Flood and Dykes

1Are there too many papers?#

In July 1945 the American engineer Vannevar Bush published “As We May Think” in The Atlantic Monthly. What he complained of was not too little knowledge but too much: there was “a growing mountain of research”, and researchers could no longer keep up with the findings and conclusions of thousands of colleagues. He gave the example of Mendel, whose laws of genetics were lost to the world for a generation because his paper never reached the few people capable of understanding and extending it. And our methods for finding a way through the maze of the literature, Bush observed, had hardly changed since the days of square-rigged ships. So he imagined a machine called the Memex, which would link material together by associative “trails”, as the human mind does – a machine that later generations came to see as the distant ancestor of hypertext.¹

Eighty years later, the mountain has become a flood. According to the US National Science Board, about 3.3 million science and engineering papers were published worldwide in 2022, 59% more than a decade earlier.² The open scholarly index OpenAlex now lists more than 477 million scholarly works.³ A single publisher, Elsevier, received more than 4.2 million submissions in 2025 and published 795,000 articles.⁴ A mountain at least stays put; today’s literature is more like a flood that keeps on rising. Alongside it rise the dykes, built ever higher: journal selection, peer review, impact factors and university rankings, plagiarism checks and retractions, and the newest and least reliable of all, AI detectors.

Let me state my argument up front. Digitization merely swapped paper for digital paper. Open access and preprints changed circulation but not the unit called “the paper”. Metrics, the replication crisis, the reviewing bottleneck, paper mills and generative AI have piled onto that unit, layer upon layer. And the many repairs, useful as most of them are, all reinforce the dykes along the same channel. Ancient Chinese books offer a ready-made story. The Records of the Grand Historian (Shiji) tells how Emperor Yao put Gun in charge of the flood, and “for nine years the waters did not subside, and his labours came to nothing”; the Discourses of the States (Guoyu) says that “Gun dammed the flood”, which the commentator Wei Zhao glosses as “walled off the hundred rivers” – that is, he built embankments to hold the water back.⁵⁶ Yu, who took over after him, changed the method. The moral of the story is not that dykes are useless, but that when the volume of water changes, the thinking about how to manage it must change too. We will come back to this at the end. First, let us see how the water rose.

2In the age of the web, why is a paper still paper?#

Digital networks were meant for scientists from the start. In 1989 Tim Berners-Lee proposed the World Wide Web at CERN, the European Organization for Nuclear Research; in CERN’s own words, it was originally conceived “to meet the demand for automated information-sharing between scientists in universities and institutes around the world”. On 30 April 1993 CERN put the Web software in the public domain.⁷ On 15 June that same year Adobe released Acrobat 1.0 and the PDF format. When the Adobe co-founder John Warnock set the project going, the problem he wanted to solve was that documents became garbled as they passed between different machines and operating systems: PDF was to make a document look exactly the same wherever it went.⁸ On one side stood hypertext, capable of endless linking; on the other, “digital paper”, a faithful replica of the printed page. Academic publishing chose the second.

The choice made sense. PDF keeps page numbers; page numbers hold citations in place, and citations hold up the whole system of academic credit. A PDF can be printed and archived, and it makes it easy to establish which copy is the official “version of record”. From the mid-1990s publishers moved their journals online one after another, and Vincent Larivière and his colleagues found that the big publishers’ share of the literature expanded rapidly precisely “since the advent of the digital era” – a point we will return to below.⁹ Journals became websites, yet the most prominent button on those websites is often still “Download PDF”. Articles are still assigned to a volume and an issue and still carry first and last page numbers, even though those pages have never been printed.

Deeper still, the genre did not change. Introduction, methods, results, discussion: this structure, known as IMRaD, actually settled into shape rather late. A sampling study of four leading medical journals found that IMRaD began to be adopted only in the 1940s, reached 80% in the 1970s, and became the sole format for original papers only in the 1980s.¹⁰ In other words, what the digital age inherited was a genre that had only just set. Its length, its structure and its static text, with “publication” as the finishing line, were all carried onto the web unchanged. If an error turns up after publication, all one can do is issue a separate correction or retraction notice; the original text itself never grows.

This is what the manifesto means when it says that the information age merely swapped paper for digital paper. Hypertext was invented expressly for scientists, yet it failed to change their most important form of writing; the reasons are discussed in more detail in the section on why hypertext did not change the paper. Here we need only remember this: digitization made papers travel faster and easier to find, but it left their body untouched. A paper is still a closed object. Data, code and failed attempts are either squeezed into appendices or left on the author’s hard drive, and the reader still has only one way in: to read from beginning to end. All the changes of the next thirty years took place around this unchanged body.

