Models 3 min readVerify a claim
Does a source link make an AI answer trustworthy?
A citation can be relevant while supporting only half of the answer. Check each claim against the passage.
The short answer
A source link gives you somewhere to check. Split the answer into claims and ask what the passage actually supports. An unanswered claim should remain visibly unresolved.
On this page
One link. Two claims.
Imagine a fictional app manual that says you can edit a document while offline. An AI answers: “You can edit offline, and conflicting changes will merge automatically.” It links to the manual.
The passage supports offline editing. It says nothing about conflict resolution. The answer has borrowed the credibility of one supported claim for a second, unsupported claim. Opening a working link would not finish the verification.
Original conceptual example / Fictional manual
One link cannot support a promise it never makes.
The manual says
- Offline editing is available
- Conflict resolution is not described
The answer should say
- Offline editing: supported
- Automatic conflict merging: not established
A working link is not a completed check.
About this visual
An independently authored fictional example. Missing support is not proof a claim is false. No citation-accuracy results are measured.
The decision this changes
Our interpretation: split an answer into claims before deciding whether its citation is enough. For each claim, identify the supporting passage and any relevant condition. Label the result supported, contradicted or not established.
Missing evidence is not the same as evidence that a claim is false. In the fictional manual, we should ask for the synchronization section or acknowledge that conflict behavior remains unknown.
What retrieval contributes
Retrieval-augmented generation, or RAG, gives a model retrieved material to use while answering. The original 2020 paper combined a searchable document collection with a generation model and reported benefits on its evaluated tasks. That historical result does not establish that every modern answer with a citation is correct. [1]
For builders Methods, implementation details and source evidence
Keep the claim attached to its evidence
A proposed documentation assistant could preserve the exact source passage, document version and section beside each substantive claim. Keep quotations distinct from the model’s explanation. Ask whether the passage supports the full claim, including words such as always, automatically or every.
This check differs from finding the right section. A system can retrieve the correct document and still add a promise that document never makes. Evaluate retrieval and claim support separately.
What we don’t know
We have not measured citation accuracy or tested a commercial assistant for this edition. The manual and answer are invented examples. A supported claim can still depend on an obsolete or inaccurate source; evidence matching is one check, not a complete truth guarantee.
Try this Open the full exercise and its limits
Try this
Write a short fictional manual with three explicit capabilities and one unresolved behavior. Ask an assistant a question that combines a documented capability with that unknown behavior.
Mark each answer clause supported, contradicted or not established. Record the exact passage the model used. The expected useful behavior is a supported partial answer plus a visible uncertainty. This proposed test does not establish a failure rate.
What to watch next
Look for evaluations that expose the claim, passage and missing evidence together. When a citation supports only half an answer, can the product tell you which half?
One AI decision, made clearer.
Get a short email with one visual explanation and one practical check. Follow the full article when you want the evidence.
Send me the briefingOpens our email signup page. Confirm by email to receive the briefing. Free; unsubscribe anytime.
Preview a short email →Follow the evidence
Sources & further reading
- Lewis et al.: Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks ↗
May 2020 research; historical context, not a current product benchmark. Checked September 27, 2026.
Back to cited passage ↑
A correction or a different perspective? Write to us.