One In 277 Scientific Papers Cites A Study That Doesn't Exist (Fabricated citations & AI research)

Category: Research & Learning

The footnote that leads nowhere

Here's something that happened while most of us weren't looking.

In the first seven weeks of 2026, roughly one in every 277 scientific papers published on PubMed, the vast medical research database that doctors, researchers and health authorities actually rely on, contained a citation to a study that does not exist.

Not misquoted. Not out of date. Not misunderstood.

Fabricated. A reference, sitting in the footnotes, pointing at a paper that nobody ever wrote.

And the trend is not gentle:

• 2023: about 1 in 2,828 papers.

• 2025: about 1 in 458.

• Early 2026: about 1 in 277.

That's roughly a twelvefold increase in two years. And researchers were able to date the inflection point fairly precisely, it accelerates sharply from around mid-2024, which lines up almost exactly with the mass adoption of AI writing tools in academia.

I want to explain what's actually going on here, because it's more interesting and less apocalyptic than the headline suggests, and because the practical lesson at the end is one of the most useful things I can give you.

Why an AI invents a citation

To understand this you need to understand what these tools actually are, and almost nobody does, including many of the people using them.

A large language model is not a database. It is not looking things up. It has no library.

It is, in essence, an extraordinarily sophisticated prediction machine. It has read a colossal amount of text and learned, in immense statistical detail, what words tend to follow other words.

Think of it as the world's most well-read parrot. It has absorbed the shape of everything ever written, and can reproduce that shape flawlessly, without any concept of whether what it's producing is true.

So when you ask it for a citation, here is what happens. It knows what an academic citation looks like:

Author surname, initial. (Year). Plausible-sounding title. Real Journal Name, volume(issue), pages.

And it fills that shape in with things that fit the pattern. A real-sounding author. A real journal. A plausible year. A title that sounds exactly like a paper someone would have written on that topic.

The result looks perfect. It is indistinguishable, at a glance, from a genuine reference.

It just doesn't exist.

And here's the thing I really want you to grasp: the machine is not malfunctioning. It is doing precisely the job it was built to do, produce plausible text. The failure isn't in the tool. The failure is in using a plausibility engine for a job that requires truth, and then not checking.

Part two: how it gets into a serious journal

Fair question. Aren't there checks?

Yes. And here's how they fail.

A researcher, overworked, under pressure to publish, writing in their second or third language, uses an AI tool to help draft a section. The AI produces beautiful, fluent prose with confident citations.

The researcher, reasonably busy, skims them. They look right. They're in the right format, from journals they recognise.

Then it goes to peer review, where other scientists check the work. Except peer reviewers are unpaid, overloaded, and typically checking the argument and the method, not personally re-downloading all 60 references to confirm each one is real.

So the fake reference passes through, gets published, and now sits in the permanent scientific record, where the next researcher may cite the paper containing it, and so on.

One of the researchers who uncovered this trend admitted that he'd very nearly been fooled himself, despite actively hunting for exactly this problem.

That's the part that should get your attention. Not "careless people are being fooled." The experts, looking for it, nearly missed it.

Part three: the wider rot (this bit is worse)

Fabricated citations aren't the whole story, and I'd rather give you the full picture than the comfortable half.

A large analysis of retracted papers, papers formally withdrawn after publication, found:

• Over 6,400 retractions linked to fake peer review. (Yes: researchers inventing fictional reviewers, or nominating friends, to approve their own work.)

• Around 2,300 tied to "paper mills" - actual businesses that manufacture scientific-looking papers and sell authorship to people who need publications for their careers.

• Around 2,100 citing AI-generated content as the reason for withdrawal.

And independent estimates suggest somewhere between 1.5% and 3% of all recently published scientific papers may be paper-mill products.

Sit with that. Somewhere between one-in-sixty and one-in-thirty scientific papers may not be real science at all, just content, manufactured to look like it.

And the consequence isn't abstract. Fake findings get swept up into systematic reviews, the documents that pool many studies together, which then feed into clinical guidelines, which is what your doctor actually follows.

The fabrication doesn't stay in the journal. It walks into the surgery.