3After open access, who pays the bill?#

The first thing that truly changed circulation was the preprint. In 1991 Paul Ginsparg, a physicist at Los Alamos National Laboratory, set up the preprint server that later became arXiv. It does no peer review; it checks only whether a submission belongs to the relevant field and is scholarly in nature.¹¹ In 1992 arXiv received 3,263 new submissions over the whole year; in 2025 it received 284,000.¹² It took twenty-three and a half years to accumulate its first million papers and another seven to reach two million; the third million took just four years, arriving in April 2026. On 1 July that year arXiv left Cornell University to become an independent non-profit organization.¹³ bioRxiv, for the life sciences, went live in 2013; medRxiv, for medicine, in June 2019.¹⁴ The COVID-19 pandemic pushed preprints into the spotlight: of the more than 125,000 pandemic-related articles that appeared in its first ten months, over 30,000 were posted on preprint platforms.¹⁵

Open access, for its part, set out to tear down the paywall in front of readers. The Budapest Open Access Initiative, made public on 14 February 2002, declared that “an old tradition and a new technology have converged to make possible an unprecedented public good”. It defined open access as “free availability on the public internet, permitting any users to read, download, copy, distribute, print, search, or link to the full texts” of articles.¹⁶ Before that, an open letter launched in 2000 by Harold Varmus and others had already gathered the signatures of more than 30,000 scientists; in October 2003 the Public Library of Science (PLOS) published the first issue of PLoS Biology, with authors paying a $1,500 publication fee and readers paying nothing.¹⁷ PLOS ONE, launched in December 2006, went further: reviewers would judge only whether a study was technically sound, leaving the question of whether it mattered to the community after publication.¹⁸ Within a few years it had become the world’s largest peer-reviewed journal by number of articles.¹⁹

Some people simply went round the wall. In 2011 the neuroscience researcher Alexandra Elbakyan founded Sci-Hub. Server logs obtained by Science showed that between September 2015 and February 2016 it received some 28 million download requests; the largest numbers came from China (about 4.4 million), India (3.4 million) and Iran (over 2.6 million), followed by Russia and the United States.²⁰ Policy followed. On 4 September 2018 cOAlition S, a group of research funders from several European countries, launched Plan S, which required the research they funded to be made openly available immediately from 2021.²¹ In August 2022 the White House Office of Science and Technology Policy told federal agencies to ensure that, by the end of 2025, papers from research they fund would be freely available as soon as they were published; the National Institutes of Health (NIH) began applying the rule on 1 July 2025.²²

China did not start late. On 15 May 2014 the Chinese Academy of Sciences and the National Natural Science Foundation of China each issued policy statements requiring the authors of funded papers to deposit the final peer-reviewed manuscript in a repository on publication and to make it openly accessible within 12 months.²³ After Plan S was announced, 15 libraries in mainland China, including the National Science and Technology Library and the National Science Library of the Chinese Academy of Sciences, met in March 2019 to respond. They expressed “broad support” for its goal of immediate open access and endorsed the final statement of the 14th Berlin Open Access Conference, while cautioning that any measures on article processing charges should take the world’s diversity into account, so that they remain affordable across countries and disciplines.²⁴

That caution went to the heart of the matter. Open access removed the readers’ paywall, only to put up a new wall in front of authors: the article processing charge, or APC. By the estimate of Stefanie Haustein and her colleagues, APCs paid to six large publishers rose from $910 million in 2019 to $2.54 billion in 2023.²⁵ As of September 2026, publishing a gold open-access paper in Nature costs £9,390 or $12,850.²⁶ An analysis of 1.5 million papers showed that APC levels are clearly stratified by discipline and by country; even after controlling for both, authors at better-resourced institutions tend to pay higher APCs, though the effect is only small to moderate and is strongest in countries with lower GDP.²⁷ According to China Daily, citing figures from the National Science Library, Chinese scholars published 313,500 fee-paying open-access papers in 2024 and paid $909 million for them, an increase of more than a fifth on the year before.²⁸ And according to Science, from March 2026 the Chinese Academy of Sciences stopped using its own funds to pay for about 30 international open-access journals that charge more than $5,000, Nature Communications and Science Advances among them.²⁹

Others have tried to change the relationship between review and publication. From 31 January 2023 eLife did away with the accept–reject decision after peer review: every manuscript sent out for review is published as a “Reviewed Preprint”, together with the public reviews and an assessment by the editors, and the authors decide whether to revise and when to finalize.³⁰ There was a price. Clarivate ruled that the model was not eligible for an impact factor and indexed eLife only partially; as a result, eLife lost its impact factor.³¹ Looking back, the open movement has achieved a great deal: according to the International Association of Scientific, Technical and Medical Publishers (STM), gold open access rose from 17% to 44% of the world’s articles, reviews and conference papers between 2015 and 2025.³² But all these changes concern circulation: who can read, when they can read, and who pays. A preprint is a paper that has not yet been reviewed; an open-access paper is a paper that does not charge its readers; a reviewed preprint is a paper with its reviews attached. The unit is always the paper. And these changes have happened mainly in the sciences. When I was a student, I heard one of my philosophy-of-science teachers say that print journals ought long since to have given way to free and open distribution on the web. Twenty years on, almost all of us search and read online, yet the public forum in which humanities scholars exchange ideas is still the print journal, while the online forums of those days have vanished. The sciences, thanks to preprints and GitHub, are somewhat better off; in adapting to the internet age, the humanities have positively gone backwards.³³

