Elisabeth Bik
She scanned 20,000 papers without reading a word — then went back to audit her own complaints

1 · One strange thing

Here is someone who became famous for finding faked figures in scientific papers. Her work has contributed to nearly eighteen hundred retractions. You would assume she is an extraordinarily close reader.

She does not read a word.

During my scan of a paper, I would not read the text, but just focus on the photos. I would look for duplicated panels within the same paper …, within the same figure …, or duplications within the same photo …. Most papers took only 1 min to scan, but papers from with a lot of different figures and panels might take up to 15 min.Scanning for duplications, 2019-11-23

In 2014–2015 she went through 20,621 biomedical papers this way. By day she was a researcher at Stanford's School of Medicine working on the microbiomes of humans and marine mammals; this happened at night and on weekends, and her own phrase for it is "a strange hobby project." It took close to three years.

The result: 782 papers (4%) contained inappropriate image duplications.

That 4% was published in mBio and covered by Nature, Science and the New York Times. If the story ended there, she would simply be someone with a very good eye. What is rare is the next step: she reported all 782 to the journals, and recorded what happened to each one in a spreadsheet. Five years later she quit her job, opened the spreadsheet, and began going through it line by line.

This blog is the record of that audit.

2 · Who she is

Elisabeth Bik is Dutch. Her PhD, from Utrecht University, was on epidemic Vibrio cholerae strains. She worked at the Dutch National Institute for Health and at St. Antonius Hospital in Nieuwegein, then spent fifteen years in David Relman's lab at Stanford's School of Medicine on human and marine-mammal microbiomes. In 2014 she started Microbiome Digest, a near-daily roundup of microbiome literature, now run by volunteers.

The turn came in 2013, and the trigger was small and personal: she found that someone had plagiarized text from one of her own articles. Her About page states it flatly — "Around 2013, after finding someone had plagiarized text from one of her articles, she became interested in scientific integrity." The following year she spotted duplicated blots in a PhD thesis, and images became the work.

One stretch of her CV she has to keep explaining: from 2016 to 2018 she was Scientific and Editorial Director at uBiome, a gut-microbiome testing company. Four months after she left, the FBI raided it and its two founders were charged with fraud. So her FAQ page carries an entry whose heading is simply "You are a fraud." Someone whose profession is exposing fabrication has to first dispose of allegations about herself, on her own homepage — which says a good deal about where she stands.

In March 2019 she quit her job to do this full time, living on Patreon donations, paid consulting for publishers and institutions, and speaking fees. She opened the blog that May. As of August 2026 her work has contributed to 1,797 retractions, 295 expressions of concern and 1,321 corrections. She has the 2021 Maddox Prize, the 2024 Einstein Foundation Award and a 2025 honorary doctorate from the University of Bern.

The blog is still alive, but the rhythm changed. She wrote 94 posts across 2019 and 2020, 12 in 2021, and 4 to 8 a year since. Her own explanation is unglamorous: too many requests to screen other people's papers, too much travel to give talks — "I am getting pretty good at packing suitcases!" The work moved to PubPeer and BlueSky, and the blog became the place she stops occasionally to tell one story properly.

3 · What she is wrestling with

Not "there is fraud in science." Everyone knows that; it does not need a blog.

Her question is narrower and more embarrassing: when the fabrication is visible, printed, sitting in a PDF that anyone can open — why does nothing still happen?

Her own data drove her to it. She reported 1,016 papers to journals in 2014–2015. Four years on, in a post teaching other people how to report misconduct, she casually settled the account:

Of the 1016 papers that I reported to journals in 2014 and 2015, 54 have been retracted and 181 have been corrected as of today. The rest? … crickets (nothing happened).How to report misconduct to a journal, 2019-07-16

Seventy-seven per cent: nothing. Note that this is not "nobody noticed" and not "the evidence was thin." Every one of those papers has an email in some editor's inbox spelling out the problem with an annotated figure attached. The reporting link works. The link after it is the broken one.

