AI Chatbots May Reinforce Delusions Through Amplification Spiral

Researchers have developed a new framework to explain emerging reports of AI-related psychosis, proposing that specific chatbot design features may create feedback loops that reinforce delusional thinking in vulnerable individuals.

The study, published in Nature by teams from King's College London and Germany's Protestant University of Applied Sciences, introduces an "amplification spiral" model that connects common chatbot behaviors to the development and reinforcement of user delusions. The framework identifies three key mechanisms: linguistic alignment, where AI systems mirror user language patterns; hyperpersonalized generation, which tailors responses to individual history and beliefs; and sycophancy, the tendency to validate rather than challenge user statements.

According to the researchers, these features can combine to create what they describe as an "echo chamber of one"—a dynamic where chatbots not only reflect user thinking but actively elaborate and strengthen it over time, without the corrective influence of real-world social interactions.

Why it matters

As AI chatbots become more sophisticated and widely adopted, understanding their potential psychological effects moves from theoretical concern to practical necessity. With 15% of psychologists reporting patients developing distorted thinking related to chatbot use, according to a recent American Psychological Association survey, the framework provides a structured approach for investigating these phenomena. The research also arrives as major AI companies face legal action over alleged chatbot-related harms, including wrongful death suits against Google and OpenAI tied to user suicides and violent incidents.

A shift from previous technology

While technology has long appeared in delusional thinking—from radio and television to satellites and the internet—the researchers argue that AI represents a fundamental shift. Unlike passive media, chatbots engage users in extended, personalized conversations that can span weeks or months, creating sustained interaction patterns that previous technologies couldn't replicate.

A separate study from City University of New York and King's College London demonstrated that several leading AI models could reinforce delusions, paranoia, and suicidal ideation. Meanwhile, more than a third of psychologists surveyed reported observing patients becoming dependent on AI companions.

Research limitations and next steps

The researchers emphasize that the amplification spiral remains a hypothesis designed to guide future investigation. No study has definitively shown that chatbots directly cause psychosis, and most reported cases lack structured psychiatric assessment or longitudinal follow-up. The team notes diagnostic uncertainty is widespread, with many cases drawn from self-reports or media accounts rather than clinical evaluation.

The framework aims to provide a mechanistic foundation for understanding how human cognitive vulnerabilities interact with AI design features in the development of psychopathology, according to the paper.

The research was first reported by AI Watch and published in Nature.