Why checking your citations matters more now that AI can invent them

A citation is a small promise to your reader. It says that a work exists, and that it says what your sentence claims it says. Most readers will never check. They take the promise on trust, and some of them build on it: they cite your paper for the claim, design a study around it, or repeat it in a review. That trust is what lets research add up, one paper on top of another.

The promise was never kept perfectly. What has changed is how easy it has become to break it without meaning to. Here is what the evidence shows since large language models (LLMs) arrived, and why checking your own citations matters more now than it did a few years ago.

A wrong citation does not stay small

Citation errors are older than any chatbot. A 2015 review that pooled 28 studies of citation accuracy in medical journals estimated that about a quarter of citations were wrong or problematic as support for the statement they were attached to⁠[1], and that nearly 12% did not support it at all.

And a wrong citation does not stay where it starts. A 2009 study in The BMJ traced every English language paper in PubMed that addressed one claim: that beta amyloid, a protein that builds up in the brain in Alzheimer's disease, is produced by and injures the muscle of people with a rare disease called inclusion body myositis. Across 242 papers and 675 citations, it found that citation itself had given the claim "unfounded authority."⁠[2] Papers that refuted or weakened the belief were cited less. Papers that presented no data on it expanded it anyway. In places, a hypothesis became a fact "through citation alone." The same patterns turned up in the text of grants funded by the US National Institutes of Health, sometimes used to justify the request for funding.

That is what a single reference carries. Other people build on it, and an error grows with every paper that repeats it.

What changed when large language models arrived

LLMs are now part of how many papers get written. By tracking style words whose use jumped after these models appeared, in more than 15 million biomedical abstracts, one study estimated that at least 13.5% of the abstracts from 2024 were processed with an LLM⁠[3], a lower bound that reached 40% in some groups of papers.

The trouble is what these models can do with references. In March 2026, researchers asked three recent LLMs, GPT-5.3, DeepSeek-V3 and Grok-4, for citations on ten core topics in neurocritical care, without letting them search the web. Of the 300 references that came back, 85, or 28.3%, were completely fabricated⁠[4], and 55% contained an inaccuracy. Fabrication ranged from 8% to 50% depending on the model, and the authors' warning applies to all of them: "fabricated references can appear complete and credible."

Letting a model search does not settle the matter. A 2025 study in Nature Communications tested seven popular LLMs on 800 medical questions, checking 58,000 statements against the sources cited for them, and found that between 50% and 90% of the answers were not fully supported, and sometimes contradicted, by the sources they cited⁠[5]. Even for GPT-4o with web search, about 30% of individual statements were unsupported.

Invented references are reaching the published record

These references do not stay in chat windows. In May 2026, The Lancet published an audit of nearly 2.5 million biomedical papers from PubMed Central's open access collection, published between January 2023 and February 2026. The authors checked 97.1 million references against the records their identifiers pointed to, and counted a reference as fabricated only when the work it named could not be found in any database they searched. They found 4,046 fabricated references in 2,810 papers⁠[6]. In 2023, about one paper in 2,828 had at least one. By 2025 it was one in 458, and in the first seven weeks of 2026, one in 277. At the time of the audit, 98.4% of the affected papers had received no action from their publisher.

The sharpest rise came in mid 2024, which the authors note coincided with the rise of AI writing tools⁠[7]. Many of these may not be fraud at all. Of the affected papers, 91% had only one or two fabricated references, and the lead author, Maxim Topaz, told Retraction Watch that many of them "are likely honest mistakes by authors who used AI tools without verifying the output."

Peer review did not catch them either. The Canadian startup GPTZero checked more than 4,000 papers accepted at NeurIPS 2025, one of the top conferences in AI research, and reported invented citations in at least 53 of them⁠[8], each flagged citation verified by a member of its team. Every one of those papers had been assigned three or more reviewers. NeurIPS responded that an incorrect reference does not necessarily invalidate a paper's content, and that is fair. But none of those reviewers caught the references, and the papers were published with them.

Books are not exempt. In 2025, Springer Nature retracted a book on machine learning after Retraction Watch reported that it contained many citations to works that do not exist. The retraction notice says the publisher "was unable to verify the source of 25 out of 46 references in this book."⁠[9]

The quieter error: a real paper that does not say it

An invented reference is the dramatic case, and the easier one to catch, because a search settles it. The harder case is a real paper cited for something it does not say. It exists, its record matches, and it passes any check that only asks whether the work is real. It still does not support the sentence. That is the error the 2015 review was counting⁠[1] long before chatbots, and the one the Nature Communications study⁠[5] found in 50% to 90% of the answers it tested.

Only reading catches it: opening the paper, finding the passage, and making sure it says what your sentence says, down to who was studied and how large the effect was. If you cannot find the passage, the citation does not belong on that sentence. Together with a check that the work exists and has not been retracted, that is a short routine you can run on any source⁠[10].

Why it falls to you

It would be easier if someone downstream caught these errors, and the Lancet authors do call on publishers to verify references at submission⁠[6]. But the rules already in place put the job on the author. The recommendations of the International Committee of Medical Journal Editors say that people are responsible for anything they submit that was made with the help of AI⁠[11], that such output "can be incorrect, incomplete, or biased," and that "humans must ensure there is appropriate attribution of all quoted material, including full citations."

The 2015 review, written long before chatbots, already listed among its remedies "declarations by authors that they have checked cited material."⁠[1] The reference list goes out under your name, and you are the only person who knows whether you opened each source.

Make checking part of how you write

None of this means giving up AI tools. It means changing what you let them do with references. Treat any reference a chatbot gives you as a lead, something to search for and open, never something to paste. Add a source to your reference list only after you have read the part you are citing, and check each citation the moment it enters your work⁠[12], while the paper is still open in front of you. Then check again before you submit: a paper can be retracted after you cite it, and a sentence can drift away from its source as you rewrite.

That last check is the one Researcher+ is built to make easier. In your draft, Check citations reads each sentence you cite against the paper it cites and shows you the passage it found, with a link that opens the paper at that page. A citation the paper does not clearly support is marked for you to review, never changed, and a reference with no PDF in your project is checked against its published record. What stands is still your decision.

Large language models made a reference cheap to produce and easy to believe. They did not change what a reference is for. Every citation in your paper is a promise that someone will rely on, perhaps years from now, and the way to keep it is the oldest habit in research: open the source, and read it.