How to Spot AI-Fabricated Sources in Student Work

You open the bibliography of a seminar paper and something feels off. The title fits the topic perfectly, the author sounds credible, the year lines up — yet you cannot locate the item anywhere. This is not a coincidence. The language models students use while writing can generate citations that look entirely real but point to no existing publication. For academic staff this is a new and tiresome challenge: how do you tell carefully gathered literature apart from entries a model simply "filled in"?

This article is not about blaming students. Most hallucinated citations end up in papers unintentionally — a student asks an AI assistant to "complete the bibliography" and trusts that if an entry has an author, a title and a year, it must exist. Your job as a supervisor is not to hunt for guilt but to safeguard scholarly rigor. And to do that, you need to know what to look for.

Why AI invents sources

A language model has no bibliographic database it "copies" references from. It predicts the next words based on probability. When you ask it for a citation, it assembles the most plausible-sounding set: a typical name from the field, a title tailored to the topic, a sensible year, a correctly formatted DOI. All of it is statistically believable — which is exactly why it is so hard to catch by eye. The entry does not look suspicious; it looks too good.

A principle worth adopting: no single entry can be judged "fabricated" from how it sounds alone. It can only be confirmed or not confirmed against source databases. That distinction matters in conversation with the student too — we say "this entry could not be verified," not "you faked a reference."

Warning signs

Certain features should raise a flag and prompt you to check the entry in a database:

• A perfect-sounding but untraceable title — it reads as if written precisely for the paper's topic, yet you cannot find it in Google Scholar, the library catalog, or with the publisher.
• A DOI that leads nowhere — the doi.org/... link returns a "DOI not found" error or resolves to an entirely different article.
• A mismatched page number — a quotation supposedly from page 214 in a work that has 150 pages, or a reference to a chapter the volume does not contain.
• An author with no track record on the subject — the name exists, but there is no publication close to the cited title.
• A phantom journal — the periodical's name sounds credible but appears in no index or on any publisher's list.
• Suspicious consistency — the whole bibliography is "too even": every entry in identical, textbook format, without the small inconsistencies typical of human research.

One sign is not proof. But several at once in the same paper is a strong reason to move to systematic verification.

How to verify it

As an illustration, take a made-up, fabricated entry (deliberately untrue): Kowalska A. (2021), "The Impact of Digital Media on Adolescent Identity," Quarterly of Digital Pedagogy, 14(3), pp. 45–67. It sounds credible — and that is exactly why it must be checked, not judged at a glance.

The manual procedure looks like this:

1. Paste the title in quotation marks into Google Scholar and the publisher's search. No hits on the exact title is the first signal.
2. Check the DOI if given — paste it into doi.org. It should open precisely this publication.
3. Verify the journal or publisher — does the periodical exist at all, and does it have the stated volume.
4. Confirm the page number — does the cited passage fall within the work's actual length.
5. Consult the databases — Crossref and OpenAlex for articles, Google Books for monographs.

The problem is not that this is hard. The problem is scale.

Why doing it by hand fails

One bachelor's thesis often has 30–50 bibliographic entries. Verifying each properly — title, DOI, journal, page — takes a few minutes per item. For a single paper that is an hour, maybe two. But a supervisor rarely runs a single paper. With 30 seminar students we are talking about hundreds, if not over a thousand, references to verify in a single session. By hand this is simply not feasible — which is why, in practice, nobody does it, and hallucinated citations pass through.

What you need is a tool that runs the same procedure automatically, across the whole bibliography at once.

How cytado verifies a bibliography

cytado takes the full reference list from a paper and checks every entry against source databases — Crossref, OpenAlex and Google Books — exactly the way you would by hand, only in seconds. For each entry it returns a clear status: confirmed (found in a database, details match) or unconfirmed (could not be located). Crucially, cytado never declares an entry "fabricated" — it only flags what could not be confirmed and leaves the judgment to you. You know the context of the paper and can follow up with the student.

For departments and faculties we are building batch checking: you upload a folder with an entire session of papers, and cytado verifies all bibliographies at once and returns a consolidated report in the panel — with a list of entries needing attention in each paper. That turns unfeasible hours of manual work into a single review. You can read more about a department rollout on the for institutions page.

Summary

Hallucinated citations are not a passing fad but a permanent side effect of writing with AI. You cannot eliminate them with appeals — you have to catch them. Knowing the warning signs lets you flag individual cases, but at the real number of supervised papers you need to verify whole bibliographies, and ultimately — in batch, for the entire department. This is not about hunting students down; it is about making sure that what stays in a paper is what actually exists.

If you want to check the bibliography of a specific paper, try the cytado source checker. And if you are considering a solution for a whole unit, see the offer for institutions.