An AI-generated answer may read smoothly even when some details need checking. An uncertainty indicator uses visual or linguistic cues to show users where reliability is unclear, rather than leaving them to treat the whole output as equally dependable. section 10.2.2
Mark what needs verification
In generated text, highlight the tokens or factual claims a user should verify. The cue directs attention to specific parts of the output without requiring the user to interpret a score. section 10.2.2
Choose the cue with the user’s next action in mind. Confidence scores are another way to communicate uncertainty, but fine-grained numbers add little when users can only accept or reject an output. Footprints and source citations offer separate ways to investigate how an answer was produced. section 10.2.2
