Hallucinated Citations: Why AI References Can't Be Found

# Hallucinated Citations: Why AI References Can't Be Found

A hallucinated citation is a reference produced by an AI chatbot such as ChatGPT that looks real but matches no real paper, or points to a real paper with the wrong details. It has real researchers' names, a plausible journal and correct formatting. If you can't find a reference a chatbot gave you, the reason may be that it doesn't exist. A language model writes a reference by predicting likely text, and on its own it doesn't look the reference up in a database. The result looks like a citation, but nothing guarantees that it describes an actual publication.

How common are hallucinated citations?

They're common enough to be studied, measured and screened for.

• In a 2023 study, two researchers asked ChatGPT to write short literature reviews on 42 topics and checked all 636 citations it produced. With GPT-3.5, 55% of the citations were fabricated, meaning no such work could be found. With GPT-4, 18% were. Among the citations that did point to real works, 43% (GPT-3.5) and 24% (GPT-4) had substantive errors in details such as authors, year, volume, issue or pages (Walters & Wilder, 2023).
• The problem isn't limited to one chatbot. A 2025 benchmark gave 8 AI chatbots, including ChatGPT, Claude, DeepSeek, Gemini and Grok, the same academic reference-retrieval tasks. The title states the result: "Grok and DeepSeek outperform ChatGPT, but none are fully accurate" (Cabezas-Clavijo & Sidorenko-Bautista, 2025).
• The problem is reaching published work. A May 2026 report found that the share of newly published papers containing fabricated citations rose about twelvefold in three years. Some publishers have started screening reference lists at submission (Mondal, 2026).
• Conferences reject papers over it. At ACL 2026, reviewers "identified many non-existent references which led to those submissions being desk rejected." The organizers later "identified over 100 papers that contained citations to non-existent literature" among accepted papers (ACL 2026, n.d.).

Why do AI chatbots make up references?

A language model writes text by predicting what is likely to come next, and references are no exception. On its own, the model doesn't look a reference up in a database. It produces text that has the shape of a reference: real researchers' names, a plausible journal title and correct formatting. Together, these parts may match no real paper. Outputs like this are usually called hallucinations.

This comes from how the text is generated. It isn't a bug in one product. And the more fluent the model, the more convincing a made-up entry looks.

Why asking ChatGPT to "double-check" doesn't work

Asking the same chatbot to verify its references doesn't solve the problem. It confirms a reference the same way it wrote it, by producing plausible text, so it can confidently confirm a reference that doesn't exist. The check has to happen in a database, not in the chat window.

The five forms of hallucinated citations

PerfectRefs has checked thousands of references, and hallucinated entries keep showing up in five forms:

1. Not found at all. The title can't be found in public scholarly databases, and the listed author has no similar paper.
2. Real title, wrong details. The paper exists and a title search finds it, but the entry's details don't match the record. Authors may be missing or swapped, the volume, issue and page numbers may differ, or even the journal may be wrong. The 2023 study counted errors like these as substantive (Walters & Wilder, 2023).
3. Real author, real journal, non-existent article. The author and the journal are both real, but that volume and issue contain no such article.
4. A DOI that doesn't resolve, or resolves to a different paper. The DOI looks well formed but leads nowhere, or it leads to an unrelated article.
5. Hybrids. The first half of the title comes from one real paper and the second half from another, or a real title is paired with someone else's authors.

Form 1 is the easiest to catch. Forms 2 to 5 are harder because the title search succeeds and a quick check stops there. The error is in the other fields, and reading the list by eye rarely catches it.

How to check whether ChatGPT's references are real

People usually take one of three approaches before they submit:

• Check by hand. Search each title in a scholarly search engine or library catalog and open the record. Compare the author spelling, year, volume, issue and pages. Paste any DOI into doi.org to confirm that it resolves to the same paper. This is the most reliable method, but a list of 100+ references takes hours.
• Ask the AI to check itself. As explained above, this doesn't work.
• Don't check and hope no one else does. Publishers screening at submission and reviewers desk-rejecting papers make this a risky bet (Mondal, 2026; ACL 2026, n.d.).

Checking a reference list with PerfectRefs

PerfectRefs does the database lookups for you. You paste your reference list and choose a check:

• Existence check looks up each reference in Crossref, OpenAlex and other scholarly databases and flags the ones it can't find. The Existence check comes with a free daily allowance.
• Style check compares the authors, title, journal, year, volume/issue/pages and DOI field by field with the database record. This catches the "real title, wrong details" cases that a title search misses. It shows the record's correct values with a source link, and you can accept a fix with one click. The Style check uses premium credits (2 per reference), and new accounts get 10 premium credits.

FAQ

What is a hallucinated citation?

A hallucinated citation is a reference produced by an AI chatbot that looks real but matches no real publication, or that points to a real publication with wrong details such as authors, year, volume, issue, pages or DOI. It happens because language models generate references by predicting likely text instead of looking them up in a database.

Why can't I find the references ChatGPT gave me?

The reference may not exist. In a 2023 study, 55% of the citations GPT-3.5 produced and 18% of those from GPT-4 were fabricated (Walters & Wilder, 2023). The paper may also be real but listed with the wrong authors, journal or page numbers, which can make it hard to match.

Does ChatGPT still make up references, or only older versions?

GPT-4 fabricated fewer citations than GPT-3.5 (Walters & Wilder, 2023). A 2025 benchmark of 8 chatbots, including ChatGPT, Claude, DeepSeek, Gemini and Grok, found that none were fully accurate (Cabezas-Clavijo & Sidorenko-Bautista, 2025).

Can I ask ChatGPT to check whether its references are real?

No. The chatbot confirms a reference the same way it wrote it, by producing plausible text, and it can confidently confirm a reference that doesn't exist. Check each reference against a scholarly database or library catalog instead.

How do I check if a reference is real?

Search the title in a scholarly search engine or library catalog and open the record. Compare the authors, year, volume, issue and pages, and paste the DOI into doi.org to confirm that it leads to the same paper. A title match alone isn't enough, because many hallucinated citations use a real title with wrong details. For long lists, a tool like PerfectRefs can run these lookups against Crossref, OpenAlex and other databases.