When you ask a language model how confident it is, the scale you choose shapes the answer more than the field has acknowledged.
Researchers tested six large language models across three datasets, asking each to report a numerical confidence score on the standard 0-100 range. More than 78% of responses landed on just three round numbers. The models weren't distributing uncertainty evenly; they were gravitating to landmarks, the way humans reach for "fifty-fifty" or "about seventy-five percent." When the team swapped in a compressed 0-20 scale and measured metacognitive sensitivity using a psychophysics metric called meta-d', the narrower format consistently outperformed the standard one. Squeezing scale boundaries inward degraded performance, and round-number clustering persisted even when researchers introduced irregular ranges to disrupt it.
The practical implication is blunt: any confidence number an LLM outputs partly depends on which options you handed it. Systems that use verbalized confidence to route decisions (flagging uncertain outputs for human review, for instance) are quietly inheriting a design bias that almost no one has tested for. The researchers argue that confidence scale design should be treated as a first-class experimental variable, not a default left in place because 0-100 feels natural.
This complicates the case for verbalized confidence as a cheap substitute for costlier uncertainty methods such as model ensembles or calibration layers. If outputs shift this much based on scale formatting, "just ask the model" requires considerably more scrutiny than the field has given it.