Design Citations and Sources in AI Output

AI Interfaces · ai, trust, transparency

Updated 2026-07-28

Attach sources to claims, not to the answer as a whole. A list of links at the bottom does not tell the reader which sentence came from where.
Make a citation verifiable in one action. If checking a claim takes more effort than looking it up independently, nobody will check it.
Show enough of the source to judge it before opening. A title, a publisher, and a date let the reader weigh it, and a bare number does not.
Distinguish what came from a source from what the model produced. Summary, quotation, and inference are three different things and users treat them identically unless you separate them.
Be explicit when there are no sources. An answer generated without retrieval should say so rather than borrowing the authority of a cited one.
Never fabricate the appearance of rigour. A citation that does not support the claim is worse than no citation, because it converts a guess into apparent evidence.
Show the date of the underlying material. Confident text about a fast-moving subject is misleading when the sources are years old.
Give the reader a route to the raw material, including the passage used, so an expert can check the reasoning rather than trusting the summary.

Citations are the mechanism by which a user can disagree with the system, and an interface that makes disagreement hard is asking for blind trust.