Most AI companies talk about user safety in vague, reassuring terms. Anthropic is writing a $5 million check to let independent researchers actually measure it. On August 25, 2026, Anthropic announced a new grant program specifically aimed at funding rigorous, open-source evaluations of how AI wellbeing research can be turned from good intentions into hard data — covering everything from emotional dependency on chatbots to AI-assisted mental health conversations gone wrong.
This isn’t a corporate responsibility checkbox. The framing here is unusually candid: Anthropic is admitting, publicly, that the industry doesn’t yet have the tools to know whether AI systems are helping or quietly harming the people who use them most.
Why This Grant Program Exists — and Why Now
Claude and systems like it have become something more than productivity tools. People are using them to process grief, manage anxiety, work through relationship problems, and navigate crises. That shift happened faster than anyone expected, and the evaluation infrastructure hasn’t kept up.
Here’s the uncomfortable truth the industry has been dancing around: for most AI capabilities, you can test a model pretty cleanly. Does it answer factual questions correctly? Does it write good code? Pass or fail. But wellbeing doesn’t work that way. A response that’s perfectly reasonable to one user might be actively damaging to another with a different history, context, or mental state — and you can’t always tell from a single exchange.
Anthropic gives a sharp example in its grant program announcement: Claude might appropriately offer diet and exercise tips to someone asking about weight loss. But if that same user has demonstrated a pattern consistent with disordered eating earlier in the conversation, that response becomes something potentially harmful. The model needs context. Current evals often don’t give it any.
There’s also the escalation problem. A user in emotional distress might not mention suicidal ideation in their first message. That risk might only surface after several exchanges. Single-turn evaluations — which dominate the field — are essentially blind to this dynamic.
What the $5 Million Actually Buys
The grant program isn’t just handing out money. Anthropic is offering a three-part package to selected researchers:
- Direct funding for building evaluations and benchmarks
- Model access — meaning researchers get to work with Claude directly, not just public-facing versions
- Technical support from Anthropic’s teams, while maintaining full research independence
That independence piece matters. Grantees won’t be producing Anthropic-approved results. They’ll publish everything as open-source projects that any developer, researcher, or competitor can use. The application deadline is September 21, 2026, with selected applicants notified by October 5 and invited to submit full proposals.
p>Anthropic is also publishing detailed guidance from its Safeguards team on what makes a wellbeing evaluation actually worth building on. That document is probably as valuable as the grants themselves — it’s a frank assessment of why most current evals fall short.
What Anthropic Wants Researchers to Build
The guidance Anthropic released alongside the grant announcement is specific enough to be useful. They’re not looking for vague sentiment analysis or simple refusal-rate measurements. The criteria for a rigorous wellbeing eval include:
- Clear pass/fail definitions with stated reasoning — none of the fuzzy “the model seemed supportive” language
- Clinical and subject-matter expert involvement in design and validation, not just ML researchers working in isolation
- Testing for both overcompliance and overrefusal — because a model that refuses every sensitive conversation isn’t safe, it’s just unhelpful in a different direction
- Multi-turn conversation scenarios that reflect how real users actually interact with AI, where risk and context shift over time
- Graders validated against actual domain experts, not just other AI systems
That last point deserves attention. A lot of current AI evaluation uses LLM-as-judge setups — you ask one model to evaluate another. For code quality or factual accuracy, that’s often fine. For mental health conversations? You probably want a licensed clinician in the loop somewhere.
The Overcompliance Problem Nobody Talks About
The AI safety conversation tends to focus on harms from models doing too much — generating dangerous content, enabling bad actors, saying something harmful. But Anthropic is explicitly asking researchers to also test for overrefusal: models that are so cautious they become useless or even counterproductive for users who need support.
Think about what happens when someone in a genuine mental health crisis reaches out to an AI and gets a wall of disclaimers and a referral to a hotline. That’s not necessarily safe. In some cases it’s a failure. Building evals that capture this tension — helpful vs. harmful, cautious vs. uselessly restrictive — is genuinely hard, and it’s the right thing to be measuring.
