Mark Lavercombe - clinician-educator and developer of Scholar Sidekick. I read citations every week, and the gap between “verified” and “correct” is the one that most often sends a reader to the wrong paper.
A citation that a tool calls “verified” can still send your reader to the wrong paper. The word hides two different guarantees: that the cited work exists, and that the identifier you will publish resolves to it. Most checks answer the first. The failure modes that survive - a real identifier with an invented title, and a real title with a wrong identifier - are the ones that mislead the reader who trusts your reference list.
When a tool tells you a citation is verified, it has answered one of two questions - and they are not the same:
Existence is the weaker guarantee. A paper with the claimed title can exist while the identifier you actually typed into your manuscript points somewhere else entirely. Identity is the guarantee your reader depends on, because the reader clicks the identifier, not the title.
Title matching takes the claimed title (and maybe authors and year) and searches a corpus or the web for a paper that looks like it. If it finds one, the citation is treated as real. This is fast and needs no identifier, but it answers existence, and it is blind to which identifier you will publish.
Identifier resolution takes the identifier you supplied - the DOI, PMID, PMCID, arXiv ID, ISBN, ISSN, ADS bibcode, or WHO IRIS URL - and resolves it against the appropriate registry first, then compares the resolved record to your claim field by field. It answers identity: does this exact link lead to the work you say it does.
On clean citations the two approaches agree. They diverge precisely on the citations that reach print.
Five citations, two approaches. Both approaches handle the obvious cases. They diverge on the two rows in the middle - which are exactly the cases that reach print.
| The citation | Title-matching approach | Identifier-resolution approach |
|---|---|---|
| Correct - real identifier, correct title | Found by title. Verified. | Identifier resolves; every field agrees. Verdict: matched. |
| Real identifier, invented title (the Topaz fabrication pattern) | Title is not findable, so: “does not exist / fabricated.” Right call to flag it - but it never reveals that the identifier itself is live and points at a real, unrelated paper. | Identifier resolves to a real paper whose title disagrees with the claim. Verdict: mismatch - and the true resolved paper is shown. |
| Correct title, wrong identifier (a real paper, the wrong DOI pasted in) | The real paper is found by title, so the citation is often passed as “verified, minor issue.” The supplied identifier is never resolved, so the reader is never told where that bad DOI actually leads. | The supplied identifier is resolved first - to a different real paper - while the claimed title matches another. Verdict: ambiguous, with both targets shown. This is the dangerous case: a wrong DOI sends every reader to the wrong paper. |
| Fully fabricated - made-up identifier and invented title | Neither title nor identifier is findable. Flagged. | Neither the identifier nor a title search returns a match. Verdict: not found. |
| Non-DOI identifier - a correct PMID, ISBN, or ADS bibcode | Existence can be confirmed by title, but the identifier you supplied is not itself validated unless it happens to be a DOI or arXiv ID. | The identifier is resolved against its own registry (PubMed, OpenLibrary, NASA ADS, WHO IRIS …) and confirmed to point to the claimed work. Verdict: matched, on the identifier you will actually publish. |
The first and fourth rows are easy: both approaches get them right. The fifth depends only on whether the checker validates identifiers beyond DOI and arXiv. The two rows in the middle are where the difference matters - and they are the two patterns that actually slip into published reference lists.
Fabricated titles get the attention, but a real paper cited under the wrong identifier is the quieter danger. The title is genuine, the authors are genuine, the DOI resolves perfectly - to a different paper. An existence check finds the real paper by its title and reports the citation as verified, perhaps noting a “minor” discrepancy. The reader who clicks the DOI still lands on the wrong article.
Identifier resolution treats this as a first-class warning, not a footnote. It resolves the DOI you supplied, sees that it leads to a paper whose title does not match your claim, finds the claimed title under its correct identifier, and reports both - so you can fix the link before it ships. That is the difference between the cited work exists and the link you are about to publish is correct.
Because the failure modes are distinct, a binary real/fake verdict throws away information. The Scholar Sidekick verifier reports four outcomes:
Not on its own. “Verified” can mean the cited work exists (a title search succeeded) or that the identifier you will publish resolves to it (a registry lookup succeeded). Those are different guarantees. A citation can pass the first and fail the second - for example, a real paper cited under the wrong DOI. Ask which question a tool answered before you trust the badge.
Yes, and it is more common than fabrication. Someone copies the right paper's title into a reference manager but pastes the DOI of a different paper, or an AI assistant pairs a genuine title with a plausible-looking but wrong DOI. The DOI resolves perfectly - to the wrong article. A reader who clicks it lands somewhere you never intended. Identifier resolution catches this because it starts from the DOI you supplied and reports where it actually leads.
It is the dominant AI-driven fabrication pattern documented by Topaz et al. (Lancet 2026): a real, resolvable identifier stitched to an invented title. The identifier is live, so a naive check that only confirms the DOI exists passes it. The title belongs to no real paper. Catching it requires comparing the claimed title against the title the identifier actually resolves to - not just confirming the identifier is syntactically valid.
Because existence and identity are different. Confirming a paper with this title exists tells your reader nothing about whether the link you published will take them there. The two failure modes that survive an existence check - a fabricated title attached to a real identifier, and a real title attached to the wrong identifier - are exactly the ones that mislead a reader who trusts your reference list. Verifying identity closes both.
Not uniformly - that would be an overclaim. On a fabricated title both approaches will flag the citation, and on a fully fabricated reference both will fail to find anything. They diverge on two cases: a real identifier with an invented title (both flag it, but only resolution shows the live paper the identifier points to), and a real title with a wrong identifier (resolution treats it as a first-class warning and shows the true target; an existence check tends to pass it as a minor issue). Resolution also validates non-DOI identifiers - PMID, PMCID, ISBN, ISSN, ADS bibcode, WHO IRIS - that a DOI/arXiv-only checker cannot confirm. The honest summary: same answers on the easy cases, safer answers on the cases that reach print.
On a 1,395-entry blind holdout - drawn from a recorded seed after the code was frozen, then measured once - the verifier caught every fabrication on the dominant patterns (150/150 = 100%, Wilson 95% CI lower bound ~97.6%) and made high-confidence false accusations on correctly cited papers at 0.8% (95% CI 0.4-1.4%). It also reports a measured blind spot: single-word near-miss semantic flips (caught 4/30). The fixtures, methodology, and downloadable receipts are at /citation-integrity; you can run the verifier yourself.