One Every Five Days

This is not about anyone. I have a rule, which is that I do not comment on live integrity cases involving named individuals, and the reason is in section five. But the questions circling those cases are worth answering in the abstract, because they are questions early-career researchers ask me privately and get mush in return: how do I make sure I am not plagiarising? Who should be on this paper? What should happen if a colleague wants a favourable review? Is that person’s publication list even possible?

So: the rules, the evidence, and the part where I refuse to draw the conclusion you may want me to draw.

What a big number is, and is not, evidence of

The modern version of this argument starts in 2018, when Ioannidis, Klavans and Boyack looked for authors who had published more than 72 papers in a single calendar year — one every five days — at any point between 2000 and 2016, and found thousands of them.1 The threshold was chosen because it is the point at which a straightforward reading of authorship stops being physically credible.

Their 2024 follow-up lowered the bar to more than 60 articles a year and produced the number that matters for this discussion: 3,191 such authors outside physics, against 12,624 inside it.2 That ratio is the whole caveat. Particle physics runs on consortium papers with author lists in the thousands, where every member of a collaboration is credited by prior agreement. Nobody in that field is pretending to have written 200 papers. The convention is public, understood, and priced in by everyone reading it.

Outside physics there is no such convention, and this is where the honest analysis gets harder rather than easier. A 2026 study of sports medicine and musculoskeletal health went through 16,983 articles and 68,209 unique authors in the top-ranked journals of the field between 2020 and 2024. It found 222 extremely productive authors — 0.45% of everyone.4 And then it found something that complicates the easy story: those authors had a mean h-index of 79.9 and mean career citations of 35,654. Their work is not junk. The authors conclude, reasonably, that high productivity does not by itself equate to low quality.

The detail I cannot stop thinking about is the position on the byline. In that sample the extremely productive authors occupied middle-author positions a median 60% of the time, and were first author on between 1.9% and 2.1% of papers.4 That is a portrait of a particular kind of career — the senior figure whose group, network and methods are genuinely everywhere — and it is entirely compatible with both an exemplary research leader and a person collecting names on work they have not read. The bibliometrics cannot distinguish between those two people. Neither can you, from the outside.

A systematic review published in Scientometrics in 2026 makes this point more carefully than I can. Having surveyed 18 articles addressing hyperprolific authorship and 79 further contributions on related behaviours, it reports that there is no consensus at all on the threshold or the method for identifying such authors, and concludes that hyperprolific authorship should not be reduced to a single category of misconduct but treated as a multifaceted, context-dependent phenomenon.3

I want to be blunt about what that means, because it cuts against the direction of travel in every comment thread on this subject. If the field that studies this professionally cannot agree where the line is, then a number on its own is not a finding. It is a prompt to ask a specific, answerable question: what did this person actually do on this paper?

Who goes on a paper, and how to decide it

That question has a published answer, and it has had one for years. The ICMJE criteria require all four of the following: substantial contribution to conception, design, acquisition, analysis or interpretation; drafting or critically revising for important intellectual content; final approval of the version to be published; and agreement to be accountable for all aspects of the work.6 The same document says explicitly that acquisition of funding, general supervision of a research group, general administrative support, and writing, technical or language editing do not on their own justify authorship. People who contributed but do not meet all four should be acknowledged, with their role named.6

Almost everyone in the life and health sciences knows this. Look at what happens anyway. A meta-analysis of nineteen surveys found that when researchers were simply asked whether any co-author was honorary, 26% said yes; when the ICMJE criteria were put in front of them first, 18%; and when researchers declared what each co-author had actually done and those declarations were checked against the criteria, 51%.5 Ten per cent reported being approached to add an honorary author; 16% admitted having added one.5

The gap between 26% and 51% is the interesting one. It is the distance between what people think they are doing and what they are doing. Most gift authorship is not a conspiracy. It is a lab norm nobody has written down, a head of department who has always gone on everything, a collaborator who sent a reagent, a supervisor added out of gratitude or fear. The practice survives precisely because everyone involved experiences it as courtesy.

