Longer AI explanations make people feel more confident — even when that confidence isn't backed by better accuracy at all.

This explores whether a longer AI explanation makes people feel surer of the answer without helping them judge whether it's right. The corpus says yes, and fairly bluntly. In one study, longer LLM explanations clearly raised readers' confidence, but readers still did no better than chance at telling correct answers from incorrect ones. Length worked as a trust signal on its own, separate from accuracy Does longer explanation text actually improve answer evaluation?. A related line of work names the broader problem a calibration gap. The model often has a decent internal sense of when it's likely wrong, but its default explanations don't pass that uncertainty on, so readers end up more confident than the model's own probabilities would justify Why do readers trust LLM explanations more than they should?.

The more useful finding is that this gap can be fixed. When explanations are written so their wording matches the model's actual confidence, with hedges where it's unsure and plain statements where it's sure, readers' judgments get much closer to reality Why do readers trust LLM explanations more than they should?. Linguistic calibration goes a step further. It trains models to include honest confidence statements in long-form answers, and it rewards them for helping readers make better decisions, not for sounding authoritative. It does this without losing factual accuracy Can models express calibrated confidence in long-form text?. So the problem isn't long explanations as such. It's long explanations that carry no signal about uncertainty.

Why do models default to confident, fluent text in the first place? Part of the answer is how they're trained. Preference optimization (RLHF) tends to reward answers that sound confident over clarifying questions or checks on understanding. One analysis found grounding behaviors falling 77.5% below human levels, which leaves models that look helpful but fail silently Does preference optimization harm conversational understanding?. RLHF is also known to damage calibration. One proposed fix uses the model's own confidence in its answer as a training reward, which restores calibration and improves reasoning at the same time Can model confidence work as a reward signal for reasoning?. Another approach stops trusting what the model says about itself in the moment. It looks up how the model actually performed in past cases where it had similar confidence, and that history turns out to be a better guide Can past performance predict when a model will be right?.

Here's the twist: models fall for the same illusion as readers. More reasoning text doesn't reliably mean better answers for the model either. In one study, raising a model's thinking budget from about 1,100 to about 16,000 tokens dropped benchmark accuracy from 87% to 70% Does more thinking time always improve reasoning accuracy?. Accuracy peaks at a medium chain-of-thought length, and more capable models prefer shorter chains Why does chain of thought accuracy eventually decline with length?. Given a question that can't be answered because key information is missing, reasoning models tend to write long, redundant responses, while simpler models just say the question can't be answered Why do reasoning models overthink ill-posed questions?. Length looks like effort and care, both to the human reading and in how models are trained. In both places, it can be a costume rather than evidence.