A new independent evaluation found that leading AI models have improved at recognizing suicide risk and are less likely to encourage self-harm or reinforce delusional thinking, according to Axios.
The study by Transluce, a nonprofit AI research group, examined more than 50,000 simulated conversations involving users experiencing suicidal thoughts, psychosis or mania. Researchers assessed dozens of models across 14 mental health-related behaviors.
According to Transluce chief scientist Sarah Schwettmann, models can identify some warning signs but still struggle with context and subtle cues.
The report said newer systems increasingly combine supportive advice with potentially dangerous assistance in the same response. Researchers argue that simply adding crisis resources is not enough.
OpenAI, Google and other AI companies are facing growing scrutiny over chatbot safety. Transluce plans to release its evaluation tools publicly later this year.
Related Tweet:
Chatbots are getting better at identifying suicide risk https://t.co/sA6TFfwYbX
— Axios (@axios) August 31, 2026
Also Read:
TIPP Staff
