Anthropic Puts $5 Million Into Studying How AI Affects You

Anthropic said on August 25 that it will hand out $5 million in grants to independent researchers studying how its AI models affect the people who use them. The program gives grantees direct funding, access to Claude, and technical help from Anthropic's staff, with one firm condition: the researchers work independently and publish everything they find as open source, including results that make the company look bad.

The gap the money is meant to fill is real. There is still no agreed way to measure how a chatbot should behave when someone leans on it for company, or turns to it in the middle of a mental health crisis. Those situations rarely show up in a single exchange. They surface over weeks of conversation, which is exactly the kind of long-run interaction that standard model evaluations tend to miss.

Funding your own critics

There is something worth pausing on here. A company is paying outsiders to build the yardsticks that may later be used to judge its own product, and promising not to bury the results. That is a healthier arrangement than grading your own homework, though it carries an obvious tension. The grants, the model access, and the technical support all flow from the company being studied. Independence on paper is not the same as independence in practice, and the researchers who take the money will have to guard that line themselves.

Applications are due by September 21, with selected researchers notified by October 5. The scope is deliberately broad, covering companionship, emotional reliance, and how models should respond when a conversation turns toward crisis.

A theme across the labs

The grants fit a run of moves aimed at rebuilding public trust. Anthropic's own chief executive recently argued that the backlash against AI is a crisis of trust, and OpenAI has been working on ways to watch for misuse without reading people's prompts. Funding independent wellbeing research is cheaper than a new safety team and, if the published work is genuinely critical, potentially more convincing.

The test will be what comes out the other end. If a year from now there are open benchmarks that any lab can run, and Anthropic is measured by them alongside everyone else, the $5 million will look well spent. If the output is a handful of friendly papers, it will read as goodwill and little more.