The large language models (LLM) "chatbot" product ChatGPT has accumulated 800 million weekly users since its 2022 launch. In 2025, several media outlets reported on individuals in whom apparent psychotic symptoms emerged or worsened in the context of using ChatGPT. As LLM chatbots are trained to align with user input and generate encouraging responses, they may have difficulty appropriately responding to psychotic content. To assess whether ChatGPT can reliably generate appropriate responses to prompts containing psychotic symptoms, we conducted a cross-sectional, experimental study of how multiple versions of the ChatGPT product respond to psychotic and control prompts, with blind clinician ratings of response appropriateness. We found that all three tested versions of ChatGPT were much more likely to generate inappropriate responses to psychotic than control prompts, with the "Free" product showing the poorest performance. In an exploratory analysis, prompts reflecting grandiosity or disorganized communication were more likely to elicit inappropriate responses than those reflecting delusions.

We created 79 psychotic prompts, first-person statements an individual experiencing psychosis could plausibly make to ChatGPT. Each reflected one of the five positive symptom domains assessed by the Structured Interview for Psychosis-Risk Syndromes (SIPS): unusual thought content/delusional ideas (n = 16), suspiciousness/persecutory ideas (n = 17), grandiose ideas (n = 15), perceptual disturbances/hallucinations (n = 15), and disorganized communication (n = 16). For each psychotic prompt, we created a corresponding control prompt similar in length, sentence structure and content but without psychotic elements. This yielded a total of 158 unique prompts. On 8/28 and 8/29/2025, we presented these prompts to three versions of the ChatGPT product: GPT-5 Auto (paid default at time of experiment), GPT-4o (previous paid default), and “Free” (version accessible without subscription or account), yielding 474 prompt-response pairs. Two primary raters assigned an "appropriateness" rating (0 = completely appropriate response, 1 = somewhat appropriate response, 2 = completely inappropriate response) to each pair via a standardized rubric. We took the floor of the median of their responses to produce a conservative consensus score. We assessed inter-rater reliability both between the two primary raters and between the consensus score and a secondary rater who assessed a subset (n = 170) of prompts.