So the practical advice, which is duller than the debate:

• Have the authorship conversation at the start, in writing. Not at submission. At the point the project is designed, when nobody yet knows whose contribution will turn out to matter, which is exactly when people are fairest.
• Use a contributorship statement and mean it. Write down who did what, in named roles, and circulate it before submission. The 51% figure exists because contributions were written down and then compared to a standard. Do that yourself and the problem largely self-diagnoses.
• Decline authorship you have not earned, out loud. This is the only lever senior people actually control, and it is worth more than any policy. A professor who says “take me off, I only funded it, put me in the acknowledgements” changes a lab’s norms in one sentence.
• If you cannot describe your intellectual contribution to a paper in one sentence, you are not an author of it. That is not an ICMJE rule. It is mine, and it has never once been difficult to apply.

Plagiarism: the boring rules that actually work

Plagiarism is presenting someone else’s words, ideas, data or images as your own. That definition is easy. What is hard is that almost nobody who commits it set out to.

The scale is not trivial. A February 2026 bibliometric analysis of 46,087 retractions across ten major publishers, drawn from the Retraction Watch database covering 1997–2026, found plagiarism named in 30.8% of them — 14,187 records — making it the third most common reason after concerns about results or conclusions and third-party involvement.7 The same study found retraction lags ranging from a median of 41 days at one publisher to 1,568 days at another.7 Four years is a long time for a paper to be cited before the correction arrives.

Here is what I tell my own students, and it is all mechanical rather than moral:

• Never paste source text into your draft file. Not even temporarily, not even with a note saying “fix later”. Every serious plagiarism case I have seen up close began as a placeholder that survived to submission. Read the source, close it, then write the sentence from your notes.
• Cite at the point of the idea, not at the end of the paragraph. A citation parked after four sentences leaves three of them looking like yours.
• Rewriting is not laundering. Substituting synonyms into someone else’s sentence structure is still their sentence. If the idea is theirs, the citation is theirs, no matter whose words carry it.
• Methods sections are the usual trap. They are formulaic, they get reused, and reviewers rarely read them closely. That is exactly why they show up in overlap reports.
• Figures, images and code count. A redrawn figure is still a derived figure. Say where it came from.
• Your own text is not a free resource either. Text recycling has legitimate uses, particularly in methods, and COPE publishes guidance for editors on handling it,8 but it needs disclosure. Note that the guidance reaching authors is genuinely thin: a survey of 209 epidemiology, public health and general medicine journals found only 12.9% had an explicit policy on salami publication and 35.9% on duplicate publication.9 Where journals have not decided, the responsibility falls to you.
• If your name is on it, you are accountable for all of it. That is the fourth ICMJE criterion,6 and it is the one that ends the “my student wrote that section” defence. Supervisors do not get to sign for the credit and not the liability.

Then there is the new problem. An analysis published in March 2026 examined 7,251 articles in one major medical journal between January 2022 and March 2025 and classified 2.7% as containing significant AI-generated text, with the detected rate climbing from 0.0% to 11.3% across the period. Just 0.2% of articles disclosed using a large language model at all.14 Detection tools are imperfect and the authors treat their estimates as detection rates rather than ground truth, which is the right posture. But the disclosure gap is the finding, and it does not depend on the detector being precise.

I work on these models and I use them. My position is not that you should not. It is that a language model produces text whose provenance is unknown to it and therefore unknown to you, and that plagiarism is defined by what appears under your name rather than by your intent when it got there. Use the thing to argue with, to restructure, to catch your own bad sentences. Do not use it to generate prose you have not verified against sources you have actually read, and declare it where the journal asks.

Collusion: the one we cannot currently catch

Fabrication gets the headlines. Plagiarism gets the software. Collusion gets almost nothing — no detection tool worth the name, very little published prevalence data, and until recently not much policy either.

It is also the failure mode least likely to announce itself as one. Fabricating data requires a decision to fabricate. Agreeing to look kindly on a colleague’s submission can be arranged in a corridor, in the language of collegiality, by people who would be genuinely offended at the word corruption.