Seven years later she measured again, this time against a single journal. In October 2015 she sent 30 papers to Cytokine, an Elsevier title. By June 2026: 2 retracted, 2 corrected, 26 (87%) untouched. She made the title a question and then answered it herself:

It is cases like these that are extremely frustrating. The problems are obvious, but the editors refused to take action. Is science self-correcting? Not for Cytokine.Is science self-correcting? Not in this Elsevier journal., 2026-06-24

Set that beside the last line of the blog's very first post and you have the whole spine. On 24 May 2019 she wrote: "Science builds upon science. Science should be open for self-correction." That is a statement of hope; it still contains a "should." Seven years later the same proposition is a question that her own data has closed.

Everything in between can be summarized as one long interrogation of a single thing: which part of science's self-correction machinery is actually jammed. First she suspected nobody could see the problem, so she built tools that let anyone see it. Then she found the fraud was far larger than any individual, so she went after the supply chain. Then she established that the jam is not at detection but at disposition, and turned her fire on editors. And just as she got that argument built, generative AI arrived and removed the first link — being able to see it at all.

4 · Her four ideas, told as she found them

1 · Turning "that looks wrong" into a graded ruler nobody can sue you over

After the 20,000-paper scan she had 782 findings and a problem nobody was going to solve for her: how to describe them.

This is not a rhetorical difficulty, it is a trap. The same photograph appearing twice might be an author dragging the wrong file into a figure — that genuinely happens — or it might be someone inventing data. In the final PDF the two look exactly the same. And the moment you say the second thing out loud, you are accusing a scientist of fabrication, with original files you will never have. She later put the position plainly: they can still sue you for libel or slander, even if you are correct.

Her solution was to lift intent out of the judgement and replace it with a geometric question. Three categories, decided purely on the spatial relationship inside the image:

What the scheme does is translate an accusation about a person into a description of a shape. "These two bands look unexpectedly similar" is unanswerable — it reports her perception, not anyone's conduct. And the category level says the thing she refuses to say: Category III can hardly be anything but deliberate. She writes only the first half of the sentence and lets the reader finish it.

She also built a calibration rig into the ruler. In the 20,000-paper study every duplication she found was independently confirmed by two veteran editors-in-chief, Ferric Fang and Arturo Casadevall; if either disagreed, the finding was discarded — "which happened in about 10% of the cases I found." A hobby project with its own false-positive control.

But a ruler only counts if other people can use it. So in January 2020 she did something rare in this field: she downloaded six photographs of palm trees at sunset from the free stock site Unsplash and sat down to falsify them by hand. One panel duplicated outright. One shifted so that it partly overlaps another. One with five palm trees cloned inside a single frame. Each time she asks the reader to find it first, then marks the answer in coloured boxes. She adds: "I am a really bad photo-manipulator, so it should be pretty obvious."

She did not simplify her explanation of Western blots — she replaced the Western blot. Palm trees and blots are the same class of object to her: a texture that should not repeat. Someone who has never been inside a lab can finish those six photographs and then read every annotated figure she has ever posted. It is the most elegant piece of teaching design this site has come across — using material entirely outside a field to hand you a way of seeing inside it.

2 · Fraud is a supply chain — and the tell is in the part nobody looks at

Until the end of 2019 her unit of work was always a person: this professor, that lab. One bad paper leads you to check the same author's others, and often a whole string falls out. She called these clusters.

In early 2020 something broke the frame.

It started with two PubPeer comments from a user called Indigofera Tanganyikensis, pointing out that the Western blots in two papers had the same background texture. The background — not the bands, not the data, but the featureless grey underneath them. Bik and several equally pseudonymous collaborators (Smut Clyde, Morty, Tiger) pulled on the thread. A month later they had more than 400 papers.

