Citations in AI answers: how to check the source
A practical workflow for opening references, matching excerpts to claims, and understanding what a citation card does and does not establish.
The key idea
A citation does not establish that an answer is correct. Open the source, check whether the text supports the claim, then review its date, jurisdiction, and context.
Overview
The most dangerous failure mode of general language models in law is not tone — it is fabricated sources. A confident sentence, a made-up article, a judgment that never existed. In a professional setting, a single error of that kind is enough to collapse trust in the tool.
A useful review rule is to request an accessible reference for each material claim. A book title or article number inside generated text is not enough: you need to open the source and compare the passage with the context in which you intend to use it.
Retrieval first
Retrieving passages before drafting is one technical approach to providing context. Retrieval alone does not establish that a source is appropriate or that every generated sentence is supported. Those are separate evaluation questions.
Look for the source title, issuing body, passage location, date, and original URL. If a field is missing, record the gap instead of guessing. A citation card is a starting point for verification, not a seal of approval.
Verification after generation
Check each material claim: does the reference exist? Does the passage support the statement, or discuss an exception or a different situation? Could amendments or later decisions change the conclusion? Resolve these questions before reusing the conclusion in professional work.
Automated checks may help flag missing references or broken links. Interpreting the text and deciding whether it applies still requires professional judgment. Clear limitations are more useful than confidence without sufficient support.
Record the limits of the result
Keep the source, version, and access date you reviewed, and distinguish original text from your interpretation. Local website demonstrations are illustrative: a sample reference is not an authority to cite or proof of a live library’s coverage.
References & further reading
Background references, not endorsements of Mustchari or guarantees of its results.
- NIST — Generative Artificial Intelligence Profile
Discusses risks from confidently presented inaccurate output and the need for evaluation and verification.
- Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
Technical background on retrieval-augmented generation, not a guarantee of correct legal citations.