Confidence expresses how strongly an AI output should be trusted. It is useful when the system can explain uncertainty and show which parts require human review rather than presenting every result as equally reliable. section 10.2.1 section 10.4
A confidence-aware experience should:
- Warn users when required data is missing before displaying the result. section 10.2.1
- Show an overall score alongside granular scores for individual paragraphs, statements, or passages. section 10.2.1
- Explain low scores, such as missing data or ambiguity in regulatory text. section 10.2.1
- Let users edit, shorten, remove, or otherwise reject uncertain content before publication. section 10.2.1 section 10.4
Reviewing an AI-generated draft
A draft displays high-confidence passages in green and uncertain passages in yellow. The user opens an explanation for a low score, supplies missing information, edits the highlighted text, and reruns the check before publishing. section 10.2.1
Figure 10.10 The draft of the report section is provided together with confidence scores; lower scores indicate statements and passages that should be verified by the user.
Use the figure to inspect how overall and passage-level confidence scores direct verification. section 10.2.1
