August 25, 2026 — An organization announced a $5 million grant program to fund independent research into how artificial intelligence systems affect users’ wellbeing. The program will offer direct funding, access to the organization’s models, and technical support to grantees; awardees will develop open-source evaluations that other developers can use and will work fully independently.
Why this matters
AI systems are increasingly central to how people work, learn, and solve problems, and they also serve as conversational partners that can provide emotional support. Despite this, the industry lacks clear, widely adopted standards for how models should behave in sensitive conversations — for example, when a user begins to seek companionship from a model or uses AI to navigate a mental health crisis.
Assessing wellbeing is particularly challenging because it often requires more context than judging a single model response. A user in distress may not immediately disclose self-harm thoughts; the need for a cautious or escalated response might only become apparent over a long, multi-turn conversation. A response appropriate in one situation can be harmful in another — for instance, offering general diet or exercise advice to someone with a history of disordered eating could be inappropriate or damaging.
The organization’s Safeguards team is developing measures to detect such risky conversations and help ensure that Claude responds appropriately. The organization also publishes research on conversation types people have with Claude to inform safeguard development and evaluation. They note that these considerations are nuanced, the stakes are high, and approaches must evolve alongside models and their uses.
What the grant is looking to fund
The program aims to fund the creation of independent wellbeing evaluations and benchmarks that are published as open-source projects. Grant recipients will retain independence over their work while receiving funding, model access, and technical assistance.
Guidance from the Safeguards team outlines common qualities of useful wellbeing evaluations. Specifically, the organization seeks proposals that:
- Clearly state what they measure (what counts as pass or fail and why it matters);
- Involve clinical and subject-matter experts in design and validation;
- Test both precautions and harms (evaluate risks of both overcompliance and overrefusal);
- Reflect how users actually use AI, often by constructing multi-turn conversational scenarios where risk can escalate and context shifts over time;
- Validate their graders against real subject-matter experts.
Application process and deadlines
Applicants can learn more and apply via the program’s application form. Applications are due by September 21, 2026; applicants selected to submit full proposals will be notified by October 5, 2026.
The initiative aims to broaden participation in this emerging field by inviting clinicians, psychologists, methodologists, and others to contribute expertise, with the goal of building more effective, transparent evaluations and benchmarks for AI’s impact on user wellbeing.
