How to let users inspect sources behind an AI answer

An AI answer should let people inspect the evidence behind a consequential claim at the moment they need it. A clickable list of documents is a start, but it does not tell a reader which passage supports a date, whether the document is current, or why two records disagree. Product teams should design the answer and its evidence as one review task.

This guide concerns the user experience of source inspection. The separate AI search guide covers entry points, retrieval boundaries, and fallback behavior for an existing product.

What should an answer cite?

Cite the smallest useful unit the product can reliably expose: a passage, table row, or record field. Place its link beside the claim it supports. If the answer combines several records, connect each material statement to the relevant record rather than attaching a generic source list to the whole paragraph.

Source identity matters too. Show a recognizable title, owner or collection where available, and the source’s effective or updated date when that affects the decision. Distinguish the date of the source from the time the assistant produced the answer. A response generated this morning can still rest on last year’s policy.

Google’s People + AI guidance on explainability recommends explaining relevant data sources so users can judge when to apply their own knowledge. The interface implication is specific: expose enough context to assess a claim, then allow deeper inspection without making every answer a wall of metadata.

How can someone inspect a claim without losing the task?

Open the source at the supporting passage where the underlying system permits it. Keep the answer visible, or make returning to it obvious. When a document is long, a title-only link sends the reader into another search. Highlighting the cited passage can help, but it must reflect the actual evidence, not a plausible sentence elsewhere in the document.

A useful citation preview can show the cited excerpt and its context. It should also make clear when the assistant paraphrases, translates, or combines source material. A reader should be able to compare the answer with the source wording before copying it into another workflow.

Test the opened source with the same access permissions as the answer. A link to a document the reader cannot open is a broken review path. The team should decide whether to omit the inaccessible claim, explain the access limit, or route the user to an approved alternative. The answer itself must not reveal protected material.

What if sources disagree?

Consider an illustrative project workspace. A signed plan lists a release date; a later issue says the release was delayed but gives no replacement date. Selecting either date as the current commitment would misstate the available evidence. The answer should name the conflict, show both records, and say what remains unresolved.

Freshness is a clue, not an automatic winner. A newer comment may describe a proposal, while an older approved document may remain authoritative. Give reviewers the relevant status and owner when the product has that data. If the records do not establish which one governs, ask the responsible person or show the ordinary search results. Do not manufacture a hierarchy the organization has never defined.

Conflicting sources also deserve a distinct visual treatment. Two citations stacked under a confident sentence can look like corroboration. Label the disagreement in the sentence itself, then let users inspect it.

What belongs in a source inspection test?

Use a small, proposed matrix built from real task categories. For each row, save the question, user role, allowed records, expected answer, and the source passage that justifies it. The examples below are illustrative test conditions, not measured product results.

Ask participants to identify what they would use, what they would verify, and where they would go next. A citation click count alone cannot show that the evidence was understood.

When should the product stop at evidence?

Some questions ask for a finding the records cannot settle. In that case, the product can present the relevant material and state that the conclusion is unresolved. It may ask for a missing scope, such as which policy version applies. The guide to uncertain AI answers sets out when to clarify, qualify, or abstain.

Before adding this pattern to an AI feature in a SaaS product, define who maintains document status and permissions. A polished citation component cannot repair an unknown source of authority. The design succeeds when a person can check the claim, recognize a conflict, and continue the real task with the evidence they are permitted to see.