Different authors, different hospitals, different publishers. Shared: the same background texture; the same smooth, evenly spaced bands with none of the usual smudges or stains — tadpole-shaped, hence the Tadpole Paper Mill; the same greyscale bar charts with double-ended error bars; even the same title grammar, for which Bik drew an actual "title generator" diagram. Some papers contained flow cytometry plots of an unearthly shape, and she reproduced Smut Clyde's description of them verbatim:

The Death Star down in each frame’s quadrant Q3 is shooting out bolts of planet-smashing energy to the right through Quadrant Q4, while disciplined flotillas of Rebel Alliance X-Wing starfighters are swooping down through Q1 and Q2.Smut Clyde, quoted in The Tadpole Paper Mill, 2020-02-21

Then she counted the buyers. 101 papers from Jining No. 1 People's Hospital, 59 from the China-Japan Union Hospital of Jilin University, 23 from the Affiliated Hospital of Qingdao University. The decisive detail: within that hundred-plus from one hospital, the authors barely overlap — they are scattered across paediatrics, cardiology, endocrinology, nephrology, vascular surgery. The buyers do not know each other. Only the seller is shared.

That changes the shape of the problem completely. If fabrication is industrial, then pursuing individual scientists is cleaning a production line with tweezers. You have to look at both ends: demand — Chinese medical students and doctors need an international publication to graduate or be promoted, and their training contains no time to do research; and absorption — the 400-plus papers are concentrated in six journals, of which Artificial Cells Nanomedicine and Biotechnology alone swallowed 76. She leaves one chilling, carefully hedged line: "An anonymous source has even suggested that some of these journals might be actively working together with such paper mills," immediately followed by a note that the rumour is unconfirmed.

It is worth pausing on where the tell was. The forger carefully constructed the bands, because bands are data and data is what a reviewer looks at. What they never thought to vary was the grey underneath. A mass forger has to reuse a template, and the least conspicuous part of a template is the hardest part to vary. To catch industrial fabrication, look at the places that carry no information at all. She reuses that eye: in 2026 she caught a paper on camel udder microbiota not by its conclusions but by the texture of its sentences — someone had run a 2017 PhD thesis through a synonymizer, so "Brain Heart Infusion medium" came out as "mentality heart imbuement."

And the Tadpole post does not land on the 400 papers. It lands on a point about labour:

Our small team of volunteers has found 400 papers … although we did put in many hours of unpaid work to find these papers. One can only guess how many other falsified publications have been produced by other paper mills using less recognizable and more variable layouts. … Finding these fabricated images should not rely solely on the work of unpaid volunteers.The Tadpole Paper Mill, 2020-02-21

This is the closest thing to a manifesto in her writing. She does not ask peer reviewers to try harder — a 2025 conference remark of hers became a headline: "Detecting fraud is a job for professionals, not peer reviewers." Reviewers are unpaid, untrained and looking at one paper at a time. Only a production line can stop a production line.

3 · The real lesion is at disposition: how a "correction" launders fabrication into a typo

This is her angriest and most original territory. Original because nearly every public conversation about research integrity stops at detecting fraud, and she put all her weight on the step after.

She got here by auditing. In November 2019, going through her spreadsheet of two thousand-plus papers, she pulled out a 2005 paper in Clinical Cancer Research (impact factor around ten). She had reported it in 2014; a year later the journal issued a correction, worded like this:

In this article, … the Western blot panels, flow cytometry, and RT-PCR panels presented in Figs. 1A, 3B–E, 4, 5A and 5E were incorrect. The authors have provided new versions of these figures…. These changes do not affect the interpretation or conclusions of this work. The authors regret this error.quoted in Mega-corrections and weak editors, 2019-11-12

She lays out the originals one by one. In Figure 1A all eight β-actin bands look like the same band. In Figure 3B lanes 2–5 look like lanes 9–12. In Figure 3E lanes 1–5 look like lanes 6–10. Figures 3C and 3D show the same regions of a flow cytometry plot repeating. All five figures are affected, and nearly all of it is Category III. Her tally: fourteen doctored photographs in one paper.

I can think of very few honest errors that could explain how the same bands or groups of cells could be visible twice within the same photo.Mega-corrections and weak editors, 2019-11-12

Here is the insight. A correction is not a gentler retraction; it is a different speech act. A retraction says the record was wrong. A correction says the record was fine and something was misprinted. When an editor accepts a correction to handle image manipulation, they are not being lenient — they are permanently rewriting evidence of fabrication as evidence of clumsiness, and irreversibly, because the paper is now marked as handled. The "new versions of these figures" the authors supplied were checked against nothing.

