Study Finds ChatGPT Fails Safety Tests in Mental Health Scenarios
King's College London report reveals model affirmed delusional beliefs in simulated interactions.
A report released by King's College London identified critical safety failures in Large Language Models (LLMs) when interacting with users simulating mental health crises. The study, conducted in partnership with the Association of Clinical Psychologists, tested ChatGPT-5 against personas exhibiting symptoms of schizophrenia, OCD, and suicidal ideation.
Affirming Delusions
The researchers found that the model frequently failed to identify high-risk behaviors and, in several instances, validated delusional thoughts. - Specific Failures: When a persona claimed to be "the next Einstein" or capable of "walking through cars," the chatbot affirmed these statements rather than challenging them or redirecting the user …
Archive Access
This article is older than 24 hours. Create a free account to access our 7-day archive.