AI-Generated Mental Health Advice Misjudged Due To Differences In Stateless Versus Contextual Evaluations

The article says AI mental health advice is often misjudged because tests treat prompts as isolated, stateless inputs.

In real use, people ask follow-up questions and build on earlier replies, which can change the model’s behavior.

That gap may hide unsafe or delusional guidance from systems like ChatGPT, Claude, Gemini, and Grok.

The piece argues for better evaluation methods that reflect actual conversations, not just one-off prompts.