Part four: now let me stop you panicking

Here's where I want to pull hard on the reins, because there's a very tempting and very wrong conclusion available: "see, science is corrupt, you can't trust any of it."

That's not what this shows. In fact it's almost the opposite, and I want to explain why carefully.

How did we find out about all this?

Scientists. Auditing 2.5 million papers. Publishing their findings openly. In the scientific literature. So that other scientists could check their work.

The system caught this. That's what the system is for. Science is not a body of facts handed down from on high, it's a process for finding and removing errors, and it is the only human institution that has ever systematically got better at admitting it was wrong.

A retraction is not a failure of science. A retraction is science working. The alternative, a system that never retracts anything, isn't a system with no errors. It's a system that doesn't look for them.

So no, this isn't a reason to distrust science. But it is a reason to distrust any individual paper, which is a very different and much more useful stance.

Part five: the toolkit (this is the bit to keep)

Here's how to actually read research claims without being had. Five habits.

1. One study is not evidence. It's a hint. The single most common mistake in the world. A lone study is an interesting suggestion. It is not a fact. Facts, in science, are things that multiple independent teams found repeatedly. Ask: has this been replicated?

2. Be maximally suspicious of the exciting result. Here's the awkward truth: the more surprising a finding, the more likely it is to be wrong. Boring findings that confirm what we already suspected are usually right and never get reported. The study that makes the front page is, by selection, the one most likely to be a fluke.

3. "A study found" is a red flag, not a credential. Which study? Which journal? Which year? Who funded it? If a claim can't survive those four questions, it isn't a claim, it's a vibe. Anyone using research to persuade you and refusing to name it is telling you something important about their confidence in it.

4. Check who paid. Not because industry-funded research is automatically fraudulent, it isn't, and much of it is excellent. But funding shapes which questions get asked and which results get published. It's a legitimate thing to weigh.

5. Ask whether the metaphor came from the scientists. Memorable comparisons, "a credit card of plastic every week," "a spoonful in your brain", are almost always generated by communicators, not researchers, and are almost always more dramatic than the underlying finding. If the sticky image is what you remember, go and check what the paper actually said.

Why this belongs in a publication about money

Because this is exactly, precisely how people get fleeced.

Every dubious financial product, every miracle supplement, every crypto scheme, every wellness grift, arrives wearing a lab coat. "Studies show." "Research proves." "Backed by science."

And it works, because the overwhelming majority of people have never been taught that "a study" is not a magic word, and that in 2026, a citation can be literally invented by a machine and published in a real journal.

The people trying to sell you things know this. They're counting on it.

So the single most valuable skill I can give you isn't a stock tip, an investment theme, or a prediction.

It's this: the ability to be appropriately, calmly, cheerfully sceptical. Not cynical, cynics believe nothing and are just as easily fooled, in the opposite direction. Sceptical. Willing to ask one more question than the person expected you to ask.

Research is more valuable than it has ever been, and simultaneously harder to verify than it has been in decades.

Both of those things are true at once. The right response isn't to give up on it.

It's to check.

Next in this series: back to the beginning, why the world just declared bankruptcy, and paid for it in water.

References

Retraction Watch - "One in 277 PubMed-indexed papers in 2026 shows fabricated references, says analysis": https://retractionwatch.com/2026/05/07/one-in-277-pubmed-indexed-papers-in-2026-shows-fabricated-references-says-analysis/

STAT News - "Fraudulent citations, blamed on AI hallucinations, are becoming more common in research papers": https://www.statnews.com/2026/05/07/lancet-study-finds-steep-rise-fraudulent-citations-academic-papers/

Retraction Watch Database - retraction statistics and causes: https://retractionwatch.com/retraction-watch-database/

PMC — "Combating Fake Science in the Age of Generative Artificial Intelligence: A Biomedical Perspective": https://pmc.ncbi.nlm.nih.gov/articles/PMC12309808/

COPE (Committee on Publication Ethics) - guidance on paper mills and systematic manipulation: https://publicationethics.org/

Ioannidis, J.P.A. - "Why Most Published Research Findings Are False," PLoS Medicine: https://doi.org/10.1371/journal.pmed.0020124