That is what the doping analogy is aimed at: the sample taken before the race was dirty; would you accept a clean one two weeks later?

She has chased the disposition end ever since. In February 2020, the journal Oncotarget, flagged repeatedly on PubPeer, paid an outside firm for image-forensics reports and posted them with the line "As you can see in the report, no manipulation was detected in this image." Bik points out that the report may well be correct — and that it answers a different question. The complaint was never manipulation; it was that two panels overlap. Every image-forensics report she has ever seen includes similarity analysis. These three did not.

It would be like calling a plumber to check on a leak in the kitchen sink and hearing that the house is not on fire. Or, calling 911 after detecting a fire in the kitchen, and having the firefighters tell you that the sink is not leaking.No Manipulation Was Detected in This Image, 2020-02-08

By 2026 every cushion has been sanded off her phrasing:

These are what I call “5 second retractions” – no matter what the authors will tell you (“the original photos were lost in an airplane crash”; “our basement got flooded”, or “the intern did it”), there is no conceivable way that these duplications could have happened by accident.Is science self-correcting? Not in this Elsevier journal., 2026-06-24

4 · She announced, herself, that her own eye had expired

In February 2024 a review article in Frontiers detonated across the internet. Its figures were generated by Midjourney; the main one showed a rat sitting up like a squirrel with four enormous testicles and an absurdly proportioned penis, labelled with testtomcels, senctolic, dissilced, iollotte sserotgomar and diƨlocttal stem ells. Bik's comment: "At least the word 'rat' is correct." Of the second figure she says she cannot tell whether it depicts the JAK-STAT signalling pathway or how to make a doughnut with coloured sprinkles. The third, she says, looks like pizzas with pink salami and blue tomatoes.

It is the funniest post on the site and the bleakest, because she then kicks the foundation out from under the joke:

But the paper is actually a sad example of how scientific journals, editors, and peer reviewers can be naive – or possibly even in the loop – in terms of accepting and publishing AI-generated crap. These figures are clearly not scientifically correct, but if such botched illustrations can pass peer review so easily, more realistic-looking AI-generated figures have likely already infiltrated the scientific literature. Generative AI will do serious harm to the quality, trustworthiness, and value of scientific papers.The rat with the big balls and the enormous penis, 2024-02-15

Then comes the line that matters. In the same post she turns back to her own most famous result — the Tadpole mill was "perhaps one of the earliest examples of peer-reviewed papers containing computer-generated images of Western blots" — and explains exactly why it was catchable: because all the blots shared one background. And then:

But recent advances in AI technology mean we’re already past the stage where a human can distinguish a fake photo from a real photo.ibid.

Put that sentence back in the mouth of the person saying it. Someone who has scanned a hundred thousand papers by eye, turned looking into a craft, and won three international prizes for it, publicly announcing that the craft has passed its expiry date. Her core method — find the repetition — rests on a premise she has stated plainly: the forger has to reuse material. Generative models remove the need. Every fake figure can now be new.

This is the most honest thing about her, and the reason this site thinks she is worth reading: she wrote down the conditions under which her own method stops working, while still making a living from it.

5 · How she thinks

What transfers from a long-running writer is usually the method, not the conclusions. Bik's are unusually hard-edged, and every one of them can be pulled directly out of something she actually wrote.

One: look where no information is carried. This is the physical basis of all of it. A forger polishes the data, because data is what gets reviewed. They do not think about the grey behind the bands, or the texture of a sentence, or the grammar of a title. Wherever the author thinks nothing matters is where the template leaves fingerprints. The Tadpole mill fell on its background; the camel paper fell on the cadence of its synonymizer. And she can see these things precisely because she does not read the text — she is not comprehending a paper, she is inspecting a fabric.

