Ask who is responsible for making sure a citation in a published paper is real, and you get four plausible answers: the author, the reviewer, the editor, or the publisher.
In practice the answer has been none of them — not through negligence, but because the job was never assigned. It sat in a gap that did not matter much until fabricating a reference stopped requiring effort.
What each party is actually doing
The author assembles the reference list, often across months, frequently pulling entries from other papers’ bibliographies. Historically this was safe: if a citation appeared in a published paper, someone had presumably read it.
The reviewer assesses whether the method is sound and the conclusions follow. They read the reference list for engagement with the literature — whether the right prior work is acknowledged — not to audit whether each DOI resolves. Nothing in any reviewer guideline asks them to.
The editor manages the review process and makes the accept-reject decision. Reference verification is not part of that remit at any journal we are aware of.
The publisher handles production, typesetting, and formatting. Reference formatting is checked routinely. Reference existence generally is not.
Four parties, each doing their job properly, and the check falls between all of them.
Why this held up for decades
Because fabricating a reference used to be hard and pointless.
Inventing a plausible citation meant constructing a fake author, journal, volume, and year, then hoping nobody looked. It was deliberate fraud with poor odds and severe consequences. Rare enough that the system did not need a defence against it.
What changed is not that researchers became dishonest. It is that fabrication became accidental. A language model produces a citation that looks correct because producing text that looks correct is what it does. An author who trusts the output and does not check has not committed fraud. They have skipped a step that nobody previously needed to take.
An unassigned responsibility is fine until something starts exploiting the gap. Then it is a hole.
The case for putting it on publishers
Of the four parties, the publisher is the only one for whom this check is nearly free.
Reference lists already pass through automated processing at submission — formatting, style conversion, metadata extraction. Adding DOI resolution to that pipeline is a mechanical step against a public API. Crossref exposes the data. Retraction Watch data is integrated into it.
A submission system could flag every unresolvable DOI before a manuscript reaches an editor, let alone a reviewer. The author fixes them, resubmits, and the reviewer never wastes attention on it.
Compare the alternatives. Asking reviewers to verify forty references adds one to two hours of unpaid work to a task people already decline for lack of time. Asking authors to self-verify is correct and necessary — our checklist covers it — but relies entirely on diligence, which is exactly what failed to produce the current situation.
Automation at submission is the only option where the cost is near zero and the coverage is complete.
What automation would not fix
Worth being precise, because overselling this would be its own kind of error.
DOI resolution catches fabricated references and corrupted metadata. It does not catch a real, correctly cited paper that does not support the claim attached to it. That failure passes every automated check ever devised, and catching it requires reading both the sentence and the source.
It would also not catch retracted papers unless retraction status is explicitly checked — a gap that AI tools currently handle badly, and one that automated pipelines would need to be built for deliberately.
So automation would not solve the problem. It would remove the largest and most mechanical slice of it, leaving the part that genuinely requires judgement to the humans who can exercise it.
Where responsibility sits today
With you. Not because that is the right allocation, but because it is the only one currently operating.
The practical consequences are worth stating plainly.
Do not inherit citations from other papers unchecked. The convention that a published reference list has been vetted no longer holds reliably enough to lean on.
Verify before submission, not after acceptance. A correction after publication is far more costly than ten minutes before.
If you review, you are entitled to flag it. Nobody requires you to check references — but noting that several do not resolve is legitimate feedback, and it costs you one spot-check of two or three entries.
The norm will shift eventually. Automated screening at submission is too cheap and too obvious not to become standard. Until it does, the gap is real and the default assignment is the person with the most to lose from it.
This piece is analysis rather than reporting. It draws on findings covered in our earlier articles on citation survival through peer review and journal AI policies, where the underlying sources are listed. Research-based rather than hands-on — see our Methodology page.