AI-Invented References: How to Catch Citations That Don't Exist

AI writing tools often produce references to papers that were never published, and those references look as convincing as real ones. The reliable defence is to check each reference against published records, then have a person look at the ones that do not match.

Similarity checking catches copied text. It does not catch a reference list that was made up. As more students and researchers draft with AI tools, schools, research offices and journal editors are finding citations that point nowhere. This article explains the problem, what to check, and how Veritext, the originality platform we built ourselves, now helps.

The short answer

Treat every reference as a claim that can be checked. A real reference matches a published record: the authors, title, journal and year line up. An invented one usually fails that test somewhere. Checking by hand is slow, so most reference lists go unchecked. Veritext now checks the references in a paper and flags the ones it cannot match to a published record, so the instructor or editor knows exactly which to look at.

How common are invented references?

Common enough to plan for. In a 2023 study in Scientific Reports, Walters and Wilder asked ChatGPT to write short literature reviews and checked the 636 references it produced. They found that 55% of GPT-3.5 references and 18% of GPT-4 references were fabricated, and many of the real ones contained substantive errors. Newer models have improved, but the problem has not gone away. A 2026 study by Spennemann in the journal Publications examined 110 AI-generated essays and found that close to 30% of their references were entirely invented, with many more partly wrong.

The invented references are rarely obvious. They tend to combine real authors, real journals and plausible titles, which is exactly why they slip past a reader who is not checking.

Why a similarity check alone misses this

A similarity report answers one question: how much of this text matches existing sources? An invented reference is not copied from anywhere, so it produces no match at all. A paper can come back with a low similarity score and still rest on a reference list that does not exist. The two checks answer different questions, and an integrity review needs both.

What to check in a reference list

• Does the work exist? The title should be findable as a published article, chapter or book.
• Do the details match? Authors, year, journal and volume should agree with the published record.
• Does the identifier lead to the right paper? A DOI that resolves to an unrelated paper is a warning sign.
• Does the source say what the paper claims? Only a person reading the source can confirm this, so keep it for the references that matter most.

What Veritext now does

Veritext now includes three features aimed at research and publishing work:

• Citation checking. Veritext checks the references in a paper and flags those it cannot match to a published record. A flag is a prompt to look, not a finding of misconduct: a reference can fail to match because of a typing error or a source that is not widely indexed.
• A scholarly index. Manuscripts can be checked against published research, not just against web pages.
• Journal folders. A school's research office or journal can keep its own past issues in Veritext and check new submissions against them.

Checks are unmetered under a fair-use policy, with no per-student fee, so drafts can be checked as often as the teaching or editing calls for. Papers are never sold, shared or used to train AI.

Keep the judgement with people

A flagged reference, like an AI-writing score, is a signal, not proof. The fair response is a conversation: ask the author for the source. A student who used an AI tool carelessly and a researcher who mistyped a page range both deserve the chance to explain. Our article on why an AI-writing score is not proof covers the same principle for AI detection.

Frequently asked questions

Can AI tools really invent references?

Yes. Published studies have found AI tools producing references to papers that do not exist, often built from real author names and real journal titles.

Does a low similarity score mean the references are genuine?

No. Similarity checking looks for copied text. An invented reference is not copied from anywhere, so it needs a separate check.

Does a flagged reference prove misconduct?

No. It means the reference could not be matched to a published record. Ask the author for the source before drawing any conclusion.

Can we try citation checking in the free demo?

The free demo shows the similarity report only. Ask us for a walkthrough of the full platform, including citation checking.

Where to start

1. Add reference checks to your integrity policy. Say when they apply, for example theses and journal submissions.
2. Run the check on drafts, not just final papers. Problems caught early are teaching moments, not disciplinary cases.
3. Review flags by hand. Ask for the source before concluding anything.
4. Check the sources that carry the argument. Read the key ones to confirm they say what the paper claims.
5. Teach it. Show students how to verify an AI-suggested reference before they use it.

Onprem2Cloud IT Solutions Co. is a Microsoft CSP partner based in Muntinlupa City, Metro Manila. See the Veritext product page, our other in-house software on the software page, or contact us for a walkthrough.