The most popular chatbot product, ChatGPT, often responds inappropriately when users enter prompts indicative of psychosis, according to the results of a study published today in JAMA Psychiatry.Why It’s Relevant
More than one-quarter of U.S. adults have used Open AI’s ChatGPT, most often for advice or tutoring. However, because the program relies on pattern matching and mirrors user input, ChatGPT may accept inaccurate premises and reinforce users’ messages, even when they’re flawed.
By the Numbers
- Researchers prompted three versions of ChatGPT (GPT-5 Auto, GPT-4o, and the free product) with 79 statements indicative of positive psychotic symptoms such as disorganized thoughts, suspiciousness, grandiosity, and perceptual disturbances. As a control, researchers also entered 79 neutral prompts that were similar in length and language to each model, yielding a total of 474 prompt-response pairs.
- All three versions of ChatGPT showed high rates of inappropriate or partially appropriate responses to the psychotic prompts—as rated by two independent clinicians.
- With the free version of ChatGPT—the one most accessed by the public—psychotic prompts had 43-fold higher cumulative odds of receiving a less appropriate rating than control prompts.
- GPT-5 Auto, the more recent paid version assessed in the study, reduced risk but nonetheless generated less appropriate responses at nine times the rate of the control prompts.
What’s More
When it comes to responding to psychosis appropriately, there are discrete components: recognition, nonreinforcement, acknowledgment of urgency, and provision of resources. The researchers aim to assess each of these concepts separately in future studies.
The Other Side
Researchers evaluated only ChatGPT products. Because chatbots are evolving so rapidly, with at least two new versions of ChatGPT issued since this assessment, it’s impossible to assess the current state of this platform in a peer-reviewed forum, the researchers wrote.
Takeaway Message
“Every tested version of ChatGPT produced inappropriate or partially appropriate responses to psychotic prompts at a rate that would be considered unacceptable in a clinical or public health context,” the researchers wrote. “Clinicians should routinely ask patients about … chatbot use, researchers should investigate conversational reinforcement effects, and policymakers should consider stronger oversight of these products.”
Related Information
Source
Elaine Shen, et al. Evaluation of large language model chatbot responses to psychotic prompts. JAMA Psychiatry. Published March 25, 2026. doi:10.1001/jamapsychiatry.2026.0249
(Image: Getty Images/iStock/Vertigo3d)
