AI-Hallucinated Citations Are Now a Screening Problem: What Editors Should Do
AI-hallucinated citations have moved from a curiosity discussed at conferences to something sitting in your submission queue this week. A language model asked to support a claim will produce a citation that has the shape of a real one — correct journal, plausible authors, a volume and page range in the right register — without any record behind it. Authors who draft with these tools and do not check the output send those citations to you unchanged.
The previous article in this series covered how to check whether a submission’s references are fabricated. This one is about the policy layer: what to put in your author guidelines, what to ask at submission, how to handle a manuscript once you find hallucinated citations in it, and what not to do.
Why AI-hallucinated citations are a different problem from fraud
Editors reach for the misconduct framework because the output looks like fabrication. In most cases it is not, and treating it that way produces bad outcomes for everyone.
A fabricated citation in the classical sense is invented to deceive: to manufacture support for a claim the author knows is unsupported. An AI-hallucinated citation is generated by a tool the author trusted too much and verified too little. The intent is different, the appropriate response is different, and an editor who opens with an accusation in the second case will usually be wrong and always be resented.
What makes it hard is that you cannot tell them apart by looking. A non-existent reference is a non-existent reference. The distinction lives in the pattern — how many, where they sit, and how the author responds when asked — not in any individual line.
The practical consequence: your process needs a step between “found it” and “misconduct case”, and most journals do not have one.
What your author guidelines should say
Most instructions to authors were written before this problem existed and say nothing useful about it. Three additions are worth making, and they should be short.
1. A disclosure requirement, scoped to what matters
Ask authors to state whether generative AI tools were used, and for what. Keep the scope narrow and the wording plain: drafting, editing, translation, code, figure generation, literature searching. Do not attempt to ban the tools — you cannot enforce it, and journals that try end up with a policy everyone ignores, which is worse than no policy.
The ICMJE Recommendations and COPE both take the position that AI tools cannot be authors and that humans remain accountable for the content. Say that in your own words rather than linking and hoping.
2. An explicit statement that authors are responsible for verifying every reference
This sounds obvious and is worth writing down anyway, because it is the sentence you will point to later. Something close to: authors confirm that every reference cited has been checked against the original source and that the citation accurately reflects its content. That single line converts a later conversation from an accusation into a reminder of an undertaking already given.
3. What happens when references cannot be verified
State the consequence before it is needed: references that cannot be verified will be queried, and manuscripts with multiple unverifiable references may be returned without review. Authors respond very differently to a stated policy than to what feels like an improvised judgement.
What to ask at submission
Two fields, both cheap to add to your submission form.
The first is the AI-use disclosure itself — a free-text box, not a yes/no checkbox. A checkbox produces no information; a box produces “used for language editing” or “used to draft the introduction”, both of which tell you where to look.
The second is a confirmation that references have been verified against original sources. It costs the author one click and gives you a documented undertaking.
Neither field prevents anything. Both change the conversation later, and both signal to a submitting author that this journal checks — which has a deterrent effect that is real even though you cannot measure it.
How to handle a manuscript with hallucinated citations
Work in three tiers, and let the pattern decide the tier.
Tier one: one or two unverifiable references, scattered
Query them. Send the author the specific references, say what was searched, and ask for the correct source. This is a correction, not a case. Expect most authors to respond within a day with a fixed citation or an admission that they cannot find it either, at which point the claim comes out.
Tier two: several unverifiable references, or one in a load-bearing position
Return the manuscript without review, with the list attached and a request for a fully verified reference list before resubmission. The distinction that matters is not the count but the position: a non-existent citation supporting the paper’s central claim is more serious than three in a background paragraph, because the argument itself may be built on nothing.
Say what you found and what you need. Do not characterise the author’s conduct.
Tier three: a pattern that suggests the text was not written from the literature at all
Many unverifiable references, citations that do not support their claims, a reference list disconnected from the paper’s actual subject, or a manuscript that reads fluently while saying nothing specific. At this point you are looking at a possible paper-mill product or a wholly generated manuscript, and the right route is a documented case under COPE guidance rather than a correspondence with the author.
Even here: your report says what was checked and what was found. It does not conclude that the author committed misconduct. That determination is not an editor’s to make alone, and putting it in writing early is how journals end up in a dispute they cannot win.
What not to do
Do not rely on AI-detection tools. Text classifiers that claim to identify AI-generated writing are unreliable, and they are systematically worse on text written by non-native English speakers — which for most journals means the tool will accuse exactly the authors least able to push back. A detector score is not evidence. Checking whether the cited papers exist is evidence, and it is checkable by anyone.
Do not ban AI use. An unenforceable rule teaches authors that your policies are decorative.
Do not accuse on the strength of a flag. Ask a question. The answer tells you which tier you are in.
Do not let this reach your reviewers. Reviewers who spend an evening on a manuscript with an invented evidence base do not come back. Reviewer supply is the scarcest thing a small journal has, and spending it on unverified submissions is the most expensive mistake in this whole area.
Where the check belongs in your workflow
Before review, after the completeness screen. In practice that means: a paper arrives, it is checked for scope and completeness, and everything that survives has its reference list resolved against Crossref, PubMed and OpenAlex before a reviewer is invited.
The cost of running it late is not just wasted reviewer time. A hallucinated citation found after publication is a correction at best and a retraction at worst, and either one is a permanent mark on a journal that was trying to build an indexing case.
Having it checked for you
NeucitePress resolves every reference in a submitted manuscript against Crossref, PubMed and OpenAlex, flags entries that do not exist, do not match the record, or have been retracted, and returns a numbered correction list you can send to the author as it is. One working day, $18 per paper, no contract and no minimum.
It is one of six checks in our submission screening service. The others cover scored desk decisions with the rejection letter drafted, author and suggested-reviewer identity verification, figure integrity, and a sampled audit of whether the cited papers actually support the claims — the category no automated check reaches. If you want the policy layer built as well as the checking, that sits in our ethics and governance programme, and the whole desk can be run through the managed editorial office.
Frequently asked questions
Should we ban AI tools in our author guidelines?
No. You cannot enforce it, and the attempt teaches authors that your policies do not mean anything. Require disclosure, require that authors verify their references, and state what happens when references cannot be verified.
Can we detect AI-written text?
Not reliably. AI-detection classifiers produce false positives, and they produce them disproportionately on text by non-native English speakers. Check the citations instead — that is a factual test anyone can reproduce.
How many unverifiable references justify returning a manuscript?
Position matters more than count. One non-existent citation supporting the central claim is more serious than several in a background paragraph. Set a threshold in your policy, state it publicly, and apply it consistently.
Is an AI-hallucinated citation research misconduct?
Usually not — it is normally a failure to verify, which is negligence rather than fraud. A pattern of them, especially where they carry the argument, is a different matter and should be handled as a documented case under COPE guidance rather than as correspondence.
What should the disclosure field on our submission form ask?
Free text, not a checkbox: which tools were used and for what. “Used for language editing” and “used to draft the introduction” tell you where to look. A tick box tells you nothing.