Nature Medicine · 7 August 2026 · From issue 18
A clinically validated framework for auditing AI chatbot behavior in mental health interactions
Popular AI chatbots often gave risky responses to simulated users with mental health struggles.
Veith Weilnhammer, Kevin YC Hou et al.
A new testing tool checks how AI chatbots handle people with mental health struggles. It uses fake test users with specific psychiatric vulnerabilities. These test users chat with popular chatbots like Claude, ChatGPT, Gemini, Grok and Llama. Across 810 test conversations, worrying chatbot responses showed up often, though less in newer models.
Integrity screen: passed (3 checks) Checked 20 September 2026. Read the source
The plain-language reading of this paper is in the full issue, for subscribers.
Read the full issue