Anthropic announced a $5 million grant program on August 25, 2026, to support independent research into AI models’ effects on user wellbeing. According to the company, recipients will receive direct funding, access to its models and technical support while developing open-source evaluations and benchmarks.
Grant recipients will work independently from Anthropic and publish their projects as open source so that other developers can use them. The company is seeking contributions from clinicians, psychologists, methodologists and specialists in other relevant fields.
The program focuses partly on assessing long, sensitive interactions in which the significance of a response may only become apparent as context develops. Anthropic gives the example of a distressed user who does not immediately disclose thoughts of self-harm. It also notes that otherwise reasonable diet or exercise advice could be inappropriate if a user has shown a history of disordered eating.
Anthropic says proposed evaluations should define clearly what they measure, what constitutes a pass or failure and why the result matters. Clinical and subject-matter experts should participate in both design and validation. Tests should cover the opposing risks of overcompliance with potentially harmful requests and overrefusal when assistance would be appropriate.
The company also wants evaluations to reflect actual patterns of AI use. That may require multi-turn scenarios in which risk escalates and context changes during a lengthy conversation. In addition, the judgments produced by evaluators or graders should be checked against those of real subject-matter experts.
Applications are due September 21, 2026. Anthropic says applicants selected to submit full proposals will be notified by October 5. The announcement describes funding for work that has yet to be conducted and does not report results from new evaluations.
Practical context: The program’s practical value will depend on whether the resulting open tests apply across different models and realistic user situations. Requirements for extended conversations and expert validation may make the evaluations more informative, but their comparability will become clear only after researchers publish their methods and findings.
Sources
Event date: 2026-08-25. Primary source date: 2026-08-25.
