Token Vibes: Translating User Perceptions of AI Behavior into Mechanistic Constructs

Token Vibes: Translating User Perceptions of AI Behavior into Mechanistic Constructs Description: This preprint introduces the Token Vibes Translation Framework, a conceptual and semi-formal method for translating subjective user experiences of interacting with large language models (LLMs) into mechanistic interpretations and associated risk profiles. Users frequently describe AI outputs using anthropomorphic or experiential language (e.g., “hallucinating,” “sycophantic,” “creepy,” “off”). While these descriptors are intuitively meaningful, they often obscure the underlying computational processes and can contribute to misinterpretation, over-attribution of agency, and instability in trust calibration. This work proposes that the issue is not the use of subjective language itself, but the absence of a consistent translation layer connecting those experiences to system-level behavior. The framework operationalizes this translation through a three-part structure: User Perception → Mechanistic Interpretation → Risk Profile, allowing experiential reports to be preserved while enabling technical clarity and evaluability. In addition, this paper introduces SMSTEP (Superfluous Model-Side Token Expenditure Percentage), a preliminary metric for evaluating signal inefficiency in model outputs. SMSTEP quantifies the proportion of generated tokens that do not contribute to task-relevant meaning, offering a practical lens for analyzing verbosity, redundancy, and interactional degradation over time. The constructs presented here emerged from sustained, high-frequency interaction across multiple LLM systems and reflect an attempt to formalize “power user” perceptual sensitivity into a communicable framework. The included glossary expands these concepts into a broader taxonomy of interaction patterns, including phenomena such as downstream token malfeasance (DSTM), token drift, context collapse, and token hygiene. This preprint intentionally occupies a hybrid space between formal research and exploratory conceptual work. While grounded in existing literature on alignment, anthropomorphism, and cognitive load, it also incorporates informal and translational terminology to capture real-world interaction dynamics more precisely than strictly technical language often allows. This is a non–peer-reviewed, exploratory preprint intended to provoke discussion, offer shared language for describing AI interaction phenomena, and bridge the gap between user experience and mechanistic understanding. All interpretations, frameworks, and conclusions are the responsibility of the author.