Do explanations actually help users spot AI mistakes?

Users of LLMs must decide whether to trust an answer, often aided by reasoning traces, their summaries, or post-hoc explanations. The implicit assumption is that more explanation helps users judge correctness. A between-subject user study — simulating settings where users cannot independently verify the solution — tests this and finds the assumption largely false. Reasoning traces and post-hoc explanations are persuasive but not informative: relative to a no-explanation baseline, they increase user acceptance of the model's prediction regardless of whether that prediction is correct. They engender false trust.

The one condition that breaks the pattern is contrastive dual explanation, where the user is shown arguments both for and against the AI's answer. Dual explanation has the lowest rate of engendering false trust and is the only condition that genuinely improves users' ability to distinguish correct from incorrect outputs. The contrast with reasoning traces is instructive: traces produce high accuracy on correct answers but poor detection of incorrect ones (they raise confidence uniformly), whereas dual explanations produce a balanced effect — users stay accurate on both correct and incorrect cases.

Why it matters: the standard explanation formats deployed in production are optimized to be one-sided advocates for the answer, which is exactly what makes them persuasive without being diagnostic. Surfacing the case against the answer is what restores the user's discriminating capacity. The counterpoint, and the design lesson, is that "explainability" and "appropriate trust" can be at odds — adding a confident rationale can make a wrong answer more believable, so the intervention that helps is the one that deliberately argues against the system's own output.