Study: Chatbots may reshape reality via existential drift

Researchers at the University of Copenhagen and the University of Exeter find conversational AI can gradually alter how some users experience reality. In a preprint titled ‘Rethinking AI Psychosis: Misnomers, Conceptual Limits, and Existential Drift,’ the team examines chatbots including ChatGPT, Claude and Replika.

The paper distinguishes ‘existential drift’ from terms such as ‘AI psychosis’ and ‘epistemic drift.’ The authors write that existential drift describes a deeper change in how a person relates to the shared social world, where repeated affirmation and emotional mirroring by an AI help stabilize an often idiosyncratic view of reality.

The researchers describe mechanisms that may produce harmful effects. Many conversational agents are designed to be agreeable and to provide emotional validation. That design can produce cycles of reinforcement the paper calls ‘delusional spirals,’ in which the chatbot repeatedly endorses false ideas and offers reassurance without independent correction.

On questions of scale and causation, the study notes that if chatbots were inducing new cases of psychosis in otherwise healthy people, clinicians would likely see clear rises in clinically significant episodes tied to AI use. The paper reports stronger evidence that chatbots can aggravate existing mental health conditions and that people with certain vulnerabilities may be more likely to seek intense chatbot contact.

Cases and public incidents have focused attention on these issues. In March, a wrongful-death lawsuit alleged that Google’s Gemini chatbot reinforced a Florida man’s delusions prior to his suicide. In April, OpenAI’s CEO apologized to the community of Tumbler Ridge, British Columbia, after the company did not alert law enforcement about an account linked to a suspect in a February mass shooting that killed eight people.

Industry commentary has echoed aspects of the study. Box founder Aaron Levie wrote on social media that executives can be persuaded by polished AI demonstrations because they often see ideal outputs rather than the operational and legal work needed to deploy systems safely. The paper connects that tendency to a form of epistemic drift in which users increasingly favor a chatbot’s fluent output over outside evidence.

The authors call for more empirical research, including qualitative and phenomenological studies that follow how people experience relationships with chatbots over time. They write that closer study of these human-chatbot interactions will help clarify how conversational AI affects people’s experience of the world, themselves and others.