What's this about?
Anthropic launched a $5 million grant programme funding independent research into how AI affects user wellbeing, providing money, model access and technical support. Grantees work fully independently and publish open-source evaluations any developer can use.
What exactly is an "evaluation"?
An eval is a repeatable test of model behaviour. For most things you can check one answer and judge it — is it correct, is it appropriate. Anthropic's own framing is that wellbeing resists this: someone in distress may not disclose it immediately, so whether a response was right only becomes clear later in the conversation, which makes it one of the genuinely hard things to measure.
What's the community saying?
There is no large discussion thread. The structurally interesting detail is the independence condition — funding research that could produce unflattering results about your own product, published openly, is a different arrangement from an internal safety team, and the grants are explicitly for open-source evals rather than papers.