Summary: Anthropic announced a $5 million program on August 25, 2026 to fund independent, open-source evaluations of AI’s effects on user wellbeing, with model access and technical support. Its guidance urged multi-turn scenarios, expert involvement, clear pass/fail definitions, testing both overcompliance and overrefusal, and validation against subject-matter experts; applications closed September 21, 2026.
The Announcement: $5 Million for Independent Wellbeing Research
Anthropic announced the $5 million grant program on August 25, 2026, with the aim of funding independent research into how AI impacts users’ wellbeing. Grantees receive direct funding, access to Anthropic’s models, and technical support while they build open-source evaluations. The point of those evaluations is to give the AI industry a way to measure how models affect the people who use them. The teams keep their independence, and everything they produce ships as open-source projects that any developer can use.
The announcement arrives at a moment when AI systems have become central to how many people work, learn, and solve problems. They have also turned into conversational partners, and at times into a source of emotional support when things are difficult. Yet the industry still has no clear standards for how models ought to behave in these conversations—for instance when a user starts seeking companionship from a model, or leans on AI to navigate a mental health crisis.
Alongside the program, Anthropic’s Safeguards team published a guidance document. It sets out what they believe makes a wellbeing evaluation rigorous enough to build on, plus the common challenges that can limit how useful an evaluation turns out to be. For any developer whose AI features touch emotional support or mental health, that guidance is a practical resource.
The Difficulty of Evaluating Wellbeing
Judging most model behaviors takes only one answer: is it accurate, and is it appropriate? Wellbeing does not work that way; it demands far more context. Anthropic offers a concrete example: a user in distress might not share thoughts of self-harm right away, and only as a long conversation unfolds does the need for a more cautious response become clear.
Context carries enormous weight here: a reply that looks fine in one setting can do damage in another. Ask Claude about losing weight, and it may answer with advice on balanced diets and workout routines. Should that same user have a demonstrated history of disordered eating, the very same answer could be inappropriate and potentially actively harmful.
A single exchange is therefore never enough. Evaluations have to stage multi-turn conversations, where context shifts and risk builds as the exchange goes on. Good evaluations, the guidance says, should do five things:
- Spell out what is being measured (i.e., what counts as a pass or fail, and why it matters)
- Bring clinical and subject-matter experts into the design and the validation
- Weigh both precautions and harms (i.e., look at what overcompliance and overrefusal each put at risk)
- Mirror how users actually use AI (often this means constructing scenarios that represent multi-turn conversations)
- Check their graders against real subject-matter experts
Those five points work as a checklist for anyone building or evaluating conversational AI, not only for grant applicants.
Inside the Grant Program
September 21, 2026 is the application deadline, and applicants who make it through to the full-proposal stage will hear back by October 5. Researchers, clinicians, psychologists, methodologists, and anyone else eager to contribute to this emerging field are welcome to apply.
Money is only part of what Anthropic provides. Grantees also get access to Anthropic’s models and technical support. The work itself, though, has to stay independent, and the results have to be published as open-source projects. That combination keeps the evaluations from being biased by Anthropic’s own interests and lets the entire industry benefit from the findings.
Product builders should read the program as a signal: Anthropic treats wellbeing as a measurable quality of AI systems. Whether or not you ever apply, the guidance document makes a valuable reference when you design evaluations of your own.
What Product Builders Can Do With This
The Safeguards team’s guidance applies directly to anyone building products that involve conversational AI—above all features that might offer emotional support or engage with users in distress. Putting it to work looks like this:
- Pin down your evaluation targets. Decide what a “good” response means inside your product, decide what counts as a failure, and put both definitions in writing.
- Get experts involved. If mental health is anywhere in scope for your product, enlist clinicians or psychologists to help design and validate your evaluation scenarios.
- Test overcompliance and overrefusal alike. A model that agrees with everything can do as much harm as one that refuses to engage, so your evaluations need to catch both extremes.
- Simulate real conversations. Sending one prompt at a time is not enough. Stage multi-turn scenarios in which the user’s state shifts as the conversation goes on; that is where risks tend to surface.
- Validate your graders. Automated or human, whoever does your rating has to agree with the judgments of actual experts in the field.
None of this is academic. Each step is a practical way to reduce the risk of your product causing harm, and together they match the kind of rigor that regulators and users are increasingly expecting.
Where the Approach Falls Short
Anthropic is upfront that the right way to evaluate wellbeing will keep changing as models and the ways people use them change. No solution settles the question once and for all. The guidance marks a beginning, not an ending.
One limitation: the evaluations are meant to be open-source, yet the program centers on Anthropic’s own models, which leaves room for a bias toward Anthropic’s safety priorities. The grantees’ independence and the open-source requirement exist to hold that in check.
Measuring wellbeing at all is a second trade-off. Accuracy and safety can be pinned down; wellbeing is subjective and tied to context. Experts or no experts, some nuances will slip past any evaluation, and the guidance says as much when it stresses continuous evolution.
So don’t count on this program to hand you a finished framework. Take it as raw material for one of your own, and plan to revise it as you learn how people actually talk to your AI.
The Bottom Line
Wellbeing stands to become a measurable, improvable quality of AI systems, and Anthropic’s $5 million outlay is a significant step in that direction. For product builders, the safest immediate bet is the Safeguards team’s guidance—run it as a checklist over your own evaluation processes. If conversational AI is what you build, begin by making sure your evaluations span multi-turn conversations and draw on real experts. None of that is marketing speak; it is one concrete way to cut risk and hold on to your users’ trust.
Researchers and clinicians should consider putting in an application. September 21 is the deadline, and the chance to shape how the industry measures AI’s impact on human wellbeing is a rare one. Whether or not you take part, the open-source evaluations this program produces should prove valuable to the community as a whole.