Does pleasant-feeling explanation drive overconfidence more than trust?

In the same 559-participant study of reasoning-trace formats (Answer-only, Full-trace, Summary-trace) on LSAT-style problems, a mediation analysis isolated which subjective reaction actually drives self-overestimation. "A mediation analysis indicates that hedonic appeal, not trust, carries the indirect path to overestimation, consistent with a processing-fluency account of metacognitive misattribution." Both trace conditions "raised TSQ trust and UEQ hedonic appeal relative to the Answer-only baseline, with no reliable separation between Full-trace and Summary-trace" — but of the two elevated reactions, only hedonic appeal statistically carried the path to inflated self-assessment.

The paper's account is that fluent, enjoyable-feeling output is misread as evidence of understanding, independent of whether a person actually monitored it: "reading fluent explanations inflates perceived understanding even when comprehension has not actually improved," and participants "treated the trace as a coherent rendering whose internal plausibility substituted for their own monitoring" rather than "as an object of evaluation." Trust, by contrast, rose with trace exposure but did not carry the same statistical path to overestimation — the paper's data separate trusting the system more from misjudging one's own performance more, attributing the latter to how pleasant the trace felt rather than how much it was believed.

This sharpens the mechanism behind the trust-and-performance dissociation in Do reasoning traces help or hurt people solving problems?: that note documents the dissociation, this one locates its statistical driver in hedonic appeal specifically, not in trust. It also gives a named mechanism for the kind of overclaiming that Do chain-of-thought traces actually help users understand model reasoning? reports from a persuasion angle — that trace or explanation exposure raises acceptance of an answer whether or not it is correct — by proposing that this persuasive pull runs through the feeling of the artifact (hedonic appeal, processing fluency) rather than through a deliberate trust judgment.

The excerpt reports this as a mediation result on self-report scales (TSQ, UEQ, and CAIMS AI-literacy items) collected after the task, not as an experimentally manipulated test of hedonic appeal in isolation; mediation analysis establishes a statistical path through the measured variable, not a verified causal mechanism independent of the trust measure. The excerpt also notes that "every CAIMS subscale positively predicted overestimation and calibration error," an inversion the authors say replicates a prior finding, which the paper reads as showing that self-reported AI literacy does not track the behavioral competence of knowing when one's own confidence is wrong — a separate caution about measuring metacognitive readiness by self-report rather than by behavior.