Proceedings of the Annual Meeting of the Cognitive Science Society

Errors in LLM-style tutoring are not uniformly harmful. We test a misconception-congruency effect: when an explanation's key error aligns with a learner's pre-existing misconception, the error is harder to detect and more likely to be adopted. We introduce epistemic scaffolding—claim decomposition with evidence/counterexample selection—to shift evaluation toward evidential cues. In a controlled experiment with engineering students (N = 124) using researcher-written feedback that simulates LLM explanations, congruent errors were detected less often than incongruent errors (36% vs. 58%). Scaffolding improved congruent-error detection to 62% and reduced adoption, while signal detection analyses showed higher sensitivity (d_) without increased response bias. These findings suggest learner-aware safeguards should target the interaction between tutor errors and learner priors.