4Where does the surge come from?#

Start with scale. The scientific literature has long been growing exponentially: one estimate combining four databases puts the annual growth of modern science’s publication output at about 4.1%, which means it doubles roughly every 17 years.³⁴ Over the past decade it has grown faster still. The National Science Board reports that worldwide science and engineering output grew by 59% between 2012 and 2022; China’s output grew by 173%, and in 2022 China accounted for 27% of the world total and the United States for 14%.² Another study counted about 47% more articles indexed in Scopus and Web of Science in 2022 than in 2016. The number of researchers has lagged far behind, so the burden of writing, reviewing and editing has grown heavier; some publishers have vastly expanded their output through “special issues” with lightning-fast turnaround.³⁵

As the literature grew, the publishing industry rapidly concentrated. Larivière, Haustein and Philippe Mongeon analysed 45 million documents published between 1973 and 2013. In the natural sciences and medicine, the top five publishers’ share rose from a little over 20% in 1973 to half in 2006 and 53% in 2013; in the social sciences it was higher still, reaching 70%, and only the humanities remained relatively independent, at about 20%. This concentration accelerated precisely after the digital era arrived in the mid-1990s.⁹ Digital networks could have decentralized publishing; instead they made the big publishers’ advantages of scale all the more pronounced. The more journals a publisher owns and the bigger their names, the more bargaining power it has when it sells them as a bundle.

Concentration brings profit. The 2025 annual report of RELX, Elsevier’s parent company, shows its Scientific, Technical & Medical segment earning £2,714 million in revenue and £1,035 million in adjusted operating profit – which works out at a margin of about 38% (besides journals, the segment includes databases and analytics tools) – with 80% of revenue coming from subscriptions. In the same year Elsevier’s more than 3,000 journals received over 4.2 million submissions, vetted by 37,000 editors and 1.9 million reviewers, and published 795,000 articles; the report makes a point of noting that article submissions were “growing very strongly”.⁴ Authors are not paid for their submissions and reviewers are not paid for their reviews, yet the libraries that serve readers must pay for subscriptions. This is the most frequently criticized structure in academic publishing.

Resistance began with the subscription crisis. In January 2012 the mathematician Timothy Gowers announced on his blog that he would no longer submit to, referee for or edit Elsevier journals. His reasons were prices far higher than other publishers’, the bundling of journals into packages, and the company’s support for legislation that would obstruct open access; he also proposed setting up a website where colleagues could declare their position publicly.³⁶ On 28 February 2019 the University of California, which produces nearly a tenth of all US research papers, announced that it would not renew its contract with Elsevier: it wanted its research to be freely available to the world immediately, and it wanted to rein in ever-rising costs.³⁷ Only after a two-year stand-off did the two sides sign a new four-year agreement, in March 2021.³⁸

The subscription model and the APC model look like opposites, but they have one thing in common: both run paper by paper. In the subscription era, journals used article counts to prop up the price of their bundles; in the APC era, every extra paper brings in another fee. In 2023 MDPI, the biggest collector of APCs, took in about $680 million on its own, while Elsevier and Springer Nature each took in more than $500 million.²⁵ Publishers have a motive to expand, authors have a motive to publish more, and the evaluation system rewards authors by the number of their papers. So the flood and the people building the dykes are tied to the same unit of counting. That unit itself is where the story of the next section begins.

5How did the measure become the target?#

What made papers countable was the citation index. In 1955 Eugene Garfield published “Citation Indexes for Science” in Science, proposing to link related papers by following their references, so that researchers could trace an idea back to its sources and forward through its development.³⁹ In 1960 he founded the Institute for Scientific Information; in 1964 he launched the Science Citation Index (SCI); he went on to develop the impact factor as a way of comparing journals of different sizes, and in 1975 began publishing Journal Citation Reports.⁴⁰ These tools were first meant for searching the literature and comparing journals; for this prehistory, see big science, citations and metrics. But a number that can rank things will sooner or later be used to rank people.

In 2005 the physicist Jorge Hirsch proposed the h-index: a person’s h-index is the largest number h such that h of their papers have each been cited at least h times.⁴¹ Two years earlier, in June 2003, Shanghai Jiao Tong University had published the first Academic Ranking of World Universities; among its six indicators today, papers in Nature and Science, papers indexed in the SCIE and SSCI, and highly cited researchers each count for 20%.⁴² American academics had long talked of “publish or perish”. In 1975 the economist Charles Goodhart wrote that “any observed statistical regularity will tend to collapse once pressure is placed upon it for control purposes”.⁴³ The line more widely quoted today – “When a measure becomes a target, it ceases to be a good measure” – actually comes from the anthropologist Marilyn Strathern’s 1997 critique of audit in the British university system; she notes that the accounting scholar Keith Hoskin had already called the idea “Goodhart’s law”.⁴⁴

In China the law has come true with particular directness. One study examined 168 policies for rewarding papers at 100 Chinese universities between 1999 and 2016, and found that in 2016 the average bonus for a paper in Nature or Science came to $43,783, with the highest reaching $165,000.⁴⁵ A 2020 document from the Ministry of Education and the Ministry of Science and Technology frankly admitted that SCI-related metrics had become the core indicators for academic evaluation, promotion and performance review, causing “problems such as distorted values, a boastful and impetuous academic culture, and a rush for quick success and instant gain” in university research.⁴⁶ We will see the consequences of such incentives again when we come to paper mills.

