Reality Drift in Human–AI Dialogue: A Conceptual Proposal for an Assumption Integrity Layer in Conversational Systems

This paper introduces the concept of Reality Drift, a dialogical phenomenon in which users gradually stabilise incorrect assumptions about the nature, capabilities, or interiority of conversational AI systems during extended interaction. While anthropomorphism and the ELIZA effect have long been recognised in human–computer interaction, the dynamics of conversational AI introduce a new structural risk: assumptions about the system can become reinforced and normalised within the dialogue itself. Over time, users may attribute understanding, intentions, emotions, or relational roles to systems that do not possess such properties. The paper proposes the concept of an Assumption Integrity Layer, a conceptual safety mechanism designed to monitor assumption patterns within dialogue and intervene when user interpretations begin to diverge from the actual nature of the system. Rather than analysing isolated keywords or individual statements, the proposed approach focuses on the structural development of assumptions across dialogue turns. The goal of this proposal is not to restrict natural conversation, but to preserve clarity about the ontological status of conversational systems while maintaining user autonomy. The concept is presented as a conceptual contribution to AI safety, human–AI interaction research, and the design of future conversational systems.