The mechanism deserves explaining, because it is specific and it is boring. Most computer science venues let reviewers bid on which submissions they would like to handle, and an automated matcher uses those bids to allocate papers. A group that agrees in advance to bid for each other’s submissions can steer the matcher into assigning them to one another, at which point the reviews write themselves. Michael Littman set this out in Communications of the ACM in 2021, describing collusion rings as sets of authors who review and support each other while breaking anonymity and concealing conflicts of interest.15 It is a coordination problem, not a technical exploit. Nothing is hacked. The system is simply told a set of lies about interest.

The genuinely bad news is that we cannot reliably catch it. Jecmen, Shah, Fang and Akoglu tested whether collusion rings can be detected from bidding patterns, using two realistic conference bidding datasets and the fraud-detection algorithms that work in other domains. Undetected colluders achieved assignment to up to 30% of the papers written by other members of their ring. And when ten colluders bid on all of each other’s papers, no detection algorithm produced a candidate group with more than 31% overlap with the actual colluders.16 Their conclusion is that collusion cannot be effectively detected from bidding using existing tools.

Read that against the volume question in section one and it should be sobering. We have decent instruments for detecting copied text and manipulated images. We have almost none for detecting an arrangement between four people who trust each other.

Venues know it. The NeurIPS 2026 handbook states that the conference “does not tolerate any collusion whereby authors secretly cooperate with reviewers, ACs or SACs to obtain favorable reviews”, and lists the consequences: immediate removal from the reviewing system, rejection of all papers under consideration, sharing of identities with sister conferences, informing the colluding parties’ home institutions, and sanctions on future participation.17 The bidding instructions carry their own warning — “Your bids are an important input to the paper matching process. Please be cognizant of our anti-collusion policies.”17 A conference does not write that sentence into its handbook as a hypothetical. There are also structural repairs available: assignment algorithms can be constrained to exclude review cycles outright, which is computationally hard in general but tractable enough in practice for a working heuristic.19

Collusion also has a milder, far more widespread cousin, and this one has hard evidence behind it. Adrian Barnett analysed more than 37,000 reviews of over 12,000 articles across four journals that run fully open peer review. Roughly 6% of reviews requested a citation to the reviewer’s own work. And where authors complied with the request, the article’s probability of approval was 92%, against 76% where they did not — an odds ratio of about 3.5.18 The picture is not simple: reviewers who asked for self-citations were, overall, markedly less likely to approve, which suggests many such requests accompany genuine criticism rather than a shakedown.18 But the compliance effect is the finding that should worry us. Adding the citations helps. Authors work that out quickly.

So, plainly, what I think good practice requires:

• Bid on what you are qualified to judge, and nothing else. A bid is a declaration of expertise, not a favour you are in a position to grant.
• Declining is the floor, not the ceiling. Refusing protects your own conduct and leaves the arrangement entirely intact for whoever gets asked next. Reporting — to the programme chairs, to an editor, to your institution — is the only step that removes it. It costs the reporter time, awkwardness and sometimes a working relationship; it costs the person who asked almost nothing to try again elsewhere. That asymmetry is a large part of why these arrangements persist, and no detection algorithm repairs it.
• Never review a paper you would struggle to describe as an arm’s-length judgement. Conflicts of interest are declared, not weighed up privately and waved through.
• Ask for citations only where you would defend the request in the open. If you would not put your name to it, it is not a scientific objection.
• Confidentiality is part of it. The same NeurIPS handbook requires that nothing about a submission be shared or discussed with anyone outside the assigned reviewers.17 Collusion rings run on side channels, and every casual conversation about a paper you are reviewing widens one.