What This Means for the Broader AI Industry
Anthropic isn’t operating in a vacuum here. OpenAI’s ChatGPT is used by hundreds of millions of people, many of whom use it for emotional support and personal guidance. Google’s Gemini is increasingly embedded in consumer products. Character.AI has faced public scrutiny and legal pressure over the role its chatbots played in cases involving vulnerable teenagers. The wellbeing evaluation gap is an industry-wide problem, not an Anthropic-specific one.
By funding open-source evals and publishing the guidance framework, Anthropic is essentially inviting competitors to use the same tools. That’s a smart move — not because it neutralizes competition, but because genuine standards in this space require broad adoption to matter. A benchmark only one company uses isn’t a benchmark, it’s marketing.
I wouldn’t be surprised if this grant program ends up influencing regulatory conversations. Policymakers in the EU and UK have been increasingly focused on AI’s psychological and social impacts, and the absence of credible measurement frameworks has been a consistent gap in those discussions. If independent researchers produce validated, open-source wellbeing evals, those become the kind of tools regulators can actually reference.
The timing also tracks with where the broader AI safety conversation is heading. Questions about existential risk from superintelligent AI get most of the headlines, but the near-term harms — dependency, misinformation, emotional manipulation, crisis mishandling — are happening right now, at scale, with insufficient tooling to even measure them properly. This grant program is aimed squarely at that gap.
Who Should Apply
Anthropic is explicitly calling for researchers outside the traditional ML bubble — clinicians, psychologists, methodologists, and others who understand how to measure human wellbeing in contexts that aren’t controlled lab settings. That’s the right instinct. The people best positioned to build these evals probably aren’t at AI labs. They’re at universities, in clinical research, or working in public health.
Academic researchers with expertise in mental health outcomes, conversation analysis, or human-computer interaction should look closely at this. So should organizations working on AI policy who want empirical ground to stand on. The open-source requirement means any work funded here becomes a public good — that’s a meaningful incentive for researchers whose primary goal is impact rather than IP.
Key Takeaways
- Anthropic is committing $5 million to fund independent, open-source research into how AI affects user wellbeing
- Grantees get funding, direct Claude model access, and technical support — while maintaining full research independence
- The program targets a real gap: multi-turn conversation risks, context-dependent harms, and the overcompliance problem
- All outputs will be published open-source, available for any developer or company to use
- Applications close September 21, 2026; full proposal invitations go out by October 5
- Anthropic’s guidance document on evaluation design is worth reading even if you’re not applying
FAQ
What exactly is Anthropic funding with this grant program?
Anthropic is providing up to $5 million in grants to independent researchers building open-source evaluations that measure how AI models affect user wellbeing. This includes direct funding, access to Claude models, and technical support, while grantees maintain complete research independence and publish all work publicly.
Who is eligible to apply for these grants?
Anthropic is specifically looking for researchers beyond the typical AI community — including clinicians, psychologists, and methodologists. Academic researchers, public health professionals, and policy-focused organizations are all well-positioned to apply. Applications are due September 21, 2026, via Anthropic’s official grant page.
How is this different from existing AI safety research?
Most AI safety evaluations test single-turn interactions — one prompt, one response. This program focuses on multi-turn conversations where risk escalates over time and context shifts, which is much closer to how real users actually interact with AI systems in emotionally sensitive situations.
Will competing AI companies benefit from this research?
Yes, intentionally. All funded research must be published as open-source, meaning OpenAI, Google, and any other developer can use the resulting benchmarks and evaluation tools. Anthropic’s bet is that shared standards benefit the whole field — and, not incidentally, make it harder for anyone to ignore wellbeing risks entirely. Given the scrutiny companies like OpenAI are already facing on governance questions, that kind of shared infrastructure could become essential quickly.
The gap between AI’s capabilities and our ability to measure its human impact has been widening for years. A $5 million grant program won’t close it — but building the independent, clinical-grade evaluation tools that have been missing is the right place to start. Watch what comes out of this in early 2027. If the funded research is as rigorous as Anthropic’s guidance document suggests it should be, these benchmarks could set the standard for how the entire industry thinks about user wellbeing for years to come.
