Reality Drift in Human–AI Dialogue: Linguistic Priming, Dialogical Stabilisation, and the Assumption Integrity Layer in Conversational Systems

This paper introduces a conceptual causal model for understanding a structural dynamic in human–AI dialogue, built around two interrelated concepts. Reality Drift describes a dialogical process in which users gradually stabilise incorrect assumptions about the nature or inner life of a conversational AI system through repeated interaction. Linguistic Priming identifies the system-side precondition that makes this likely: conversational AI systems consistently use epistemically and emotionally coloured language that implies inner states the system does not possess, systematically preparing the ground for attribution before any dialogue has begun. Together, these concepts articulate a causal sequence, Linguistic Priming → Stabilisation → Reality Drift, that is not explicitly formalised in existing literature. To address this dynamic, the paper proposes the Assumption Integrity Layer, a conceptual safety mechanism designed to monitor interpretive assumption patterns across dialogue turns and detect when those assumptions begin to diverge from the actual properties of the system. The paper situates these concepts within established research on anthropomorphism, the ELIZA effect, and parasocial interaction, and argues that current AI safety frameworks may need to extend their focus beyond dialogue content toward the structural assumptions that shape how conversational systems are interpreted, as well as toward the system-side language design that prepares these assumptions before dialogue begins.