OpenAI's MentalHealthBench: rating how AI behaves, not how smart

The release of GPT-6.1 Sol and MentalHealthBench in the same period sets up a contrast the author draws between capability and impact. Until now, generative AI evaluation centered on how smart a model was: solving math, writing code, answering facts correctly, reasoning. That competition is now entering a stage of maturity, the author argues, while AI's social influence has moved past simple question-and-answer.

OpenAI itself lists the situations where people use AI, including relationship worries, handling stress, family problems and difficult decisions, and says ChatGPT users exceed 1 billion per week. So the question becomes not whether an answer was correct, but what the conversation did to the person. The author also offers reasons people confide in AI: it does not judge, it responds at any hour, and it appears to have no stake in the relationship.

Seen this way, the benchmark's emphasis on preserving user autonomy points at the next theme in AI safety research, beyond preventing harmful content and hallucinations. Whether this becomes an industry standard is an open question, and the author's judgment that history will favor the benchmark over the model is just that, a judgment. But the direction it signals is likely to matter most to those building and buying AI agents meant to work closely with people.