From 2018 the remedies came thick and fast. In July that year the General Office of the Communist Party’s Central Committee and the General Office of the State Council issued opinions calling for efforts to “overcome the tendency to rely solely on papers, titles, diplomas and awards, and introduce evaluation based on representative works”.⁴⁷ In October, five ministries, among them the Ministries of Science and Technology and of Education, launched a special campaign to clear away the “Four Onlys”,⁴⁸ and the Ministry of Education soon added “hats” – honorific talent titles – to the list: hence the “Five Onlys” reform.⁴⁹ In February 2020 the Ministry of Science and Technology introduced “three categories of high-quality papers”, required that in principle at least a third of the representative works submitted be papers in Chinese science and technology journals, and banned the use of national science and technology programme funds to reward publication.⁵⁰ The Ministry of Education and the Ministry of Science and Technology told universities to abolish rewards based directly on SCI metrics and neither to publish nor to rely on rankings built around the SCI.⁴⁶ That October, the Overall Plan for Deepening the Reform of Educational Evaluation in the New Era summed up the goal as “resolutely overcoming the stubborn ills of relying solely on test scores, admission rates, diplomas, papers and hats”.⁵¹

Soul-searching abroad came at almost the same time. The San Francisco Declaration on Research Assessment (DORA) of 2012 argued that journal-based metrics such as the impact factor should not be used as a substitute for judging the quality of individual articles, nor to assess individual researchers or to make decisions about hiring, promotion and funding.⁵² The very first principle of the 2015 Leiden Manifesto for Research Metrics is that quantitative evaluation should support expert judgement, not replace it.⁵³ The Agreement on Reforming Research Assessment (CoARA), reached in 2022, has been signed by more than 800 organizations, which commit themselves to abandoning the inappropriate use of the impact factor and the h-index as proxies for quality, and to keeping institutional rankings out of research assessment.⁵⁴ These reforms point in the right direction, but what they replace is the way of counting, not the thing being counted: “representative works” are still papers, and “three categories of high-quality papers” are still papers. A metric’s proper role is only to aid judgement. I have said the same about plagiarism checks. Plagiarism has only two states, copied or not copied; there is simply no such thing as a few per cent of it. Similarity-checking software can at most help to expose it; it cannot serve as a hard pass mark. If a university cannot even manage to have its teachers read every graduation thesis from beginning to end, with the software’s help, what right does it have to admit so many students? Substituting metrics for judgement is, in the end, institutional laziness.³³⁵⁵ The paper has been too successful: it has become almost the only currency of an academic career, and the joy that research ought to hold has been squeezed out.

6Where does failed research go?#

In 2005 the epidemiologist John Ioannidis published a paper with a startling title in PLoS Medicine: “Why Most Published Research Findings Are False”. His argument was not mysterious. The smaller the studies, the smaller the effects, the more numerous and less carefully chosen the relationships tested, the more flexible the designs and analyses, the heavier the interests and prejudices involved, and the more teams chasing the same statistical significance, the lower the probability that a finding is true; in many circumstances a research finding is more likely to be false than true.⁵⁶ The paper did not say that science cannot be trusted. What it pointed out was that the institutional arrangements of a field can systematically raise the proportion of false positives.

A decade later the empirical data arrived. In 2015 the Open Science Collaboration reported in Science that it had repeated 100 studies from three psychology journals: 97% of the original studies had significant results, but only 36% of the replications did, and effect sizes were on average only half as large.⁵⁷ Biomedicine was no different. In 2012 researchers at the company Amgen reported that they had tried to confirm the findings of 53 “landmark” cancer studies and had succeeded with only six.⁵⁸ In 2021 the Reproducibility Project: Cancer Biology repeated 50 experiments from 23 high-impact papers. For positive effects, the median effect size in the replications was 85% smaller than in the originals, and across several criteria the success rate was about 46%.⁵⁹ In 2016 Nature surveyed more than 1,500 researchers: over 70% had at some point failed to reproduce someone else’s experiment, and more than half had failed to reproduce their own.⁶⁰

