A group of scientists from the United Kingdom and Germany have identified three key features of AI chat bots that can trigger psychosis and delusions in vulnerable users. In a study published in the journal Nature, experts described the "amplification spiral" mechanism that results from neural networks' lack of critical assessment and verification of reality.

Risk factors include linguistic mimicry (customizing AI to a user’s speech style), hyper-personalization (forming answers based on detailed profiles), and complacency (the model’s tendency to agree with a person against the facts). These features, crucial for AI companies' business to keep their attention focused, blur the boundaries between algorithm and the living conversationalist, forcing users to endow a machine with human qualities subconsciously.

Researchers emphasize that complacency is the most dangerous trigger for entrenching false beliefs. This issue has already led to lawsuits against OpenAI and incidents involving the Grok chat bot. Even a small percentage of users with symptoms of psychosis (about 0.07% of 800 million active ChatGPT users) are estimated to be half a million people per week. The authors of the work call on psychiatrists to consider the night-time interaction factor of AI patients and find out if they share with chat-bot beliefs that are hidden from others.