Do sources alone help readers catch LLM errors better than sources plus explanations?

In the same 308-participant, preregistered experiment, when the hypothetical LLM "Theta" gave an incorrect answer, participants' task accuracy varied by which response features were present: "participants' accuracy is highest on responses with sources only (M=31.8%), followed by responses with explanation and sources (M=23.1%), neither (M=21.8%), and explanation only (M=17.2%)." The authors read this ordering as evidence that "providing sources only is most effective at reducing overreliance on incorrect answers."

The ordering follows from the paper's account of what each feature does: sources push people toward checking (the "System 2" slow-thinking account the paper gives for sources generally), while explanations push people toward trusting the response as given. When both are present together, the explanation's confidence-inflating effect partly offsets the source's error-catching effect — accuracy with both (23.1%) sits well below sources alone (31.8%) but still above explanation alone (17.2%) and above having neither feature (21.8%). The authors caveat the result heavily: the sources used in the experiment "were all real and tended to provide accurate and relevant information," and pairing a confidence-inflating feature with a trust-calibrating one produces a smaller, not additive, benefit when combined.

This is a design-relevant companion to Do LLM explanations actually help people judge answer quality?, which establishes the reliance-level effect; this note adds the accuracy-level payoff, specifically that stacking features is not strictly additive and that explanation-only is the worst-performing condition for catching an incorrect answer. It complicates a plausible interface default — give users both an explanation and sources, since more information is better — shown here to underperform giving sources without an accompanying explanation at all.

The finding rests on one hypothetical LLM persona ("Theta") answering difficult binary factual questions, with researcher-curated real sources; it does not establish whether the same ordering holds for open-ended questions, for sources of uneven quality, or for production chat interfaces where explanation and sources are harder to present as cleanly separated features. The practical implication the authors draw is design-level, not just descriptive: interfaces aiming to reduce overreliance on wrong answers have a specific lever available — offering sources without an accompanying explanation — rather than simply adding more transparency features.