Anthropic puts $5M toward AI mental health research

People are increasingly turning to AI for late-night companionship and emotional support, but almost nobody actually knows how these long interactions affect human psychology over time.

Anthropic just launched a $5 million grant program aimed at solving that problem. The company is funding independent researchers to build open-source benchmarks that measure AI's impact on user wellbeing, offering grantees direct cash, Claude model access, and technical support.

Applications are open until September 21, and shortlisted teams will be invited to submit full proposals by October 5.

The multi-turn minefield

Benchmarking code or math is straightforward: the answer is either right or wrong. Evaluating emotional wellbeing is vastly messier.

As Anthropic pointed out, assessing safety on a single-turn prompt misses the point entirely. A user struggling with self-harm usually doesn't announce it immediately; the warning signs only emerge gradually over a multi-turn conversation that could span hours.

Context also changes everything. If someone asks Claude for weight loss tips, standard dietary advice seems perfectly benign. But if that user has a history of disordered eating, that exact same advice can be actively harmful.

To tackle this, Anthropic is setting specific requirements for grantees. Projects must involve actual clinicians, psychologists, and methodologists. They must construct multi-turn scenarios where risks escalate naturally.

Crucially, researchers must evaluate both overcompliance and overrefusal. An AI that blindly agrees with a delusional user is dangerous, but an AI that panics and spits out a suicide hotline every time someone mentions feeling stressed is equally broken.

All tools and benchmarks resulting from the grants will be released open-source for anyone across the industry to use.

Why it matters

There is nothing for regular users to download today, but this initiative targets one of generative AI's most uncomfortable open secrets.

Millions of people are already using chatbots as ad-hoc therapists and pseudo-friends. Today, almost every model handles this terribly, vacillating between sycophantic yes-men and sterile, liability-dodging safety disclaimers.

We think Anthropic is making a smart move by recognizing that AI engineers are completely unqualified to define mental health guardrails on their own. Bringing in outside clinicians and funding open-source benchmarks gives the whole effort much-needed credibility.

That said, quantifying mental health into automated grading scripts is notoriously difficult. Clinical psychology itself lacks global consensus on many conversational boundaries, so turn-by-turn rubrics run a serious risk of becoming overly rigid.

Still, as models become more capable conversationalists, the battleground for safety is clearly shifting from simple content moderation to understanding the psychological toll on real humans.