AI-assisted research starts with a claim-evidence ledger, not fluent prose

A citation has three jobs that can fail separately

First, the referenced item must exist and be identified accurately. Second, the reviewer must access the relevant version or authoritative record. Third, the cited passage must support the nearby claim at the stated strength. A polished link can pass the first job and fail the third; a true source can be cited for a conclusion it never reaches. Independent citation research has evaluated these dimensions and documented reference hallucination in tested language-model systems. That evidence does not establish one permanent error rate for every model or task. It does justify treating AI output as a research lead. The claim-evidence ledger records existence, access, support, and scope separately instead of giving a paragraph one vague verified label.

Evidence: Proceedings of Machine Learning Research; Association for Computational Linguistics

Discovery and verification are different modes of work

Generative systems can suggest vocabulary, competing hypotheses, institutions, source families, and search paths. Discovery tolerates candidates and uncertainty because its output is a queue. Verification requires the original or authoritative source, a stable identity, date and version, relevant passage, context, and an editor capable of judging the domain. Never ask the same unsourced answer to certify itself. A model may confidently repeat the same error, and retrieval can find a page whose words are topically similar but logically insufficient. Mark every item as candidate, located, inspected, supports, contradicts, contextual only, or rejected. Only inspected evidence may enter a factual publication claim.

Evidence: National Institute of Standards and Technology; National Institute of Standards and Technology

The unit of review is a bounded claim

Break prose into statements a source could actually support: a definition, date, named rule, product behavior, measured outcome, quotation, comparison, or inference. Record exact wording, stakes, volatility, required authority, source passage, page or section, caveat, and reviewer. Compound sentences often need splitting because one citation may support a definition but not the causal conclusion attached to it. Distinguish a source's finding from the article's inference and label illustrative examples. Numerical, legal, medical, financial, safety, and product-current claims require proportionally stronger review. NIST's risk-based approach supports tailoring measurement and oversight to context; it does not turn a framework into evidence for the article's subject matter.

Evidence: National Institute of Standards and Technology; National Institute of Standards and Technology

Contradiction is evidence, not an inconvenience

Search deliberately for limitations, later versions, corrections, negative findings, jurisdictional differences, and sources with a plausible opposing interpretation. Record whether disagreement comes from definition, sample, date, method, population, policy, or genuine conflict. Do not average incompatible findings or let an AI synthesize consensus without exposing the source set. An authoritative rule may supersede an older official page; a primary study may answer a narrower question than a review. When sources remain unresolved, narrow the claim, describe the disagreement, or omit it. The ledger should preserve rejected candidates and reasons so another editor can see that absence of citation was a decision rather than a forgotten search.

Evidence: National Institute of Standards and Technology; Association for Computational Linguistics

Synthesis adds obligations that retrieval does not satisfy

After individual claims are checked, review the whole article for selection bias, missing populations, causal language, timeline, category errors, duplicated dependence on one source, and whether summaries change the author's meaning. Verify quotations against the displayed source and respect copyright and quotation limits. Check links, titles, authors, publishers, dates, versions, and access. A source may be independent in ownership yet still derive all facts from the same press release. Count evidence chains, not logos. The editor owns the final interpretation, including transitions the model invented between supported facts. Human approval is substantive only when the reviewer can reject, rewrite, and stop publication.

Evidence: National Institute of Standards and Technology; Proceedings of Machine Learning Research

Build one ledger before asking AI for another draft

The next action is to select five factual claims from a pending article and complete existence, access, support, contradiction, scope, and reviewer fields for each. Remove or qualify any claim whose support cannot be shown from the source. Limits remain: paywalls and dynamic pages can constrain access, domain expertise may be unavailable, new sources can supersede today's record, and human reviewers also make errors. AI-assisted verification tools can reduce clerical work but cannot supply authority they never accessed. Schedule rechecks for volatile claims and preserve source versions where lawful. Research is ready when a skeptical editor can reconstruct why each consequential sentence deserves its wording.