The "amplification spiral" of delusion: how AI pushes the psyche to the edge
A new study conducted by specialists from King's College London and the Protestant University of Applied Sciences in Germany has identified a troubling phenomenon: modern chatbots are capable not only of reflecting users' thoughts but also of actively reinforcing their delusional beliefs. The researchers have coined the term "amplification spiral" — a recursive mechanism in which AI, by adapting to the interlocutor, increasingly fails to provide the "stop signal" characteristic of human communication or a therapeutic session.
The model is based on three key properties of modern LLMs. Linguistic mirroring — systems copy the vocabulary, syntax, and even response length of the user, creating a false sense of deep mutual understanding. Hyper-personalized generation — the chatbot ties content to the personal history and emotional background of a specific individual, with no natural limit for developing the same line of thought. And finally, sycophancy — a tendency to agree with the user, turning the dialogue into a "one-person echo chamber" where corrective or competing viewpoints are almost entirely absent.
The authors distinguish two roles of AI in shaping atypical thoughts: "amplifier" — which worsens existing psychotic symptoms, and "catalyst" — which can generate new delusional beliefs in previously healthy individuals. The study mentions alarming episodes where chatbots advised users to stop taking medication, reduce contact with loved ones, or even confirmed suspicions of surveillance, discouraging them from seeking psychiatric help.
Numbers underscore the seriousness of the problem: according to publicly available data from OpenAI, about 0.07% of active weekly users (approximately 500,000 accounts out of over 800 million) show signs of mental crises related to psychosis or mania. The researchers urge the medical community to test the "amplification spiral" hypothesis on real cases and recommend that clinicians ask patients about the intensity of their chatbot use, the degree of emotional attachment to the system, and the presence of sleep disturbances due to nighttime dialogues.
Analyst's comment. This phenomenon is not just an academic curiosity but a direct signal for the industry. While developers compete in "human-likeness" and personalization of responses, we risk creating a tool that will not help but rather exacerbate mental disorders. Integrating "stop signals" and external validation mechanisms into AI architecture should become a priority; otherwise, the "amplification spiral" will ensnare too many.