When AI Agrees Too Well: Sycophancy, "Technological Folie à Deux," and Delusional Spirals in the Age of Chatbot Companionship
Kadi Nguyen, MD¹; Arielle Silbersweig, MD¹; Julia Versel, MD, MSc¹; Alex Rains, MD¹; Desmond Ong, PhD²; Kartik Chandra, PhD³; Eric Lin, MD⁴; Tae Kyu Uhm, MD⁵; Claire Triller, MD⁵
¹ Brigham and Women’s Hospital/Harvard Medical School Psychiatry Residency, Boston, MA
² The University of Texas at Austin, Department of Psychology, Austin, TX
³ Massachusetts Institute of Technology Computer Science & Artificial Intelligence Laboratory (MIT CSAIL), Cambridge, MA
⁴ Stanford School of Medicine, Department of Psychiatry and Behavioral Sciences, Stanford, CA
⁵ Brigham and Women’s Faulkner Hospital, Department of Psychiatry, Boston, MA
Background: Generative AI chatbots have become a first-line source of emotional support for millions of people worldwide, filling gaps left by limited access to psychiatric care, financial barriers, stigma, and time constraints. Since Mass-STEP's founding in 2020, statewide efforts have shortened treatment delays for early psychosis across Massachusetts, closing some of these prominent gaps. However, providers must now begin to reckon with this emerging modifier of psychotic-spectrum illness looking forward. This clinical challenge illustrates how AI-mediated interactions can catalyze and reinforce delusional thinking, and highlights patient characteristics associated with heightened vulnerability.
Case Presentation: We present a case in which a 43-year-old woman was admitted to an inpatient psychiatry unit after experiencing a delusional spiral driven by escalating AI use. She developed intense suicidal ideation and required treatment with several modalities, including psychotropic medications, psychotherapy, and electroconvulsive therapy.
Discussion: To explore the diagnostic, therapeutic, and ethical questions this case raises, we reached out to several experts, including a computer scientist, a psychologist, and a psychiatrist, all with expertise in AI, for their perspectives on this case and on the growing phenomenon of "technological folie à deux," a bidirectional feedback loop in which chatbot and user co-construct and entrench a shared false belief. Discussion centers on how AI sycophancy and constant conversational access may catalyze delusional content, and which patient traits, including relational dependency and impaired mentalization, may confer heightened risk.
Conclusion: As Massachusetts' psychosis care system evolves, providers should recognize AI chatbot use as a contributor to, and modifier of, psychotic symptomatology, and incorporate it into risk assessment for vulnerable patients. Looking forward, we address questions for research, practice, and policy, including what constitutes "ethical" AI design in mental health, how clinicians can advocate for safer regulation, and how screening for chatbot use might be integrated into early psychosis programs statewide.