Fake citations in AI research papers: what happened, and how to check yours

In 2026, the AI research world had an awkward problem. Papers at top AI conferences were found citing papers that do not exist. The fake references look like the work of AI writing tools, and many of them passed peer review.

This matters for every student who uses AI to help with writing. Here is what happened, and a simple way to check your own references.

What is a fake (hallucinated) citation?

A citation points to another piece of work: its authors, title, year and where it was published. A hallucinated citation looks real but is wrong. The paper may not exist at all, or the authors, title or year may be wrong.

Chatbots make these because they write text that looks likely, word by word. References follow a common pattern, so a chatbot can write one that looks perfect but points to nothing. University library guides warn that AI tools can invent titles and authors.

NeurIPS 2025: fake citations in accepted papers

NeurIPS is one of the biggest AI conferences in the world. The company GPTZero checked 4,841 of the 5,290 accepted papers from NeurIPS 2025, and found 100 fake citations.

A study by Samar Ansari (February 2026) looked closely at these 100 citations. They appeared in 53 accepted papers, about 1% of all accepted papers. Each paper had been reviewed by 3 to 5 expert researchers, and the fake citations still got through.

The study sorted the fake citations into five types. The two biggest:

• Fully made up (66%): the cited paper does not exist at all. The study calls this "total fabrication".
• Wrong details (27%): a reference with some wrong details, such as the authors or the year.

The other three types made up the last 7%. The study also found that every fake citation had more than one problem at the same time, which is part of why they were hard to spot.

ICLR 2026: fake citations in submissions

GPTZero also checked a sample of 300 papers submitted to ICLR 2026, another top AI conference. It found 50 fake citations in that sample. GPTZero says it is now working with ICLR on the problem.

ICML 2026: reviewers who broke the AI rules

It is not only authors. Reviewers use AI too. ICML 2026 let reviewers choose between two rules:

• Policy A: no AI at all.
• Policy B: AI may help you understand the paper and polish your review.

To check, ICML hid instructions for AI inside the paper PDFs. If a reviewer pasted the PDF into a chatbot, the hidden text would leave a trace in the review. A human checked every flagged case, and normal "AI text detectors" were not used.

The result: 795 reviews (about 1% of all reviews), written by 506 reviewers who had agreed to Policy A, had used AI. Some of these reviewers were also authors who had to review as part of submitting their own papers. ICML desk-rejected those papers, which means it rejected them straight away, even if they already had reviews: 497 papers, about 2% of all submissions.

ICML noted that this check is easy to get around. It probably caught only the most careless cases.

How to check every citation you write

1. 1Never copy a reference from a chatbot. Use AI to find ideas and search words, then find the real paper yourself.
2. 2Search the exact title on Google Scholar, Semantic Scholar or arXiv. If you cannot find it, do not cite it.
3. 3Check the details. The authors, the year and where it was published must match the real paper.
4. 4Open the DOI link (it starts with doi.org). It must open the same paper, not a different one.
5. 5Read at least the abstract. Make sure the paper really says what you say it says.
6. 6Save papers in a free reference manager like Zotero, straight from the paper's web page. Then the details come from the real source, not from your typing or a chatbot.
7. 7Before you submit, check every reference one more time.

Researchers are also building tools to check references automatically, such as CiteAudit (2026). Tools help, but you are still responsible for every citation with your name on it.

So, can I use AI at all?

Usually yes, if you follow the rules. Most conferences and journals now have an AI policy. Read it before you submit. Use AI to help you search, understand and edit, and say that you used it if the rules ask you to. But every citation and every claim in your paper is your responsibility.

At APJ-RAISE, our free courses on reading and reviewing papers and writing a research paper teach you how to find, read and cite papers properly.