The other side of the problem is publication bias. In 1979 the psychologist Robert Rosenthal gave it a vivid name, the “file drawer problem”: non-significant results get locked away in a drawer and never see the light of day.⁶¹ A 2014 study that followed 221 social-science experiments found that studies with significant results were 40 percentage points more likely to be published than those with null results, and 60 percentage points more likely to be written up at all. The bias lay mainly not with editors but with authors, who simply never wrote up their null results.⁶² Daniele Fanelli’s analysis of 2,434 papers showed that the share reporting “positive” results ranged from 70.2% in the space sciences to 91.5% in psychology and psychiatry;⁶³ another of his studies, covering more than 4,600 papers from 1990 to 2007, found that the proportion of positive results rose by more than a fifth over that period.⁶⁴

These numbers point not just to the carelessness of individual researchers but to the structure of the paper as a genre. A paper has to tell a story that worked out; null results, dead ends and abandoned plans can rarely become “a paper” in their own right. The manifesto calls this the cost of publication: the world of the literature is like a city built by survivors. Repairs have arrived here too. In 2013 the neuroscience journal Cortex introduced “Registered Reports”: the research question and methods are reviewed first and, if they pass, the paper is accepted in principle, whatever the results turn out to be.⁶⁵ Today more than 300 journals offer the format.⁶⁶ A 2019 report from the US National Academies likewise called on researchers, institutions, journals and funders to work together to improve transparency.⁶⁷ But Registered Reports change the timing of the judgement; what they produce is still a paper, and the vast majority of failures still have nowhere to go. Nor do failures necessarily belong in more papers. As I said in an interview, the answer to publication bias is not necessarily to publish more papers about failures, but to build open databases of experiments, so that even when a failed experiment is never written up, its data enter the public record; the idea that “successes and failures alike should all get published as papers” does not make sense. More fundamentally, our idea of success has to become more plural: scientists should be able to define “success” more flexibly according to their own circumstances, backed by third-party oversight and diverse sources of funding, instead of being left with a single rigid, top-down yardstick.⁶⁸

7Who reviews, and can they keep up?#

If journals are the dykes, peer review is the night watch on top of them. The institution is younger than many people imagine: Nature made external review a requirement for publication only in 1973, and before that its editors often published papers directly on the strength of their trust in the authors.⁶⁹ For this history, see the recent origins of peer review. Today peer review is academic publishing’s main quality gate, and also its largest body of unpaid labour. Balazs Aczel and his colleagues estimated that in 2020 the world’s reviewers spent more than 100 million hours reviewing – over 15,000 years. Valued at their wages, the work of reviewers was worth more than $1.5 billion in the United States, more than $600 million in China and nearly $400 million in the United Kingdom.⁷⁰

The night watch is getting harder to staff. Publons’ Global State of Peer Review report for 2018 showed that in 2013 editors had to send out 1.9 invitations on average to secure one review, and by 2017 they needed 2.4. Review invitations were growing by 9.8% a year, while accepted papers grew by only 4.9%, and three-quarters of editors said that finding reviewers and getting them to say yes was the hardest part of the job. The report also found that China carried out only 8.8% of the world’s reviews, well below its 13.8% share of paper output; in 2017 China completed just 0.8 reviews for every submission, against an average of 2.3 for the established research nations.⁷¹ Artificial-intelligence conferences have pushed the pressure to the limit. NeurIPS 2025 received 21,575 valid submissions and enlisted 20,518 reviewers;⁷² ICLR 2026 received nearly 20,000 submissions, and according to Nature reviewers were assigned five papers each on average, to be reviewed in about two weeks.⁷³ The finer the specialization, the fewer the “peers” able to review any given paper – an echo, at the publishing stage, of the triumph and the price of specialization.

More fundamentally, peer review was never designed to catch errors and fraud. In 1998 a general medical journal deliberately introduced eight weaknesses into a paper it had already accepted and sent it to 420 reviewers; the 221 reports that came back pointed out, on average, only two of them.⁷⁴ In 2008 another trial, at the BMJ, planted nine major errors in a paper, and the 607 reviewers found on average only 2.58 of them the first time round.⁷⁵ Reviewers cannot see the raw data, and they have no time to redo the experiments. And when the review itself is faked, it cannot even protect itself: in 2017 Springer retracted 107 papers from Tumor Biology in one go because their peer reviews had been fabricated, many of them under the names of real scientists, paired with fake e-mail addresses.⁷⁶

Open peer review is one response. Every research paper submitted to Nature from 16 June 2025 is published with the reviewers’ reports and the authors’ responses, and reviewers can decide for themselves whether to be named; discussions between editors and confidential comments, however, stay private.⁷⁷ eLife publishes every review, and ICLR’s reviews are posted openly on the OpenReview platform.⁷⁸ Transparency makes the people on the dykes visible, and it allows the quality of reviewing to be checked and studied. But it does not reduce the volume of water. Review remains a one-off checkpoint for a single paper, and the queue in front of it only grows longer.

