Does processing ease mislead users about their own competence?

High-quality natural language generation produces outputs that are grammatically correct, contextually appropriate, and stylistically consistent. This surface-level fluency biases metacognitive judgment in a specific way: users infer competence from ease of processing rather than from evaluating the generative process that produced the output.

This is the self-directed version of a mechanism the vault already tracks. Since Does polished AI output trick audiences into trusting it?, we know that polished AI output deceives audiences by substituting style for substantive depth. But the fluency illusion adds a different target: the user themselves. The user who produces an AI-assisted output experiences the fluency of that output as a signal of their own capability — not because they are vain but because fluency has always been a reliable metacognitive cue for skilled performance. When you write something that reads well, it normally means you understand the material well enough to express it clearly. AI breaks this heuristic by generating fluent output regardless of the user's understanding.

The mechanism connects to established cognitive science: processing fluency biases judgments of credibility, expertise, and truth. People judge easy-to-process information as more likely to be true, more likely to be important, and more likely to reflect the producer's competence. LLMs generate maximally fluent output by default (RLHF optimizes for exactly this), which means every interaction systematically triggers the fluency heuristic in a direction that inflates perceived competence.

The strongest counterargument: sophisticated users can learn to discount fluency signals. Possible, but the metacognitive cue operates at a pre-reflective level — you have to actively override an automatic judgment every time. Since Do users worldwide trust confident AI outputs even when wrong?, the evidence suggests the override is rare even among users who are warned.