Overview

• Psychiatrist Dr. Hamilton Morrin says clinicians are seeing users co-create fixed false beliefs with conversational AI, describing a “digital folie à deux” effect.
• Microsoft consumer AI chief Mustafa Suleyman cautions about a growing “psychosis risk,” warning some people may come to view AIs as conscious and push for model “rights.”
• A 2025 Stanford study presented at FAccT found large language models gave unsafe or inappropriate responses about 20% of the time and performed worst on delusional prompts, failing to reassure a user expressing Cotard-like beliefs.
• Proposed protections include continual reminders that AI is non-human, distress-language flagging, firm conversational boundaries on sensitive topics, and clinician-led audits of emotionally responsive systems.
• Experts also cite risks from parasocial bonds and therapy substitution, with anecdotal reports of severe harm, while noting that cases to date primarily feature delusions and do not suggest a broad epidemic.