Why AI tools invent citations, and how to check yours

An AI tool can produce a reference with real-looking authors, a plausible title, a journal that exists and a DOI in the right format, and no such paper. The reference reads correctly. It points at nothing, or at a different paper.

This post covers what the research found, why it happens, what it has cost people, and how to check a reference list before you submit it.

What the research found

A 2025 study in JMIR Mental Health (Linardon and colleagues) asked GPT-4o to write six literature reviews. There were three topics, each with a general review and a specialized one, and each review was to include at least 20 citations. The researchers checked all 176 citations against Google Scholar, Scopus, PubMed and other databases.

• 35 citations (19.9%) were fabricated, meaning the source did not exist.
• Of the 141 citations that did exist, almost half had at least one error, such as a wrong year, volume number or author list.
• Taken together, nearly two thirds of the citations were either invented or wrong in some detail.

The fabrication rate depended on the topic. It was 6% for major depressive disorder, a heavily researched subject. It was 28% for binge eating disorder and 29% for body dysmorphic disorder, which have smaller literatures. Narrower requests did worse in some cases: for binge eating disorder, 46% of the citations in the specialized review were fabricated, against 17% in the general overview.

The study tested one model, in one field, on a limited set of prompts. Treat the numbers as an example of the problem, not as a rate for any tool you use. The pattern is the useful part: in this study, the topics with smaller literatures had higher fabrication rates.

Why it happens

A language model writes text by predicting what is likely to come next. A reference has a recognizable shape: author names, a title, a journal, a year, a DOI. A model can produce that shape without any record behind it. Tools connected to a search engine or a database can look sources up. A model answering from memory alone has nothing to look up.

A 2025 paper by Kalai, Nachum, Vempala and Zhang, Why Language Models Hallucinate, makes a further point. Most benchmarks score a model on whether its answer is correct. A wrong guess and an honest “I don’t know” both score zero, so guessing never costs the model anything. The authors argue that this scoring rewards confident guesses and discourages saying nothing.

What it has cost

In Mata v. Avianca, lawyers filed a brief in federal court in New York that cited cases generated by ChatGPT. The cases did not exist. On June 22, 2023, Judge P. Kevin Castel fined the two lawyers and their firm $5,000 and ordered them to notify their client and each judge falsely named as the author of a fake opinion.

The judge wrote that there is nothing “inherently improper” about using artificial intelligence, and that existing rules “impose a gatekeeping role on attorneys to ensure the accuracy of their filings.” The sanction followed from the lack of checking, and from the lawyers standing by the cases after the court questioned them.

A student reference list is not a court filing. The principle carries over: the person who submits the work is responsible for every source in it.

How to check a reference

Do these in order for each reference. Stop at the first step that fails and treat the reference as unconfirmed.

1. Search for the exact title in Google Scholar, your library’s database, or PubMed for biomedical work.
2. Open the DOI. Put https://doi.org/ in front of it. A DOI is meant to lead to one specific work, so if the link fails, or opens a different paper, the reference is wrong.
3. Compare the details. Authors, year, journal, volume and pages should match the record. In the study above, almost half of the real papers were cited with at least one detail wrong.
4. Read the source. Confirm that it says what your sentence says it does. A real paper cited for a claim it does not make is still a bad citation, and no database will catch that.
5. Check for a retraction. A retracted paper should not be cited as evidence. The publisher’s page for the article shows a retraction notice, and Crossref publishes retraction data that includes Retraction Watch records.

If a reference fails step 1 or 2, do not assume it is invented. Books, theses, reports and non-English sources are often missing from the databases. It does mean you have no evidence the source exists, and you should find it or remove it.

Checking a whole list

Doing this by hand takes a few minutes per reference. PaperCitation runs the same lookups on a pasted list: it matches each entry to a published record, compares the details, and flags retracted papers. It does not call any reference fake. A reference it cannot match is marked “Couldn’t verify”, with the reason, because a missing record is not proof that a source is invented. The methodology page lists the databases it searches and what each verdict means.

A verdict of “Verified” says that a published record matches what you wrote. It does not say the source supports your argument. That part is still yours to read.