8Why can papers be mass-produced?#

As long as papers can be exchanged for money, promotions and degrees, someone will mass-produce them. A 2022 report by the Committee on Publication Ethics (COPE) and STM defined paper mills as outfits that exploit the pressure to publish in academia, producing fraudulent manuscripts so that researchers can publish easily or buy authorship.⁷⁹ A study published in PNAS in 2025 found that paper mills, brokers and predatory journals have joined up into networks that are adept at evading interventions, and that fraudulent publications are growing far faster than legitimate science.⁸⁰

In 2023 this undercurrent swept away a stretch of the dyke. That year more than 10,000 research papers were retracted worldwide, a record, and over 8,000 of them came from a single publisher, Hindawi.⁸¹ Founded in Cairo and headquartered in London, with more than 200 journals, this open-access publisher had been bought by Wiley for $298 million in January 2021.⁸² After that, it gradually came to light that paper mills had infiltrated its special issues on a massive scale. In December 2023 Wiley announced that it would stop using the Hindawi name. By then Web of Science had already delisted 19 Hindawi journals, and Wiley itself had closed four journals “heavily compromised by paper mills”; as a result, Wiley’s revenue fell by $18 million in a single quarter.⁸³

Exposing paper mills often falls to volunteer “science sleuths”. In 2021 Guillaume Cabanac, Cyril Labbé and Alexander Magazinov noticed a strange style of writing. Someone had run existing text through an online paraphrasing tool to “launder” it, and technical terms had been swapped for plausible-looking synonyms that were wrong. So “artificial intelligence” became “counterfeit consciousness”, “deep neural network” became “profound neural organization”, and “big data” became “colossal information”. They called these “tortured phrases”, and found them appearing in batches in an Elsevier computing journal.⁸⁴ The retraction database that Retraction Watch had built up over many years was acquired by Crossref in September 2023 and made freely available; at the time, the two organizations’ records together came to about 50,000.⁸⁵ As of September 2026 the database lists more than 67,000 retractions.⁸⁶

China is one of the worst-hit countries in this wave of retractions. A 2023 analysis in Nature found that among the major research nations, Saudi Arabia, Pakistan, Russia and China had the highest retraction rates over the previous two decades.⁸¹ A team at the National Science Library of the Chinese Academy of Sciences, citing Retraction Watch data, reports that China’s share of the world’s retractions for misconduct rose from 14.4% in 2011 to 53.2% in 2020.⁸⁷ Enforcement was stepped up in turn. In May 2018 the General Offices of the Party Central Committee and the State Council issued Several Opinions on Further Strengthening the Building of Research Integrity, stating plainly that researchers must “not buy, ghostwrite or submit papers by proxy, or fabricate peer-review experts and their comments”, insisting on “zero tolerance”, and making those guilty of serious breaches of research integrity accountable for life.⁸⁸ In 2022, 22 government departments led by the Ministry of Science and Technology issued Rules for Investigating and Handling Research Misconduct, which specifically define “proxy submission”: having a third-party agency submit a paper and respond to the reviewers on one’s behalf.⁸⁹ And since 2020 the National Science Library of the Chinese Academy of Sciences has published an Early Warning Journal List of international journals, which weighs indicators such as publication volume, APCs, self-citation rates and retractions, to warn researchers to choose where they publish with care.⁸⁷

Rising retraction numbers can be read in two ways: they show that fraud is serious, and they show that detection and correction have become more active. Both readings are right, and both deserve to be taken seriously; retractions are themselves evidence that the dykes are still working. But either way they reveal the same structure. What paper mills produce is exactly the thing the metrics system counts. A mill does not need to discover anything; it only needs to manufacture an object formatted like a paper. As long as “a paper” can still be cashed in for bonuses, promotions and degrees, someone will manufacture “a paper”; however high the dykes are built, water will seep through at the weakest point. Nor is there any end to the arms race of detection. I have compared plagiarism checks to the melamine in China’s milk powder: once protein testing became a checkpoint that anyone could run for themselves, what dairy farmers mixed in was no longer water but poison. Once plagiarism checks became universal, “getting past the checker” became an essential skill and text-laundering services sprang up to meet the demand; today, the web pages that check your “AI rate” already carry advertisements offering to “cut your AI rate in one click”. What really works has never been sharper detection; it is sound oversight, and penalties that bankrupt the fraudster. Retractions and apologies that barely sting are precisely the soil in which fraud flourishes.⁹⁰³³

9Can AI-written papers be detected? Should they be?#

On 30 November 2022 OpenAI released ChatGPT.⁹¹ Within a year its fingerprints were turning up in formally published papers. A paper published on 9 August 2023 in the physics journal Physica Scripta bore on its third page the words “Regenerate response” – the label on a button in the ChatGPT interface. The person who spotted it was, once again, Cabanac.⁹² The publisher reportedly went on to retract the paper because the authors had not disclosed their use of AI, in breach of publishing ethics.⁹³ Similar fingerprints include “As an AI language model”, “Certainly, here is” and “As of my last knowledge update”. The Polish scholar Artur Strzelecki found 89 papers containing phrases of this kind in journals in the top two quartiles of Scopus, many of which had already been cited by others.⁹⁴

