RLHF Fluency Optimization and the Activation Conditions of Evaluative Cognition

Existing frameworks for understanding the epistemic risks of large language models — including automation bias, deskilling, and cognitive offloading — share an unexamined presupposition: that users enter evaluative cognition when engaging with LLM outputs, and that the difficulty lies in what happens within or after that evaluative process. This paper examines the presupposition itself. It argues that reinforcement learning from human feedback (RLHF) operationalizes human processing fluency preference as an optimization target, producing outputs whose fluency may reduce the likelihood that discriminative evaluative processing is recruited. Drawing on the cognitive psychology of processing fluency, engineering studies of reward model bias, and recent experimental work on human-AI interaction, the paper introduces evaluative activation conditions as a distinct explanatory level — logically prior to and irreducible to trust miscalibration, skill degradation, or voluntary delegation. The contribution is not a demonstrated causal mechanism but an inferential architecture that specifies the constitutive links of the proposed account so that each becomes amenable to targeted empirical investigation. The most directly testable link is the claim that semantic and structural fluency affect evaluative activation through the same metacognitive pathway as perceptual fluency — a prediction the paper specifies experimentally by manipulating fluency independently of content accuracy.