Fake references in AI-written papers: how to spot them and check every citation

AI writing tools can produce references that look perfect but do not exist. One invented citation is enough to sink a submission or damage a reputation. This guide explains why it happens, the warning signs, and a five-step check you can run on any reference list in minutes.

Ask a general-purpose AI assistant to write a literature review and it will usually give you a neat reference list: plausible authors, a believable journal, a year, even volume and page numbers. The problem is that some of those papers may never have been written. Researchers call these fabricated or hallucinated references, and they are one of the most common and most damaging mistakes in AI-assisted writing.

Why AI tools invent references

A language model writes by predicting likely text. It has seen millions of citations, so it knows exactly what a reference looks like: common surnames, journal titles that fit the field, a recent year. Unless the tool is connected to a real database and only allowed to cite records from it, nothing forces each reference to match a real paper. The result is a citation that is statistically plausible but factually empty.

Even carefully engineered systems are affected. In the ScientistTwo study from Google Cloud AI Research, the authors' own integrity audit found 19 fabricated or unverifiable references out of 1,840 across the papers their system generated, until they added a dedicated reference-checking agent (our summary of that study). About 1% sounds small, but a single invented source in a submitted manuscript can be treated as a serious integrity problem.

Seven warning signs of a fake reference

  1. No DOI, or a DOI that does not resolve. Most journal articles published in the last twenty years have a DOI. If https://doi.org/ followed by the DOI shows "DOI not found", treat the reference as suspect.
  2. The title returns nothing in a search. Paste the exact title, in quotation marks, into Google Scholar. Real papers almost always appear.
  3. The right authors, the wrong paper. AI often pairs real, well-known researchers in a field with a title they never wrote.
  4. A real journal, an impossible issue. The volume, issue or page range does not exist for that year, or the journal did not publish on that topic.
  5. Suspiciously perfect fit. The title restates your exact argument in your exact words. Real literature rarely lines up that neatly.
  6. The citation does not support the claim. The paper exists, but its abstract says nothing about the point it is cited for. This is a subtler error, and reviewers notice it.
  7. Inconsistent details. The in-text citation says 2021, the reference list says 2022; the author spelling changes between mentions.

A five-step check for any reference list

You do not need special software. For each reference, work through these steps and stop as soon as one confirms the paper is real and correctly described.

StepWhat to doWhat it confirms
1. Resolve the DOIOpen https://doi.org/ plus the DOIThe paper exists and the DOI is correct
2. Search the exact titleGoogle Scholar, Crossref Search or OpenAlex, title in quotation marksThe title, authors and year match a real record
3. Check the journal detailsCompare journal, volume, issue and pages with the publisher's pageThe bibliographic details are accurate
4. Read the abstractDoes it support the sentence where you cite it?The citation is used honestly
5. Check for retractionsSearch the title in the Retraction Watch databaseYou are not citing withdrawn work

For a list of 70 to 120 references this sounds like a lot, but steps 1 and 2 take seconds each, and most problems surface there. A reference manager such as Zotero can also import a DOI and fill in the details automatically: if it cannot find the DOI, that is a warning sign in itself.

Free tools that help

  • doi.org: resolves any DOI to its publisher page.
  • Crossref Search and OpenAlex: open indexes of hundreds of millions of scholarly records, searchable by title.
  • Google Scholar: fast title search, and shows who cites the paper.
  • Retraction Watch Database: lists retracted papers.
  • Zotero: free reference manager; adding a paper by DOI checks it exists.

How to use AI writing tools safely

  • Prefer tools that cite from a real database. A tool that looks up real publications and only lets the model cite those cannot invent a reference, though you should still check that each one supports its claim.
  • Never copy a reference list you have not checked. Run the five-step check before you share a draft with a supervisor, and always before submission.
  • Read the sources you cite. At minimum, the abstract. Ideally, the relevant section.
  • Follow your publisher's AI policy. Many journals now ask authors to disclose how AI tools were used.

How Soba Researcher handles references

Soba Researcher was built around this problem. Before it writes a paper, it searches OpenAlex, an open index of scholarly publications, for real papers on your topic, keeping work from established journals. The writing model can only cite those records, by an internal key; it never types author names or years itself. Each citation is then turned into a correctly formatted in-text citation, in APA 7, Harvard, Chicago or IEEE style, that links to its entry in the reference list, and each entry carries the paper's real DOI.

That removes invented references, but not the need to check. A real paper can still be cited for a point it does not quite make, so read the abstract of every source before you rely on it. You can try it free at sobaresearcher.com, and read more in our FAQ.

Key takeaways

  • General AI tools can produce references that look real but do not exist; even advanced research systems have done so.
  • Warning signs include a missing or broken DOI, a title that returns no search results, and details that do not match the journal.
  • A five-step check (DOI, title search, journal details, abstract, retractions) catches almost every problem.
  • Tools that cite only from a real database prevent invented references, but you should still confirm each source supports its claim.

Figures on fabricated references are as reported by Nam et al. (2026), ScientistTwo, arXiv:2609.19644. This article is general guidance; always follow the citation and AI-use policies of your institution and the journal you submit to.