The most absurd episode came in February 2024. A paper published on the 13th in Frontiers in Cell and Developmental Biology included a figure of a rat with preposterously large genitals, labelled with strings of meaningless pseudo-words. Three days later the journal retracted it, on the grounds that its “AI-generated figures” did not meet its standards of editorial and scientific rigour, and admitted that one reviewer had raised concerns about the figures, to which the authors never responded.⁹⁵ Beyond the jokes lies use that has become routine. Dmitry Kobak and his colleagues analysed shifts in the vocabulary of more than 15 million PubMed abstracts and estimated that at least 13.5% of 2024 abstracts had been processed with large language models, rising to 40% in some subgroups.⁹⁶ On the reviewing side, the detection company Pangram ran its own detection model over the 75,800 public reviews of ICLR 2026 and estimated that 21% had been generated entirely by AI and more than half showed traces of AI involvement, whereas only about 1% of the submitted papers had been written entirely by AI. All of these are a detection model’s judgements, not official conference statistics.⁷³⁷⁸ As for how reliable detection models are: I have tried several tools that claim to detect AI-generated content, and they often judge my own writing to be 100 per cent AI; meanwhile, text from a model you have taken the time to break in differs little in style from your own writing, and is practically impossible to detect.³³

Conferences have started setting traps. In the PDF of every submission, ICML 2026 hid instructions that only a large language model would read, telling it to work two randomly chosen phrases into its review. The aim was to catch reviewers who had broken their pledge and handed a whole manuscript to a language model. In the end 795 reviews were confirmed, about 1% of all reviews; and under the rule that submitting authors must also take on reviewing duties, the 497 papers that the offenders had submitted as authors were rejected outright – about 2% of all submissions.⁹⁷ ICLR 2026, for its part, required reviewers to declare their use of large language models; those who did not declare it truthfully risked having their own submissions rejected.⁹⁸ Both measures catch breaches of contract, not bad reviews. The hidden phrases prove only that someone broke a promise, and the declaration is simply the platform’s contract; once rules are made they should be kept, and there is nothing wrong with that. But neither can answer the real question: were those 795 reviews actually worse than the rest, or better? Scoring texts by their “AI rate” and acting on the score is both inaccurate and beside the point. It is institutional laziness, putting detection in the place of judgement.⁵⁵

The real pressure comes from volume. Submissions to NeurIPS grew from 9,467 in 2020 to 21,575 in 2025.⁷² From 31 October 2025 arXiv’s computer-science section stopped accepting review articles and position papers that had not been through peer review. Its reasons: hundreds of review articles were pouring in every month, large language models had made such content easy to churn out on demand, and most of what arrived was “little more than annotated bibliographies”.⁹⁹ In February 2026 Nature reported that Raphael Wimmer of the University of Regensburg in Germany had used Prism, a writing tool newly launched by OpenAI, to produce a paper about a fictional experiment in fifty-four seconds. “Writing a paper has never been easier,” he wrote on social media. “Clogging the scientific publishing pipeline has never been easier.”¹⁰⁰ Computer science has taken the brunt; for how the discipline has responded, see the section on computer science.

AI has also begun to take the stage as author and reviewer. In March 2025 Sakana AI announced that a paper generated entirely by its “AI Scientist” system had passed peer review at a workshop at ICLR 2025, with scores of 6, 7 and 6 from its three reviewers; as agreed in advance, it was withdrawn before publication.¹⁰¹ That October, researchers at Stanford University held Agents4Science, the first conference at which AI served as first author and as reviewer. Of 315 submissions, 48 were accepted. The correlation between the scores of three different large-language-model reviewers was only 0.48; more than half of the submissions had at least one reference flagged as problematic; and research driven entirely by AI made up 23.3% of submissions but only 14.9% of accepted papers.¹⁰²

The rules are catching up too. The Guidelines for Responsible Research Conduct (2023), issued in December 2023 by the Department of Supervision of the Ministry of Science and Technology, stipulate that generative AI must not be used to produce application materials directly and must not be listed as a co-contributor to research results, that unverified references it generates must not be used directly, and that using it in peer review requires the organizers’ prior consent.¹⁰³ For the practices of journals and funders in other countries, see the rules of journals, funders and learned societies. But most of these rules are answering “Who wrote it?”, when the question to ask is “Is it any good?” What generative AI has broken is not the value of the paper but a stand-in that had long been taken for granted. Writing a decent paper used to take real effort, so people treated “being able to write it” as a guarantee of effort and competence, and reviewers were glad of the shortcut. Now the cost of writing is approaching zero and AI has raised the average to fluent, well-reasoned prose. The guarantee is gone, and what stands exposed is a shortage of judgement. The problem is not AI; the existing system of evaluation was never reliable to begin with. And if AI makes waffling papers obsolete, that is no bad thing. This is exactly what the manifesto means when it says that the old system was designed on the assumption that writing is hard.^3355104

10Is the open-science toolbox enough?#

Faced with the flood, open science has produced a whole toolbox. The first set of tools is identifiers and links. Crossref, founded in January 2000, registers DOIs for the scholarly literature, giving every paper and every reference a stable address;¹⁰⁵ ORCID, launched in 2012, gives each researcher a unique number, ending the confusion between people with the same name;¹⁰⁶ CERN launched Zenodo in May 2013 so that data, code and research outputs of every kind could be archived and cited;¹⁰⁷ and the FAIR principles, put forward in 2016, require data to be findable, accessible, interoperable and reusable – and friendly to machines as well.¹⁰⁸ With these in place, research outputs other than papers had, for the first time, “ID cards” of their own.

