What question did this study set out to answer?

The central aim is to establish AI fixation and its consequences during human interaction with language models.

January 20, 20260 citationsOpen Access

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

Key Points

The central aim is to establish AI fixation and its consequences during human interaction with language models.
Defined AI fixation and drift concepts
Identified Undetected Constraint Contamination (UCC)
Developed metrics such as AFI, AFI-slope, FLR, CDR
Provided preregistration-ready methods for empirical assessment
AI fixation was shown to alter LLM responses based on small user-supplied attributes
Recurrent fixation episodes led to a stable attribute-centered response framework
Undetected constraints were identified as significant influences on inference

Abstract

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:

问 AI

Bookmark

View Full Paper