Anthropic has announced a US$5 million program to fund independent research into how AI systems affect users’ wellbeing. The company says grantees will work independently and publish open-source evaluations that other developers can use, while receiving funding, model access and technical support from Anthropic.
The announcement argues that wellbeing is harder to test than factual accuracy because risk can emerge across a long conversation. A response that appears reasonable in isolation may be inappropriate when earlier context shows distress, disordered eating or another vulnerability. Anthropic says strong evaluations should define what they measure, involve clinical or subject-matter experts, reflect real multi-turn use and test both inadequate safeguards and excessive refusal.
For organisations deploying chatbots or internal assistants, the practical lesson is not to treat a generic safety score as sufficient. Testing should include realistic conversation sequences, clear escalation rules and human review for sensitive scenarios. This is an Anthropic-funded initiative, so the value of the program will ultimately depend on the independence, quality and usefulness of the research it produces.