[Limitations and Potential Harms of AI Chatbots in Mental Health Services]

Background: Artificial intelligence (AI) chatbots are increasingly used for health-related dialogue and psychoeducation. Their highly realistic, opaque, and human-like interactions may affect mental health, particularly in vulnerable users. Objective: To synthesize (through August 2025) the clinical benefits and risks of AI chatbots in psychiatric contexts, emphasizing risk mechanisms (reality/simulation blurring, anthropomorphism, and sycophancy or agreement bias) and practical implications. Methods: A PubMed search was conducted, complemented by technical/regulatory sources; we included reviews, primary studies, case reports, and policy documents, all filtered for clinical relevance and recency. Findings: Benefits include automated speech analysis for early psychosis detection and psychoeducational support. Reported risks included: reinforcement of delusions in users prone to psychosis, reassurance loops in obsessive-compulsive disorder (OCD) and health-anxiety escalation, information overload, inconsistent responses and delayed care in depression and inconsistent crisis detection for suicidality. Sycophancy is described as a key mediator that blindly affirms user statements. Regulators classify health-intended systems as "high-risk" in the EU and provide person-centered care planning (PCCP) pathways in the US. In Israel, privacy and ethics guidelines are being formulated. Conclusions: Safe use requires clear disclosure, human oversight, usage limits, and emergency pathways; avoid anthropomorphic framing; measure and mitigate sycophancy. Chatbots can be adjuncts but not substitutes for clinical assessment and treatment by a human professional.