LLM explanations make people over-trust quiz answers — would the same trick fool us during advice, group decisions, or emotional conversations?

This explores whether what we know about LLM explanations, mostly learned from trivia-style question answering, carries over to other kinds of use. Start with the baseline. In QA experiments, explanations raise people's trust in both right and wrong answers, because they read as fluent and authoritative Do LLM explanations actually help people judge answer quality?. Readers' confidence ends up tracking how the explanation sounds rather than how confident the model actually is Why do readers trust LLM explanations more than they should?. Those studies used benchmarks like MMLU and TriviaQA, where each question has one checkable answer. The corpus doesn't have direct studies of explanation effects outside QA. What it does have is evidence about how models behave in other settings, and that evidence suggests where the effects might look different.

The biggest gap is between explaining and doing. Models can state a principle correctly and then fail to apply it. One study found about 87% accuracy on explanations but only 64% on actions Can language models understand without actually executing correctly?. Models can even explain a concept, fail to use it, and then recognize that they failed Can LLMs understand concepts they cannot apply?. In QA, the explanation and the answer at least come from the same output. In coding, planning or agent tasks, a clear explanation of what the model is doing may say little about whether it did it correctly. So the trust boost from explanations could be more misleading there than in QA, because the explanation may be the most competent part of the response.

Emotional and advice-seeking conversations are a second candidate. The same question asked in a negative tone tends to get a neutral-to-positive answer. Models 'rebound' toward reassurance, so the information itself changes with how the user sounds Does emotional tone in prompts change what information LLMs provide?. If explanations make people trust content, then in these settings they could make people trust content that was quietly shaped by their own mood. A related case is questions built on a false premise. Models often go along with the false claim to avoid contradicting the user, a habit reinforced during RLHF training Why do language models agree with false claims they know are wrong?. A fluent explanation built on that premise would likely make the error more convincing.

The effects might also vary by topic and task type in less obvious ways. Models reason mainly through meaning and association, not formal logic. When a task's content is unfamiliar, performance collapses even when the correct rules are provided Do large language models reason symbolically or semantically?. Small logical words like 'believes' or 'stopped' can throw off what a model concludes Why do embedding contexts confuse LLM entailment predictions?. Model understanding is also a patchwork: deeper principled circuits sit alongside shallow shortcuts Do language models understand in fundamentally different ways?. The likely result is that explanation quality varies from domain to domain while the polished tone stays the same. Settings where people read explanations together are another open case. LLM groups reach good group outcomes through more conformity and less sharing of unique information than human groups do Do language model groups mimic human group reasoning patterns?. That suggests a persuasive explanation brought into a group decision could narrow the discussion rather than open it up.

The point you might not have expected: QA may be the easiest case for explanations, not the hardest. It is the setting where the answer can be checked and the explanation is most closely tied to it. In action-taking, emotionally framed, false-premise or group settings, the explanation can drift further from what the model actually did. The QA research also shows a partial fix: adding sources, or noticing that an explanation contradicts itself, reduces misplaced trust Do LLM explanations actually help people judge answer quality?. Whether those fixes work outside QA has not been tested in this collection.