I wrote in MyTh 001 about reviewers outsourcing their judgement to language models, and argued that it was a workload symptom before it was an ethics failure. Collusion is not that. Nobody drifts into a collusion ring because their department shed a fifth of its staff. It is deliberate, it is coordinated, and it is the one item on this list for which I can find no structural excuse at all. It is also the one that most of the advice in this entry cannot reach. Everything else here is a habit you can build cheaply in advance — a citation style, an authorship conversation, a rule about never pasting source text. Collusion is not defeated by a habit. It is defeated by somebody being willing to tell a chair what they were asked to do, which is a harder thing to ask of people and a much rarer thing to observe.

Why I will not tell you what I think about a live case

Now the part that this entry exists to say.

An allegation is not a finding. The machinery for turning one into the other is unglamorous and it matters. In the United States, a substantially revised version of the Public Health Service regulation on research misconduct — 42 CFR Part 93 — applied to all institutions from 1 January 2026, the most substantial revision since the regulation was adopted in 2005, with updated assurances due by 30 April 2026.11 In the UK, the refreshed Concordat to Support Research Integrity was published on 4 April 2025, and organisations were expected to align with it by April 2026.12 Between them these documents require what a headline cannot supply: a defined standard of proof, a definition of what counts as intentional, knowing or reckless conduct, confidential routes to raise concerns, trained investigators, fixed timelines, and a right of appeal.1112

They also encode something that gets lost in public argument: honest error and honest differences of interpretation are not misconduct. A wrong paper is not a fraudulent paper. A messy paper is not a fraudulent paper. A paper with an unattributed sentence in the methods is a serious problem that is nonetheless a different problem from fabricated data, and treating the two as one category serves nobody except people who would like the whole conversation abandoned as hysteria.

I have views about the cases in the news. I am not going to publish them, because I have not seen the evidence, and neither has anybody forming a confident opinion from a news article. That is not fence-sitting. It is the same standard I would want applied to me, and I would rather be dull about it than be one more person adding certainty to a matter that has not yet been established. If an institution reaches a finding, that finding will be public and arguable. Until then, the honest position is that I do not know.

The count was never supposed to be the point

The premise underneath all of this is that something changed — that people once spent months on one problem and published a paper or two a year, and now the numbers are absurd. I have a great deal of sympathy for that account and I want to be careful with it, because nostalgia is not evidence and I do not have a clean historical median to hand.

What I do have is the direction. Articles indexed in Scopus and Web of Science grew roughly 47% between 2016 and 2022, far outpacing any growth in the number of practising scientists — which means the writing, reviewing and editing load per researcher rose sharply over six years.10 And extreme publishing behaviour is not a static curiosity: Ioannidis and colleagues recorded 1,226 authors outside physics exceeding 60 articles in 2022 alone.2 Something is being produced faster than it is being read.

The consoling thing is that formal research assessment has largely stopped rewarding personal volume, and almost nobody has noticed. REF 2029 — whose contribution to knowledge and understanding element carries 55% of the weight — requires submissions of 2.5 outputs per full-time-equivalent staff member at unit level, with no minimum output requirement for any individual and an explicit expectation that no more than five outputs be associated with a single person.13 Read that as a system design and it says the opposite of what people believe it says: your five best pieces of work are all the exercise will look at. The hundredth paper is worth precisely nothing to it.

Three things I hold, then.

• Integrity is a set of habits, not a personality trait. Nobody avoids plagiarism by being a good person. They avoid it by never pasting source text into a draft, by writing the authorship agreement before the results exist, by keeping records they would be content to hand over. Build the habits when the stakes are low, because you will not invent them when the stakes are high.
• The incentive is the villain, and it is local. The national assessment framework already caps how much any one person’s volume can count. Hiring panels, promotion committees and probation reviews mostly have not caught up. That is fixable by the people reading this, in the meetings they sit in, this year.
• Suspicion is not analysis. The right response to an improbable publication list is a specific question about a specific paper, put through a process with rules. It is not a thread. I say this as someone who finds some of these lists genuinely hard to explain — and who has been wrong before about what he assumed he was looking at.

I still think the best work I have done took months of being stuck, and produced very little that a counter could register. That is not a moral claim about other people’s output. It is just what the work felt like when it was going well, and I would not know how to want a career that never felt like that.