Funding better evaluations of AI's impact on wellbeing

Key Findings

• — $5 million programme; initial applications due 21 September 2026, with full-proposal invitations by 5 October 2026
• — Funded evaluations must state clearly what they measure, meaning what counts as a pass or fail and why it matters
• — Funded evaluations must involve clinical and subject-matter experts in design and validation, and must validate their graders against real subject-matter experts
• — Funded evaluations must test both precautions and harms, explicitly evaluating the risk of overcompliance and of overrefusal together
• — Funded evaluations must reflect actual use, which the announcement glosses as multi-turn scenarios where risk escalates and context shifts over the course of a long conversation
• — The announcement states that the industry is still working toward clear standards for model behaviour when a user seeks companionship or uses AI to navigate a mental health crisis

Methodology Notes

No study, sample or results: this is a funding announcement carrying a stated set of evaluation-design criteria and an application deadline. The publisher is the funder, so the criteria describe what it will pay for rather than a validated or consensus standard. Date taken from the date string on the artifact page and the news index. Note the slug: the page is at /news/wellbeing-research-grants, not a slug derived from the title.

Sources