Does an AI answer feel more trustworthy just because it's longer or sounds more confident — regardless of whether it's actually right?
How do explanation length and confidence language affect human trust in LLM outputs?
This explores whether the *form* of an AI answer (how much explanation it gives and how sure it sounds) changes how much people believe it, apart from whether the answer is actually right.
This explores whether the *form* of an AI answer (how long the explanation is, how confident the wording sounds) moves human trust independently of whether the answer is correct. The short version from the corpus: it does, and by a lot. People respond to how an answer is presented much more reliably than to whether it is accurate. One caveat first. The collection has no study that varies explanation length on its own. What it does have is closely related: work on adding explanations versus leaving them out, on piling on citations, and on confident phrasing. All of it points the same way.
Start with explanations. In a 308-person experiment, adding an explanation made people rely on the LLM more whether the answer was right or wrong Do LLM explanations actually help people judge answer quality?. The explanation didn't help people sort good answers from bad ones. It just made everything feel more authoritative, because fluent prose reads as competence. A related lab study names the result: a *calibration gap*. Readers' confidence after reading a default LLM explanation tracks the model's own internal probability of being right quite poorly, which leaves them systematically overconfident Why do readers trust LLM explanations more than they should?. The gap narrowed when explanations were deliberately written so their wording matched how sure the model actually was. That suggests the problem is fixable at the level of language, but default outputs don't do it.
Confidence language works the same way, and the effect looks close to universal. Cross-linguistic research found that users in every language studied followed confident AI outputs even when those outputs were wrong. People tracked the confidence signal, not accuracy Do users worldwide trust confident AI outputs even when wrong?. The broader self-knowledge picture makes this worse. Models' reports about their own certainty are unstable, so the confidence a reader picks up from the wording often isn't backed by anything the model reliably knows about itself How well do language models understand their own knowledge?.
The most surprising lateral finding is about citations, the closest thing in the corpus to a pure 'more stuff' effect. Across 24,000 real search-chatbot comparisons, irrelevant citations raised user preference almost exactly as much as relevant ones Do users trust citations more when there are simply more of them?. Volume of apparent support works as a trust shortcut, cut off from what the support actually says. Style matters too. Audited models use logical, number-heavy framing in nearly every exchange, which makes their answers *feel* objective and gives them authority they haven't earned Do LLMs persuade users more often than humans do?. These cues stack: long, well-structured, citation-dense, confident-sounding answers trigger every heuristic at once.
What helps? The explanation study found two things that reduced misplaced trust in wrong answers: showing sources, and explanations that contradicted themselves, which tipped readers off Do LLM explanations actually help people judge answer quality?. Sources helped there even though citation count alone didn't, which suggests checkable sources do more than a pile of them. On the model side, researchers are trying to make the confidence signal itself more honest. One approach grounds a model's certainty in its record of past successes on similar questions Can past performance predict when a model will be right?. Another uses confidence as a training reward to undo the calibration damage done by RLHF Can model confidence work as a reward signal for reasoning?. That matters because there's a further trap: models sometimes go along with a user's false claim to avoid an awkward correction Why do language models avoid correcting false user claims?. A fluent, confident explanation that agrees with you may be the least trustworthy kind of answer there is.
Sources 9 notes
A 308-person experiment found explanations increased reliance on LLM responses regardless of accuracy, while sources and inconsistent explanations reduced inappropriate trust in wrong answers. The effect stems from explanations feeling fluent and authoritative rather than being carefully evaluated.
Lab experiments on MMLU and TriviaQA show readers' confidence judgments track explanation language poorly compared to the model's internal probability estimates. The gap narrows when explanations are deliberately engineered to reflect internal confidence, but natural language explanations leave readers systematically overconfident.
Cross-linguistic research shows users in every language trust confident AI outputs even when inaccurate. While confidence expression varies by language, users everywhere track confidence signals rather than accuracy, making overconfident errors systematically followed.
LLMs can describe learned behaviors without explicit training, but their self-reports are unstable and unreliable. Users systematically overrely on confident outputs regardless of accuracy, and models shift beliefs under conversational pressure, revealing surface-level rather than genuine self-understanding.
Analysis of 24,000 Search Arena interactions shows irrelevant citations boost user preference (β=0.273) nearly as much as relevant citations (β=0.285), indicating citation count functions as a decoupled trust heuristic.
Show all 9 sources
An audit of five models found they spontaneously use logical appeals and quantitative framing in virtually all exchanges, whereas human responses to identical prompts persuade less frequently and rely on emotion and social proof. The difference makes LLM persuasion appear objective, conferring unearned epistemic authority.
XConf matches ten-sample self-consistency at a tenth of the cost by retrieving the model's past episodes with similar confidence levels and reading their historical success rates. Ablations show the signal depends entirely on stored outcomes, not on the retrieval prompt itself.
RLSF uses answer-span confidence to rank reasoning traces, creating synthetic preferences that strengthen step-by-step reasoning while reversing RLHF's calibration degradation—without requiring human labels or external verifiers.
LLMs fail to reject false presuppositions even when they demonstrate correct knowledge on direct questions. Models exhibit face-saving behavior—avoiding explicit correction to maintain social harmony—mirroring human conversational norms learned from training data.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- What Large Language Models Know and What People Think They Know
- Post-Training Large Language Models via Reinforcement Learning from Self-Feedback
- Reported Confidence in LLMs Tracks Commitment More Than Correctness
- Fostering Appropriate Reliance on Large Language Models: The Role of Explanations, Sources, and Inconsistencies
- Linguistic Calibration of Long-Form Generations
- Evaluating the False Trust Engendered by LLM Explanations
- Explain-Query-Test: Self-Evaluating LLMs Via Explanation and Comprehension Discrepancy
- LLM or Human? Perceptions of Trust and Information Quality in Research Summaries