AI Fixation → AI Drift: Attribute Dominance, Undetected Constraint Contamination, and Serial Feedback Dynamics in Human–LLM Interaction

This paper proposes AI fixation (McQueen, 2026) as a distinct prompt-sensitivity phenomenon: small user-supplied attributes—especially evaluative or epistemic qualifiers—can dominate an LLM’s interpretation and response stance even under neutral user intent, shifting outputs toward judgment, reassurance, or justification. The paper formalizes AI drift as a downstream, interaction-level multi-turn process in which repeated fixation episodes compound into a stabilized attribute-centered frame. A central mechanism is Undetected Constraint Contamination (UCC), wherein casual adjectives function as implicit constraints that remain unnoticed while shaping inference. The manuscript provides preregistration-ready methods and metrics (AFI, AFI-slope, FLR, CDR) and recommends scientific-method prompting as scalable mitigation.Released under CC BY 4.0. Reuse, distribution, and adaptation are permitted for any purpose, provided appropriate credit is given to the author and source.Please cite as: McQueen, K. K. (2026). AI fixation → AI drift: Attribute dominance, undetected constraint contamination, and serial feedback dynamics in human–LLM interaction (Version 1) [Preprint]. Zenodo. DOI: Concept DOI (always latest): https://doi.org/10.5281/zenodo.18279192