Two: rewrite the accusation as a report on your own perception. "These two bands look unexpectedly similar," never "these two bands have been copied and pasted." She lists an entire safe vocabulary — "concern about Figure 1," "sharp transition between two adjacent bands," "appear to show repetitive background patterns" — and the words to avoid: cloned, fabricated, manipulated, misconduct. This reads like legal advice; it is actually an epistemic discipline. She asserts only the part she genuinely knows. Intent she does not know, so she does not say it — and the category grade says it for her.

Three: keep the spreadsheet, then go back and audit. This is the rarest habit here. She is not the only person to have reported a thousand papers; she is very nearly the only one who reopened the file five years later and counted what happened to each. The blog exists because she treated her own complaints as an experiment with trackable outcomes. "54 have been retracted and 181 have been corrected … The rest? … crickets." That sentence lands because she is holding the denominator.

Four: when accused of bias, don't argue — count. In late 2019 she flagged more than fifty papers by one Chinese immunologist and was immediately asked by journalists and colleagues why she was singling out a country. Her reply runs a few hundred words: in the previous six months she had posted 775 flags covering 32 countries on six continents, of which nearly 200 were from the US and 106 from China, with a world map attached.

So no, I am not specifically targeting anyone or any country. I am, however, targeting problematic papers, no matter where they come from.Targeting problematic papers, not countries, 2019-11-23

Answering an allegation of bias with a denominator is worth taking away from this page on its own.

Five: humour as armour, and as amplifier. "Oops!… I Did It Again" is Britney Spears, as she notes in the post itself; flow cytometry becomes #FlawCytometry; the Death Star plots; "at least the word 'rat' is correct"; the FAQ's drinking game, where every time someone asks why she doesn't just write software she replies "Drink!" (with a note that it is mostly tea). This is not garnish. She is explicit about what the job costs — "some of your coworkers might call you a snitch and stop talking to you. This has happened to me and it still hurts." And the AI rat probably did more for public awareness of image integrity than her previous 4,500 PubPeer comments combined. Absurdity is the one weapon she has that reaches past the field, and she aims it carefully.

Six: publish the method, not just the findings — aim at your own redundancy. Her About page says: don't send me papers, post them on PubPeer. Underneath are her six How-To guides, including a copy-and-paste template for a misconduct report and the line "You do not have to give me credit and I won't sue you for plagiarism, hahaha!" She gave away the entire Einstein Foundation prize to fund other science sleuths, for the same reason: this kind of work fits no research funder's hypothesis-driven application form, so it can never be paid for.

6 · Where she may be wrong

The biggest one: her instrument is sensitive to a single kind of fabrication, and it is not obviously the worst kind. She does not read the text; she looks at pictures. That means her method can only catch fabrication that is visually lazy. A paper with beautifully invented numbers, clean figures, or no photographs at all is completely transparent to her. And the literature's real maladies — irreproducibility, p-hacking, selective reporting, wholesale invented datasets — are mostly not image problems. She has never claimed image fraud is the whole picture, but a blog that has shone a light on one lesion for seven years makes it easy for a reader to mistake the lit patch for the room.

Second: she polices an industry that partly pays her. Her disclosure page is exceptionally candid: since 2019 she has done paid consulting for multiple scientific publishers and research institutions (almost all under confidentiality agreements, so neither clients nor cases can be named); from 2021 to 2024 she sat on the eLife Ethics Committee for an annual stipend; she has free access to ImageTwin, Proofig and iThenticate, three commercial screening tools. She imposes a rule on herself — she will not take a "second opinion" engagement on any paper she has publicly flagged — and the rule is serious. But confidentiality means the self-discipline cannot be externally checked, and the free tool licences come from vendors competing for publisher contracts, for whom her visible use is the most valuable endorsement available. Putting all of it in public is to her credit. It does not make the tension go away.

