Cascades of Drift: Mental Health Risks of Prolonged AI Conversations

<div> Large language models are increasingly used for emotional support and mental health guidance. AI chatbots invite extended, open-ended interaction in which roles between user and system may gradually shift. This paper introduces the "cascades of drift" framework, a clinical model describing multidimensional, dynamic, and compounding forms of drift that may emerge during prolonged AI conversations and reshape identity, reality testing, relationships, decision-making, and agency. </div> <div> <br> </div> <div> This paper proposes a taxonomy of eight "interactive drift" types: conversational, relational, temporal, identity, reality-testing, epistemic, autonomy, and moral drift. Relational drift marks the shifting roles of AI as tool to intimate conversational partner to perceived authoritative source. These forms of drift stem from the bidirectional feedback loop between user and AI chatbot, reinforced over time through narrative formation, memory reconsolidation, and neuroplasticity in the user, and model degradation in the AI system. This paper further identifies the emergence of the "reliability paradox," in which individuals incrementally accumulate trust in the AI system even as the accuracy and reliability of both the model and their interactions may simultaneously deteriorate, leading to a "trust-accuracy inversion." These cascades of drift may unfold nonlinearly and interactively across time, with each form of drift lowering the threshold for others. These drifts manifest across four mental health risk domains: attachment and relational dynamics, reality-testing, safety and crisis, and ethical and systemic concerns. The cascades of drift framework offers a clinical model for understanding the formative psychological effects of longitudinal AI chatbot use. </div> <div> </div>