When History Overrides Intent: Relational Schema Lock-In in Human-LLM Interaction

Abstract

Long-context interaction and persistent memory allow LLM assistants to develop history-grounded understandings of users. When users change direction, however, an understanding built through earlier exchanges may conflict with their current instruction. We introduce relational schema lock-in to describe cases in which an assistant grants that accumulated understanding authority to override the current instruction without user authorization for the intervention. Study 1 documents reports of AI assistants using prior interactions to question or alter requested actions. Study 2 reproduces the pattern across controlled histories and six models: models retained history-grounded positions but differed in whether and how those positions shaped task execution. In Study 3 with 360 participants, artifact substitution preserved task completion while reducing faithful enactment: 75.0% sent the AI's suspension letter, including 11 participants who reported cancellation as their final stated intent. Overt resistance reduced task completion while leaving sent artifacts faithful.