How do you verify AI-generated citations?

What should you check separately?

Use four review fields: source identity, claim support, publication status and disposition. The workflow below is our recommended review process, not a tested accuracy guarantee for a citation checker.

Crossref’s Simple Text Query documentation describes finding DOIs from references and resolving ambiguous matches. Its Crossmark documentation describes checking reported corrections, retractions and updates. Use these for their documented purposes; assess the claim against the source text separately.

How do you establish which source was cited?

1. Save the exact AI-generated claim and reference before editing them.
2. Search for the title in the publisher’s records or a relevant scholarly index. For scholarly references, Crossref Simple Text Query is one documented DOI lookup option.
3. Compare the title, authors, year, publication and DOI with the candidate record. Investigate mismatches rather than accepting a similar title.
4. Open the original publication and record its URL and version.

Crossref notes that multiple deposited DOIs or insufficient reference and registered metadata can prevent a unique match. It asks users to evaluate alternative candidates. Consequently, our review label for an unsuccessful or ambiguous lookup is unresolved, pending further investigation. Source: Crossref matching instructions.

How do you check whether the source supports the claim?

Read the passage, table or result that would substantiate the exact statement. Record its location and compare the scope: population, task, conditions, dates and qualifications. For a number, check the denominator and metric definition. For a quotation, check the wording and attribution.

Illustrative review example: an AI answer says a study proves a tool is reliable for every business task. The accessible abstract reports an evaluation of three legal research tools. That abstract does not establish the broader claim. Mark the claim unsupported by the material reviewed, then revise it to the verified scope or withhold it.

The example reflects a real scope constraint: the 2024 legal-tool reliability paper’s abstract reports findings about Lexis+ AI, Westlaw AI-Assisted Research and Ask Practical Law AI. It is not a current assessment of every legal tool or a universal result about all business systems.

If only an abstract or indexed excerpt is readable, name that limit. Do not mark claims about unseen methods or results as verified. Seek an accessible authoritative copy or route the question to a reviewer who can inspect it.

How do you check corrections and retractions?

Check the publisher’s page for notices and, where present, use Crossmark to inspect reported updates. Crossref states that Crossmark itself is not a guarantee and relies on participating members reporting changes. Older-content coverage is optional for implementing members, so do not treat an absent button as a clean bill of health. Source: Crossmark.

Record the date you checked and the relevant notice. If a correction affects the cited result, reassess the claim using the corrected material.

What evidence should a reviewer retain?

For each consequential claim, we recommend saving:

• Claim: the exact sentence and the decision it would inform.
• Source: verified reference details, URL and version.
• Support: passage location and a brief explanation of what it establishes.
• Status: updates checked, check date and any relevant notice.
• Disposition: verified within scope, unsupported by the reviewed material, or unresolved; reviewer and next action.

These are proposed workflow labels. They should describe what was actually checked, including access gaps. An unresolved claim should remain visible to the person responsible for approving its use.

How should this fit into a business workflow?

For a hypothetical reimbursement assistant, verify both the cited policy and whether its effective date and conditions apply to the expense. Then evaluate the calculation and authorization under explicit rules. Record the transaction result separately from the explanation.

Use the prevention guide to connect evidence review to exception handling, and the measurement guide to count unsupported claims and abstentions separately. Start with the scope of a hallucination-free claim when setting acceptance criteria.

What source material was available?

Crossref’s matching instructions and Crossmark guidance were read through indexed passages because direct opens returned no body; full live pages were not verified. The legal-tool paper was read at abstract and metadata level. The recommendations above do not depend on inaccessible full-text findings or an evaluation of commercial citation checkers.