ChatGPT Fabricates References: Why It Happens, How Common It Is, and How to Check a Source Really Exists

Why ChatGPT and other chatbots invent papers and references that look real, what research found about the rate of fabricated references, the four kinds of errors, and a five-minute checklist to verify every source before it goes into your paper.

Part of the series: AI in Academic Writing: The Complete Guide to Using ChatGPT Correctly in a Seminar Paper or Thesis

A student asks ChatGPT for "five papers on test anxiety" and gets a perfect list: real researchers in the field, real journals, plausible years, even DOIs. Except some of the papers don't exist. It's the mistake that turns a good paper into a disciplinary case fastest - and one of the easiest to prevent.

Why does it happen?

A language model doesn't "look up" papers - it predicts the next plausible text. It has seen millions of references, so it knows what a reference looks like: a known researcher's name, a title in the right style, a fitting journal. The result is a reference that looks entirely credible - but was assembled from parts, not retrieved from a database.

How common is it?

In a study published in Scientific Reports, researchers had ChatGPT write short literature reviews on 42 topics and checked 636 references. 55% of GPT-3.5's references and 18% of GPT-4's were entirely fabricated. Even among the real ones, 43% (GPT-3.5) and 24% (GPT-4) had substantive errors - wrong year, journal, pages or authors (Walters & Wilder, 2023).

Newer models are better, and tools with web search err less - but no study has found a rate of zero. One source is enough to break a paper.

Four kinds of error - from obvious to dangerous

1. A paper that doesn't exist at all - easiest to catch: one search.
2. A real paper, wrong details - wrong year, journal or DOI. Looks like careless APA.
3. A real author, a paper of theirs that doesn't exist - harder: the researcher really works on the topic.
4. A real paper - but the claim isn't in it. The most dangerous, and it happens with database-backed tools too: the paper exists but doesn't say what AI attributed to it. Only reading catches it.

Verification checklist: five minutes per source

1. Search the exact title in quotation marks in Google Scholar or your library catalogue. No result? The source probably doesn't exist.
2. Check the DOI: paste it into doi.org. Does it lead to the same paper, with the same authors and title?
3. Compare the details: year, journal, volume, pages - against the paper's page on the journal's site.
4. Read at least the abstract - and if the claim is central to your paper, the relevant section. Does the paper say what you write in its name?
5. Only list what passed every check, and take the details from the paper itself - not from the AI's text.

Using AI without falling in

• Ask for search terms, not sources: "which English terms should I search for on this topic?" - then search yourself.
• Use database-backed tools (Semantic Scholar, Elicit, Consensus) - their papers exist. See AI tools for academic research.
• Work on papers you chose: upload them to a tool like NotebookLM and ask only about them.
• Don't copy a reference from a chatbot - export it from Google Scholar or Zotero, and check against APA templates.

Already submitted a draft with sources from ChatGPT?

Go through the whole list with the checklist now. A source that can't be found comes out, replaced by a real source saying something similar (or the claim is deleted). If the draft is already with your supervisor, it's better to tell them first that you found errors and fixed them than to have them find out for you.

FAQ

Does ChatGPT make up sources? Yes, still - less than before, but not zero. Every reference that comes from a chatbot must be verified.

How do I check whether a paper exists? Search the exact title in Google Scholar, and check the DOI at doi.org. If neither leads to the same paper, it probably doesn't exist.

Do Elicit and Consensus get things wrong too? Their papers are real, but the summary or extraction can attribute something to a paper that it doesn't say. Read the source before citing it.

What happens if I submit a fabricated source? At most institutions it counts as an academic-integrity breach, even without intent - Bar-Ilan, for example, explicitly holds students responsible for errors that come from AI.

Want someone to go through your reference list? We verify every source - that it exists, its details, and that it really says what's attributed to it - and bring it into APA. Get a quote. For the full guide to working with AI correctly: AI in academic writing.

Source

• Walters, W. H., & Wilder, E. I. (2023). Fabrication and errors in the bibliographic citations generated by ChatGPT. Scientific Reports, 13, Article 14045. https://doi.org/10.1038/s41598-023-41032-5

Last updated: 2026-10-01