Users do not experience a model’s internal uncertainty. They experience an interface.
A clean paragraph with no citations feels certain. A percentage feels measured. A green badge feels approved. Even the decision to return one answer instead of three changes how much confidence the system projects.
Confidence is assembled
It comes from model behaviour, retrieval quality, prompt design, formatting, labels, defaults and what the interface hides. A 92% score may describe token probability, a classifier calibration, a heuristic or nothing defensible at all.
Use evidence states instead
- Supported: decisive claims link to current primary evidence.
- Contested: credible sources disagree or the governing fact is unresolved.
- Incomplete: required information was unavailable.
- Inferred: conclusion follows from evidence but is not directly stated.
- Out of scope: the system is not approved to answer.
These states help a decision-maker act. A decorative confidence number often does not.
Field instrument
Take five consequential outputs. Remove the confidence label and inspect the evidence. Then show the output to a user with and without the interface cues. If their willingness to act changes more than the evidence changed, the interface is manufacturing authority.
Uncertainty should not make every answer timid. It should make the boundary between evidence and presentation visible.