Check an AI citation in four stages: confirm the source exists, confirm its identity and authority, locate direct support for the claim, and assess date and context. A real URL is not enough. Record what the source supports and revise any wording that goes beyond it.
Who this is for: Students, editors, researchers, and professionals responsible for publishing or submitting AI-assisted factual work.
- Verify bibliographic identity before spending time interpreting a supposed source.
- Match each citation to a specific claim and inspect the original passage in context.
- Use current primary sources when facts, product behavior, or guidance can change.
Start with source existence and identity
Open the link or locate the work through the publisher, organization, journal, or library. Check title, author, issuing body, publication date, and stable identifier where applicable. Fabricated citations often combine a real author, plausible title, and nonexistent publication. A redirect to an unrelated page is also a failed citation.
If the model provides only a title, search exact phrases and then broader combinations, but do not infer that a similar result is the cited work. Record 'not located' after a reasonable search. Ask for better evidence or remove the claim rather than completing the citation from guesswork.
Assess authority for the claim
Authority depends on the question. Official API documentation is appropriate for current API behavior. An original research paper is stronger for its own methods and findings than a blog summary. A regulator's publication may establish its guidance but not prove that every organization follows it. Match source role to claim role.
Inspect whether the page is maintained, who is accountable for it, and whether it cites primary material. A provider's marketing page may accurately announce a feature but overstate comparative value. Use independent or primary evidence for broader performance claims. For safety, legal, health, or financial conclusions, seek qualified domain review.
Test direct support and context
Find the exact paragraph, table, figure, or documentation section that supports the statement. Compare scope, population, conditions, and strength. 'Can improve' does not support 'will improve.' A result from one tested setting may not support a universal claim. Read surrounding limitations and definitions before quoting.
Watch for citation laundering, where one secondary article cites another until the chain ends without evidence. Follow the chain to the primary source. If a quotation is accurate but the surrounding source attributes it to someone else or rejects it, the citation still misrepresents the record.
Record and communicate the check
Maintain a source table with claim, URL, source identity, supporting location, access date, and verdict. This record lets another reviewer repeat the check and helps update volatile pages later. Save appropriate excerpts or document locations according to copyright and organizational policy, not entire unauthorized copies.
Revise the prose to match verified support. Correct the citation, narrow the claim, add conditions, or remove it. If evidence remains uncertain, say so plainly. For current software details, point readers to official documentation and include a review date because interfaces and capabilities can change after publication.
Verify a citation in a training memo
An AI-written memo says NIST requires every company to ban generative AI and links to the AI Risk Management Framework.
- Open the official NIST page and confirm the framework's title, issuing organization, and publication context.
- Search the framework and related playbook for the claimed mandatory ban language.
- Distinguish voluntary risk-management guidance from a legal requirement and record that the cited page does not support the claim.
- Replace the sentence with an accurate description of the framework's risk-management purpose.
- Have the policy owner review whether separate laws, contracts, or internal rules create obligations for this organization.
Citation audit record
Complete these fields for every source-bearing claim that matters.
- Identity check: title, author or issuer, publication, date, identifier, and working official URL.
- Authority check: why this source is appropriate for this exact type of claim.
- Support check: quoted passage or table location, scope, conditions, and relevant limitation.
- Currency check: update date, superseding guidance, archived status, and next review date.
- Editorial verdict: keep, narrow, correct, replace, remove, or send for specialist review.
Common mistakes
- Treating an official-looking domain as proof that the page says what the draft claims.
- Citing a search result snippet without opening and reading the underlying source.
- Using one citation at the end of a paragraph containing several unrelated factual claims.
Try one
A generated report cites current model documentation for a claim about lower business costs. What evidence checks are required?
The reviewer should confirm the documentation describes the specific model and current behavior, then determine whether it actually establishes business cost rather than technical capability. Cost claims need relevant usage assumptions and current pricing or internal measurements, which should be checked separately. A strong answer narrows or removes the conclusion if the cited documentation does not directly support it.
Sources
- NIST AI Risk Management FrameworkNIST framework for governing, mapping, measuring, and managing AI risk.
- NIST AI RMF PlaybookNIST suggested actions for applying the AI Risk Management Framework.