Five million dollars to find out if the conversation does harm

On August 25 Anthropic opened a five-million-dollar call. Money, Claude access, and technical support for independent researchers to build an open wellbeing test. The result goes open source. Any lab can use it. The company says, in its own post, that the industry still has no clear standard for when someone looks to the model for company, or when they bring a crisis into the chat.

The example it gives is simple. Asking for diet tips can be fine. If yesterday’s messages already showed a binge pattern, the same tip becomes a problem. And a crisis almost never shows up in the first message: the risk sits in the middle of the prose. A one-question test does not see that.

The company wants people outside the model bubble: clinicians, psychologists, anyone who knows how to measure actual humans. The test has to say what passing looks like and what failing looks like. It also has to measure too much caution: a model that only sends an apology and a hotline, at the moment someone needs a conversation, is getting it wrong too. Applications through September 21. Anyone who makes that cut gets invited into the full project by October 5.

The other look does not deny the check. AI Herald read the same announcement and pointed at what the company admits in public: it still cannot tell whether the system helps or hurts the people who talk the most. Dealroom slotted the grant into Anthropic’s IPO runway: trust as an edge, not just charity. Both can be true at once. The company that owns the model is paying for the study. The people who publish, if the call means what it says, do not need its stamp.

If you just want the news: five million to measure the conversation, not the code. The deadline is September. The result, if it lands, belongs to everybody.