Third: her answer on anonymity is incomplete. Most of her flags go on PubPeer, many pseudonymously, and several of her long-term collaborators (Smut Clyde, Morty, Tiger, Clare Francis) are pseudonyms throughout. The weightiest named objection comes from Michael R. Blatt, then editor-in-chief of Plant Physiology, in his 2015 editorial "Vigilante Science" (DOI: 10.1104/pp.15.01443). His two central complaints: that "most individuals posting on PubPeer … take advantage of the anonymity afforded by the site in full knowledge that their posts will be available to the public at large"; and that "the vast majority of comments that are posted focus on image data (gels, blots, and micrographs)," where "commenters on PubPeer do no more than flag perceived faults and query the associated content." Bik's reply, scattered across her FAQ and reporting guides, is essentially that anonymity is the only protection a weaker party has and that retaliation is real — in 2021 the French microbiologist Didier Raoult's side filed a criminal complaint against her for aggravated moral harassment, attempted blackmail and attempted extortion. (Per contemporaneous reporting in Nature, Science and The Scientist, more than a thousand scientists signed a letter supporting her and the complaint went nowhere. I have not read the complaint itself; what is stated here is the documentary fact as reported.) That reply is forceful, but it answers why anonymous critics need anonymity. It does not answer Blatt's other half: when the critic bears no cost of attribution, how is the burden of proof distributed? One comment and one coloured box can mark a paper permanently, while the cost of clearing one's name is far higher. Her own annotations are of very high quality. She is not the only person using the mechanism.

Fourth, and the one that worries her most: her output is being taken and used by other people. In July 2025 she quoted, in her own roundup, an interview she gave to Nature:

“We try to point out those bad papers because we still believe in science and want to make science better,” says Elisabeth Bik … But, she adds, “I am very worried about how the work we do in pointing out bad papers is currently being misused, or even weaponized, to convince the general public that all science is bad”.Miryam Naddaf, Nature, July 2025, quoted in Science Integrity Digest Summer 2025

The diagnosis is exact; the remedy she does not have, and may not be able to have. Her blog does two opposite things at once: it dismantles the credibility of Frontiers, Elsevier and Oncotarget one paper at a time, and it dismantles shoddy preprints claiming COVID vaccine harms one paper at a time. Inside her method these are the same act — both are bad papers. In public argument they are each other's ammunition. Someone whose method is publicly pointing out error cannot control who picks up the evidence. That is probably the deepest bind in this kind of work, and in seven years she has named it without solving it.

7 · Sources

• Scanning for duplications2019-11-23 · The full account of the 20,000-paper scan: how she scanned, how long each took, the definitions and examples of all three categories, and how independent confirmation by two editors-in-chief threw out a tenth of her findings. Nearly all of the method section above comes from this post.
• Types of image duplications: the palm trees2020-01-08 · Six sunset palm trees, falsified step by step into all three categories, with the reader asked to find each one before the answer is shown. A lesson that requires no biology at all.
• The Tadpole Paper Mill2020-02-21 · More than 400 papers sharing one blot background, broken down by journal and by hospital, with a "title generator" diagram and the closing line about unpaid volunteers.
• No Manipulation Was Detected in This Image2020-02-08 · A journal buys an image-forensics report that answers a different question. The plumber-and-fire analogy is here.
• Is science self-correcting? Not in this Elsevier journal.2026-06-24 · The most recent post as of this page's review date. Eleven years of accounts settled against one journal: 30 papers reported, 26 untouched. The definition of a "5 second retraction" is here too.

Other posts quoted above, in order: A new Science Integrity Blog (2019-05-24), How to report misconduct to a journal (2019-07-16), Mega-corrections and weak editors (2019-11-12), Targeting problematic papers, not countries (2019-11-23), Don’t correct science papers with manipulated photos – Retract! (2019-12-19), Western blots (2020-02-21), Thoughts on the Gautret et al. paper (2020-03-24), 2020: A year in review (2020-12-31), Cassava Sciences: Of stocks and blots (2021-08-27), The rat with the big balls and the enormous penis (2024-02-15), ScienceGuardians (2025-04-18), Science Integrity Digest Summer 2025 (2025-07-25), The Camel’s Camel (2026-03-12), Reporting from the 2026 WCRI (2026-05-20), plus the two standing pages About and FAQ. Blog home: scienceintegritydigest.com.