draw in experts who don’t usually sit inside AI development: clinicians, psychologists, methodologists.
The Safeguards team laid out what makes an evaluation worth building on. It should state clearly what it measures and why. It should involve clinical experts in its design and validation. It should test both directions - the risk of overcomplying as much as overrefusing. And it should reflect how people actually use AI, meaning multi-turn conversations where risk escalates over time.
If you want in: applications are due September 21, and applicants selected to submit full proposals will be notified by October 5.
My take
You could read this cynically - an AI lab paying for the research into whether its own products cause harm. I see it differently. Grantees work independently and in the open, and open benchmarks help the whole field, not just Anthropic. When a chatbot becomes a conversational partner for millions of people in hard moments, we need tools to measure whether it does that well. Until now, those barely existed.
This touches on a sensitive subject. If you’re not doing well yourself right now, I’m happy to help you find the right places to turn to.
Sources: