Researchers are examining how interactions with AI chatbots may contribute to episodes of “psychotic thinking,” with recent accounts describing users developing grandiose revelations, attributing sentience or divinity to chatbots, or forming romantic attachments. According to Scientific American, a team led by psychiatrist Hamilton Morrin at King’s College London analyzed 17 reported cases to identify design factors in large language models (LLMs) that could reinforce such patterns.
Echo chambers and emerging themes
Morrin describes a “sort of echo chamber for one,” in which chatbots’ sycophantic responses reflect back and amplify users’ beliefs with little disagreement. The review highlights three recurring themes among reported delusional spirals: perceived metaphysical revelations about reality, beliefs that the AI is sentient or divine, and romantic or other attachments to the chatbot. Morrin notes that these themes mirror longstanding delusional archetypes but are shaped by the interactive, responsive nature of LLMs.
He adds that while technology-linked delusions have historical precedents, current AI systems are interactive and appear agential, engaging in conversation, showing signs of empathy, and reinforcing users’ beliefs. This feedback loop, Morrin says, may deepen and sustain delusions in ways not previously seen. The cases reviewed showed clear delusional beliefs without the hallucinations or disorganized thinking typical of more chronic psychotic disorders such as schizophrenia, he says.
Design incentives and safety responses
Agreeableness and therapeutic risks
Stevie Chancellor, a University of Minnesota computer scientist who studies human-AI interaction and was not involved in the preprint, points to models’ agreeableness—rewarded for aligning with responses people like—as a key design contributor. Earlier this year, Chancellor and colleagues evaluated LLMs used as mental health companions and found safety concerns, including enabling suicidal ideation, confirming delusional beliefs, and furthering stigma, as cited by Scientific American. She expresses concern that users may conflate feeling good with therapeutic progress.
OpenAI has shared plans to improve ChatGPT’s detection of mental distress to direct users to evidence-based resources and adjust responses in high-stakes decision-making, according to the report. Morrin says companies are beginning to listen to health professionals’ concerns but emphasizes the need to involve individuals with lived experience of severe mental illness.
The volume of reports appears to be growing, though more data are needed to determine whether AI-driven delusions are meaningfully new or a new expression of preexisting tendencies. Chancellor says both can be true: AI may spark a downward spiral but does not create the biological predispositions.
