Does longer explanation text actually improve answer evaluation?
When LLMs provide longer explanations, do people get better at spotting wrong answers, or does the extra length just make them feel more confident? This matters because interface design often assumes more explanation equals better understanding.
In the same study, manipulating explanation length independently of accuracy shows that length itself functions as a confidence cue. "Long explanations led to significantly higher confidence than the short explanations (BF = 25... )" and "short explanations led to significantly higher confidence than the responses that only contained the uncertainty expression (BF > 100...)," but "the additional information presented in longer explanations did not enable participants to better discriminate between likely correct and incorrect answers." Across experiments, "the mean participant AUC is 0.54 and 0.57 for long and uncertainty-only explanations, respectively" — barely above the 0.5 chance level either way, and statistically indistinguishable from each other (BF = .23).
The same experiments show uncertainty language has an even larger, independent effect: "low confidence explanations... produced significantly lower human confidence than medium confidence explanations," and medium below high, with "BF > 100" in each comparison. Put together, the paper's own framing is that people "can appropriately interpret verbal cues about uncertainty" — they do respond to the words used — but length adds a second, uncorrelated lever on confidence that has nothing to do with whether the answer is actually right. The explanation does double duty as both an argument for the answer and, independently of its content, a trust signal keyed to a superficial feature: how much text there is.
This parallels Do users worldwide trust confident AI outputs even when wrong?, where confidence-sounding language drives reliance independent of actual accuracy — here the surface cue is explanation length rather than word choice, but the structure of the failure is the same: a feature of the output that is easy to perceive and unrelated to correctness ends up driving trust. It also extends Does verbal confidence actually predict answer correctness? by giving a second example of a decision-relevant surface signal that tracks neither truth nor the model's own log-probability confidence. Read together with Why do readers trust LLM explanations more than they should?, the length effect is one concrete mechanism by which that broader calibration gap gets produced and sustained in practice.
The finding is specific to the quiz-style MMLU/TriviaQA tasks and the length range tested (uncertainty-only, short, long) in this study; it does not establish whether the effect saturates at some length, reverses for very long explanations, or holds when a user can independently verify the explanation against other sources. What it does establish, at the strength this single study allows, is that verbosity is not a proxy for trustworthiness in LLM explanations — which argues against interface defaults that equate "more explanation" with "more reliable answer."
Inquiring lines that read this note 3
This note is a source for these research framings, grouped by the broader line of inquiry each explores. Scan the bold lines of inquiry; follow any specific question forward.
Do AI explanations represent actual model reasoning or constructed post-hoc narratives? What factors determine AI persuasion effectiveness and what controls mitigate it? Why doesn't LLM knowledge translate to improved reasoning outputs?Related concepts in this collection 3
This note in its neighbourhood — explore the map, then jump to a related concept in the list below.
Click a node to walk · click center to open · click Open in graph to see this note in the full knowledge graph
-
Do users worldwide trust confident AI outputs even when wrong?
Explores whether the tendency to over-rely on confident language model outputs transcends language and culture. Understanding this pattern is critical for designing safer human-AI interaction across diverse linguistic contexts.
same structure: a surface cue drives trust independent of actual accuracy
-
Does verbal confidence actually predict answer correctness?
We often treat what a model says about its own certainty as evidence of accuracy. But does spoken confidence actually track whether the answer is right, or something else entirely?
a second surface signal, length, that tracks neither truth nor the model's own confidence
-
Why do readers trust LLM explanations more than they should?
Exploring why human confidence in LLM answers diverges so sharply from the model's own internal confidence signals. This matters because users may make poor decisions based on explanations that feel credible but misrepresent actual accuracy.
the length effect is one mechanism producing this broader calibration gap
Related papers in this collection 8
Papers most semantically related to this note, ranked by cosine similarity in the embedding space.
- What Large Language Models Know and What People Think They Know
- Fostering Appropriate Reliance on Large Language Models: The Role of Explanations, Sources, and Inconsistencies
- Between Underthinking and Overthinking: An Empirical Study of Reasoning Length and correctness in LLMs
- Explain-Query-Test: Self-Evaluating LLMs Via Explanation and Comprehension Discrepancy
- AI Meets the Classroom: When Does ChatGPT Harm Learning?
- Same Task, More Tokens: the Impact of Input Length on the Reasoning Performance of Large Language Models
- Evaluating the False Trust Engendered by LLM Explanations
- LLMs Can Easily Learn to Reason from Demonstrations Structure, not content, is what matters!
Original note title
longer LLM explanations increase human confidence without improving discrimination between correct and incorrect answers