The Damage Is Already Done

In 2023, a fabricated reference in a biomedical paper was rare enough to be a fluke: one in 2,828. By 2026, you'd find one in every 277 papers published in the first seven weeks of the year. Twelvefold, in two years. The number comes from an audit of nearly 2.5 million PubMed-indexed papers, published as a letter to The Lancet by Maxim Topaz of Columbia University's Data Science Institute and colleagues, and reported by Retraction Watch.

A bibliography is supposed to be the boring part of a paper. Nobody peer-reviews the reference list with the same scrutiny they give the methods section. That turns out to be exactly the point.

What they actually did

Topaz's team used AI to do the checking, which is its own small irony. The tool's job was to distinguish genuine fabrications from formatting discrepancies. An informally abbreviated title is just sloppy typing. Sorting the two apart at scale is the entire trick.

They verified 97.1 million references. Out of that pile, they identified 4,406 fabricated references, appearing across 2,810 papers. Small numbers next to millions. Large enough, apparently, to move the needle from one-in-thousands to one-in-hundreds inside three years.

The team traced the sharpest increase in hallucinated references to mid-2024. They note that this coincides with the rise of AI writing tools, a correlation the letter states plainly, without needing to oversell it. The literature had help. Something new got handed to millions of authors, and a fraction of them used it without checking the output.

The part that should bother you more than the number

Here's the detail that outlasts the headline: over 98% of the articles found to have fake references had seen no publisher action at the time of the audit in February, based on data from the Retraction Watch Database and elsewhere. The fabrication gets published, gets cited by someone else's honest bibliography and nobody touches it.

The finding has company. Nature reported the previous month that tens of thousands of publications from 2025 might include invalid references generated by AI. Two separate investigations, two similar shapes: the fabrication happens quietly, and the correction, if it comes, happens later, if at all.

What to do about it, and nobody agrees

A commentary accompanying the analysis, written by Howard Bauchner (former editor of JAMA) and Frederick Rivara (former editor of JAMA Pediatrics), argues for the hard line: retract. Their reasoning is procedural. Researchers "incur responsibility for the entire content of that paper" the moment they agree to be listed as authors. Retraction, they write, might lead to greater scrutiny of references by authors going forward, accountability as a training exercise, applied after the fact.

David Resnik, an integrity researcher at the National Institutes of Health, disagrees with the blanket policy. He argues the decision depends on the role the citation plays in supporting the study's results. A fabricated reference propping up the central claim is a different animal from one buried in the introduction, cited for context nobody's contesting.

Topaz himself lands somewhere in between, and it's the most useful position in the whole piece. He told Retraction Watch an article should be retracted when the hallucinated references are central to the conclusions of the paper, and gave an example from the analysis where 18 of 30 references appeared to have been fabricated. That's a paper built mostly on citations to nothing.

But Topaz also noted that 91% of the articles in his group's dataset of problematic papers had only one or two fabricated references, the kind he says are likely honest mistakes by authors who used AI tools without verifying the output. Most of the contamination comes from someone trusting a tool that was never built to know the difference between a real paper and a plausible sentence shaped like one.

Then Topaz said the line that gives this post its title. The contamination of over 4,000 fabricated references his team found survives the AI getting better. The damage is already done. Fixing the model doesn't un-publish the paper.

The critics have a point too

Not everyone bought the audit at full strength, and the letter's own honesty about that is worth noting. Ella Flemyng, head of editorial policy and research integrity at Cochrane, called the findings serious but flagged that the AI approach was validated on only 500 records, and said considerable details about the methods were lacking. A twelvefold headline built on a 500-record validation set is a headline worth reading carefully before you repeat it.

Mohammad Hosseini, who studies AI ethics and research integrity at Northwestern University's Feinberg School of Medicine, went further, calling the analysis "simplistic", low-hanging fruit, the tip of the iceberg. His point, reading between the lines the article gives us: catching fabricated citations is the easy version of this problem. The harder version is everything the AI wrote that sounds plausible and can't be checked against a database at all.

Both critiques land without contradicting the underlying finding. The number is real. Whether the number captures the full scale of the problem is a separate, harder question. The researchers themselves don't pretend to have closed it.

What the publishers said, and didn't

A Taylor & Francis spokesperson told Retraction Watch the publisher is investing in technology, specialist staff and processes to catch problematic citations, and said submissions may be rejected if concerning references substantially impact manuscript integrity. Renee Hoch, head of publication ethics for PLOS, said they're exploring options for system-wide reference integrity screening, and noted PLOS treats a fabricated reference as something short of automatic misconduct, a distinction that matters if you're the author who trusted a bad AI output and didn't intend to deceive anyone.

Elsevier, Wiley, Springer Nature, IEEE and Sage were also contacted. They didn't respond within the short timeframe the Lancet embargo allowed. Draw your own conclusions about five academic publishers going quiet on the exact week their journals were the subject of the story.

The 2am version

Ninety-one percent of the contaminated papers in Topaz's dataset had one or two fabricated references. That's the actual shape of this problem: a large number of authors who asked a tool for sources and didn't check before it went into the bibliography. The tool was fluent. The paper it pointed to didn't exist. Nobody involved set out to lie; the sentence just looked exactly like the ones that were true.

That's the trap for anyone finishing a draft at 2am with a reference list that needs six more sources and a deadline in the morning. A fabricated citation from a language model looks like every other citation in your document, the same format, with a publication year that happens to fit the argument you're making. You find out it isn't real when someone downstream goes looking, the way the audit went looking across 2.5 million papers and found 4,406 that weren't there.

Checking every citation against the live record before you submit leaves the contaminated literature contaminated. Topaz is right that the damage already done stays done. It stops your name from being added to the pile, before a reviewer, an editor or somebody's own audit two years from now finds the empty space where your source should be.