Observable signal
Confidence rises with answer length, polish, or decisiveness while verification behavior falls.
Trust calibration
The confidence a human places in an AI answer can be shaped by how the answer is expressed, not only by whether the answer is reliable.
Underlying MD-006 · informs MD-01101 / DEFINITION
Persuasiveness and accuracy are different variables.
Fluent explanations can transmit confidence. A user may therefore become more certain without becoming better able to distinguish correct outputs from incorrect ones.
Trust calibration concerns the relationship between actual system reliability and the confidence users place in its outputs. Research has shown that longer explanations can increase human confidence even when the added length does not improve answer accuracy or people's ability to discriminate correct from incorrect answers.
This creates a practical design problem. Explanation is valuable when it exposes assumptions, evidence, uncertainty, and verifiable reasoning. Explanation becomes hazardous when rhetorical completeness is mistaken for epistemic support.
Not the same as
Trust earned through repeated independent verification, reliable provenance, or measured performance under relevant conditions.
Useful control
Separate presentation quality from evidence quality; make uncertainty and source support inspectable rather than implicit in tone.
02 / EVIDENCE
What the research supports.
Human perception of model confidence can diverge from model accuracy, and explanation style can widen or narrow that gap.