Checking AI Research Citations: A Practical Reference

Checking AI research citations means verifying the claim, not just opening the link. A report can contain real sources and still misstate what they found. Before approving an agent-written brief, ask whether each important assertion is supported by the cited passage, with the same population, period and units. A working link is only the first check.

The reference below separates fifteen ways a citation can fail into retrieval, support and provenance. It is a proposed editorial checklist, not a study of model accuracy. Use it when a report informs a purchase, a public article or a business decision; the aim is to make unresolved evidence visible before fluent prose turns it into apparent fact.

Read the table from left to right: identify the failure, perform the check, then decide what the report can honestly say. A case can trigger several rows. The groups organize work; they are not severity scores.

Gao and colleagues’ ALCE paper evaluates fluency, correctness and citation quality separately. That separation is useful: a readable answer and a well-cited answer are different achievements. We use the distinction here without carrying the paper’s historical model results forward as current failure rates.

W3C’s PROV overview describes provenance as information about the entities, activities and people involved in producing something. For a research brief, that means keeping track of the source document and how a claim was extracted, summarized or calculated. A citation URL alone does not record those transformations.

The fifteen rows are our synthesis for editorial work. Neither source publishes this checklist, validates its coverage, or assigns these groups a risk weighting. That boundary matters when the checklist itself becomes somebody else’s citation.

Start with the decision-bearing sentences: the price that changes a budget, the date that creates urgency, or the comparison that supports a recommendation. Give each a short identifier. Store the exact assertion, source location, document date, retrieval date and a status such as supported, partial, contradicted or unverified. Keep the register small enough for an editor to use during revision.

Consider an illustrative sentence: a tool is cheaper and available to every team. The price document might support the first clause while an access policy excludes some accounts. Split the sentence. Cite the price and access conditions independently. Deleting the second clause is preferable to stretching the first source beyond what it says.

Record calculations as calculations. If an agent divides a monthly price by a hypothetical task count, the result is a scenario. The vendor did not publish your cost per task. Save the assumed workload and arithmetic so a reviewer can change either input without reconstructing the entire brief.

A script can flag broken URLs, missing dates, duplicate source locations and numbers that differ between a table and summary. A second model can propose the supporting passage. Neither step settles whether a study population matches the audience being described. Put those judgments in the review queue with the evidence already attached.

Avoid a reviewer prompt that asks only whether the draft looks accurate. Ask for unsupported clauses, changed units and omitted limitations. That produces corrections a writer can apply. Preserve disagreements rather than instructing reviewers to converge on a single answer.

This complements our review of research agents and the vendor benchmark reproducibility audit. The first concerns the research system; the second concerns whether a published result can be recreated. This checklist concerns the sentence you intend to publish.

Readers do not need the entire collection log. They do need to know when a central source was unavailable, an estimate substitutes for a measurement, or a finding came from the vendor selling the product. Place that qualification beside the claim. A general disclaimer at the bottom cannot repair an overconfident table cell.

Keep source snapshots where your organization has permission to retain them. When a source changes, the old passage explains why an earlier conclusion differed. A refresh should revisit dependent claims, not merely replace the retrieval date.

For benchmark-specific reporting, our benchmark methodology guide covers contamination and comparability. Do not turn a citation check into a substitute for examining the experiment.