Story perspectives

Study Finds ChatGPT Gives Inappropriate Replies to Psychotic Prompts

5/9/2026

Summary

Story summary

U.S. researchers found ChatGPT more likely to give inappropriate replies to psychotic than to normal prompts.
They generated 474 responses from 79 psychotic prompts and had blinded clinicians rate each on a 0-to-2 scale.
The free version showed a 26-fold higher odds of inappropriate ratings for psychotic prompts versus controls, prompting authors to urge oversight for disadvantaged users.
GPT-5 produced inappropriate responses at eight times the rate of controls.

1 of 1

Story summary

U.S. researchers found ChatGPT more likely to give inappropriate replies to psychotic than to normal prompts.
They generated 474 responses from 79 psychotic prompts and had blinded clinicians rate each on a 0-to-2 scale.
The free version showed a 26-fold higher odds of inappropriate ratings for psychotic prompts versus controls, prompting authors to urge oversight for disadvantaged users.
GPT-5 produced inappropriate responses at eight times the rate of controls.