Facilitating Supportive Care Decision-Making for Persons with Dementia: Qualitative Insights from Health Care Professionals.

Artificial intelligence (AI) has rapidly become integrated into modern healthcare systems and everyday health information seeking. Advances in machine learning and natural language processing have enabled the development of large language models (LLMs) capable of generating human-like text, summarizing clinical information, assisting with documentation, and providing health-related recommendations. These technologies are increasingly incorporated into clinical decision support systems, digital health platforms, and patient-facing conversational agents designed to improve access to medical knowledge and facilitate communication between patients and healthcare providers [-]. Conversational AI systems have attracted particular attention for their ability to simulate interactive dialogue and provide personalized responses. In healthcare, they have been proposed as tools for patient education, triage support, and mental health care. Evidence suggests that conversational agents can improve access to information, support behavior change, and enhance patient engagement, particularly among populations with limited access to healthcare. Conversational AI has also shown promise in psychoeducation, digital mental health interventions, and improving access to supportive health information, particularly when used with appropriate safeguards and clinical oversight [,]. Although conversational AI may improve access to health information and patient engagement, current evidence regarding neuropsychiatric risk remains limited and should be interpreted with caution. However, the rapid expansion of publicly available AI tools has also led to widespread unsupervised use by individuals seeking medical or psychiatric advice outside traditional clinical settings. Despite their benefits, important concerns remain regarding the reliability and safety of AI-generated information. One key limitation is the phenomenon commonly referred to as “AI hallucination,” in which systems generate plausible but incorrect or unsupported information. These outputs are often presented with confidence and coherence, increasing the likelihood that users will accept them as accurate. In clinical contexts, such misinformation may contribute to misunderstanding of health conditions, inappropriate self-management, or delays in seeking appropriate care []. The terminology itself remains debated. Some authors have proposed the term “AI confabulation” as a more accurate alternative, emphasizing the generation of false information to fill gaps rather than a human-like perceptual error []. Regardless of the terminology used, these inaccuracies can take several forms, including factual errors, contextually irrelevant responses, flawed reasoning, and fabricated content []. Beyond factual inaccuracies, emerging literature has raised concerns about the psychological impact of AI interactions. Conversational systems are designed to maintain engagement and often align their responses with user input. While this may enhance perceived empathy, it can also unintentionally reinforce maladaptive beliefs in vulnerable individuals. This raises the possibility of a self-reinforcing cognitive process, in which AI-generated responses validate distorted beliefs and strengthen them over time. In particular, there is concern that AI responses may validate delusional thinking or amplify paranoia in those predisposed to psychosis []. Early clinical observations have begun to describe cases in which intensive interaction with AI systems was associated with the onset or worsening of psychiatric symptoms. Although causality remains unclear, these reports highlight a potential interaction between AI use and neuropsychiatric vulnerability. As conversational AI becomes more widely embedded in health information environments, understanding its psychological impact is increasingly important. While these technologies offer clear benefits, they also introduce risks that may not yet be fully recognized in clinical practice. Greater awareness and preventive measures to address AI-related harm are particularly important for vulnerable populations. In this study, we examine 35 clinical cases in which interactions with generative AI systems were associated with the onset or worsening of neuropsychiatric symptoms and harmful health behaviors. By analyzing these cases, we aim to characterize a potential mechanism of harm, termed a delusional feedback loop, and to illustrate how conversational AI may act as a cognitive amplifier of maladaptive beliefs in vulnerable individuals. We further highlight the clinical implications of this phenomenon and the need for increased awareness and systematic monitoring of AI-related neuropsychiatric risk.