Anthropic Launches $5M Grant Program for AI Well-Being Research

Anthropic is putting $5 million behind a question the AI industry has largely avoided answering rigorously: what does sustained interaction with an AI actually do to the people using it? The company announced a new grant program this week that will fund independent researchers to build open-source evaluations of how AI models affect user well-being — with a deadline of Sept. 21 for initial applications.

The program offers more than just money. Grantees will receive direct funding, access to Claude and technical support from Anthropic’s team. But the company is emphatic that grantees will operate with full independence — Anthropic does not direct their research — and all outputs must be published as open-source projects available to any developer in the industry. Selected applicants will be notified to submit full proposals by Oct. 5.

Why Standard Benchmarks Fall Short Here

Most AI evaluations are designed to test whether a model gives a correct or appropriate answer to a single prompt. Well-being doesn’t work that way. Anthropic’s announcement points to the example of a user with a history of disordered eating asking about weight loss: a response that looks entirely reasonable in isolation could be actively harmful given context that only emerges across a longer conversation. Similarly, a user in emotional distress might not mention self-harm until several exchanges in.

The company’s Safeguards team has published guidance alongside the grant announcement outlining what a rigorous well-being evaluation actually requires: clear definitions of what counts as a pass or fail, involvement of clinical and subject-matter experts in design and validation, testing for both overcompliance and overrefusal, multi-turn conversation scenarios where risk escalates over time, and grader validation against real experts — not just automated scoring.

“AI systems have become central to how many people work, learn, and solve problems. They’ve also become conversational partners and can be sources of emotional support during difficult times. But as an industry, we are still working towards developing clear standards for how models should behave in these conversations,” wrote Anthropic.

A Genuine Gap in the Evaluation Landscape

To understand why this matters, it helps to look at what already exists. OpenAI’s HealthBench, introduced earlier this year, is designed to measure the accuracy of health information AI provides — a capability benchmark. Widely used open-source frameworks like DeepEval, deployed by companies including OpenAI, Google and Microsoft, cover safety metrics such as bias and toxicity. Neither category addresses the longitudinal, context-sensitive psychological dimension Anthropic’s grants are targeting.

The gap isn’t just a technical one. A WHO-affiliated workshop in Jan. 2026 convened more than 30 experts in AI, mental health, ethics and policy specifically to address responsible AI in mental health contexts. The concern is real and recognized at the policy level. What has been missing is a funded, systematic effort to translate that concern into open, independently produced measurement tools rather than proprietary internal evals that labs keep to themselves. Anthropic’s program is structured precisely to fill that space.

What This Means for Students and Early-Career Researchers

The program is an unusually direct on-ramp for people who don’t fit the standard AI researcher profile. Anthropic explicitly names clinicians, psychologists and methodologists as target grantees alongside technical researchers. That’s a signal: this is not a program looking for another capability benchmark built by ML engineers. It’s looking for people who understand how to measure human psychological outcomes and can apply that expertise to AI behavior.

For graduate students and early-career researchers sitting at the intersection of AI and behavioral or clinical science, the timing is notable. The Sept. 21 deadline lands at the start of the fall academic semester, leaving enough runway for a doctoral student or small research team to scope a project. Because all outputs must be published as open-source projects, a successful grant produces a public, citable body of work — the kind of portfolio asset that matters for academic job markets, graduate school applications, or positions at AI safety organizations.

Recognized methodological approaches — structured patient simulations, scenario-based evaluations grounded in clinical guidelines, expert-informed annotation protocols — are realistic paths for researchers who don’t have deployment infrastructure or proprietary model access. The grant provides the latter.

The Takeaway

Anthropic is not the first company to acknowledge that AI’s impact on emotional well-being is poorly understood. It is, however, one of the first to fund independent researchers specifically to build the measurement tools the entire industry is missing — and to require that those tools be published openly. For students with backgrounds in psychology, clinical research or human-computer interaction, this is a rare moment when industry funding is actively chasing your expertise rather than the other way around.

Full application details and Anthropic’s Safeguards team guidance are available on the Anthropic website.

Source: Anthropic