Why an AI Answer Can Cite a Source and Still Be Wrong

Conversational can show the passages used to write an answer. That is useful. It does not mean the answer is correct. The model can retrieve the wrong file, retrieve an earlier version of a fact, or draw a conclusion the passage does not support.

The three examples below are fictional notes. They are not a product measurement.

You ask: When is the project review? The workspace also contains a note titled "Friday lunch order." The model answers Friday and cites the lunch note. The citation is real. It is not about the review.

Check: does the cited passage mention the review at all? If not, the link is not evidence for that question.

The demo file has a 10 March note that says the review is Tuesday, and a 12 March note that moves it to Friday. If the answer says Tuesday and cites 10 March, the citation is accurate as history and wrong as current state.

That is a stale-fact error: serving a superseded value as current. Paper 1 studies this class. See Why RAG cannot tell time.

The Friday note says the review moved to Friday. The model answers "Friday at 4pm in the boardroom" and still cites that note. The time and room were not in the source. The citation covers the day. It does not cover the extra details.

Check: every load-bearing phrase in the answer should appear in the selected evidence, or the product should show that evidence is limited.

After an answer, open Why this answer. Read the passage. If the source is empty, wait until indexing is Ready and ask again. The setup guide walks through the Tuesday/Friday demo.