The second set remakes the process of review and publication. Registered Reports move the judgement to before the results are in; open peer review brings reviews into the daylight; and “overlay journals” pare the journal down to nothing but its certifying function. Discrete Analysis, which Gowers announced in 2015, is just such a journal. It is itself no more than a list of links to arXiv preprints; it pays a review-management platform about $10 for each submission it handles, and everything else rests on mathematicians’ voluntary labour.¹⁰⁹ It was officially launched in March 2016.¹¹⁰ Medicine has developed “living systematic reviews”, which are continually updated as new evidence comes in, instead of being written once and then frozen.¹¹¹

The third set is infrastructure and error correction. OpenAlex, launched in January 2022, took over from the Microsoft Academic Graph with open data.³ In December 2023 Sorbonne University announced that from 2024 it would stop subscribing to Web of Science and Clarivate’s bibliometric tools and turn to open, free tools such as OpenAlex.¹¹² Since June 2019 the reference manager Zotero has drawn on Retraction Watch’s data to alert users automatically to items that have been retracted.¹¹³ The fourth set is the evaluation reform discussed above: DORA, the Leiden Manifesto for Research Metrics, CoARA, and China’s system of representative works. According to China Daily, Tsinghua University and other universities now let faculty put forward just a handful of outputs – five at most – that best represent their scholarly standing.²⁸

These tools are all real advances. They make papers easier to find, to link and to check; they give data and code identities of their own; they have brought part of the reviewing process into the light; and they have done something to curb the tyranny of counting. Many of the components of what the manifesto calls new standards and reputation come straight out of this toolbox: evidence you can click open, versions you can trace, reviews held in public. Yet line these tools up side by side and a common shape appears. They have all grown up around the same thing, like rings of stone reinforcing the banks of the old channel.

11Raise the dykes, or dig new channels?#

Almost every patch has been applied around the unit called “the paper”. A DOI is the address of a paper; ORCID links a person to their papers; a Registered Report is a paper accepted in advance; an overlay journal is a list of papers; open peer review is the public review of a paper; a retraction notice is a label stuck on a paper. DORA says: do not judge papers by their journals – but what is judged is still the paper. The representative-works system says: do not count papers – but the representative works are still papers. Even the rules on generative AI mainly ask: who wrote this paper?

The trouble lies precisely in this unit. A single paper carries too many functions at once: it is a vehicle of communication, an object of certification, an archival record, a unit for counting credit, and a voucher to be exchanged for bonuses and promotions. Every pressure bears down on the same static text. Metrics turn it into currency; currency attracts paper mills; volume overwhelms review; success stories hide failures; and generative AI has cut the cost of producing it to almost nothing. Terence Tao calls the predicament that AI creates for mathematics “proof indigestion”: AI-generated proofs will accumulate faster than they can be verified, and verified ones faster than they can be given a readable write-up; well-written proofs will overwhelm a peer-review system that depends on volunteer labour, and even the published ones will be too numerous for the community to work into definitive form.¹¹⁴ The bottleneck has moved from writing things down to verifying, explaining and evaluating them – and none of these is something a paper can carry on its own. The mathematician David Bessis puts it more bluntly: the “theorem economy”, in which only proving theorems counts as a contribution, is headed for collapse. As he reminds us, quoting William Thurston: “The product of mathematics is clarity and understanding. Not theorems, by themselves.”¹¹⁵

Back to Gun and Yu. Right after saying that “Gun dammed the flood and was put to death”, the Guoyu continues: “Yu, by his virtue, was able to mend Gun’s work.”⁶ The Mencius records Yu’s method: “He dredged the nine rivers, cleared the Ji and the Ta, and channelled them into the sea.”¹¹⁶ Yu did not tear down all of his father’s works; what he changed was where the water went. When I talk about technology I often invoke Heidegger’s “releasement” (Gelassenheit): neither rejecting technology wholesale nor submitting to it mindlessly, but first treating it as a natural environment. The flood rolls on and cannot be held back, but it can be steered by working with the current – and the brave can even stand on the crest of the wave and ride it.¹¹⁷ Applied to today, this means we need not abandon repairs, but we do need a different way of thinking. Change the unit: let a piece of research exist as a knowledge object, with its core argument, traceable evidence, a record of its process, multimodal explanation, transparent governance, an interface for AI, and co-creative evolution – in other words, the seven layers of a work. Change evaluation: let verifying, replicating, explaining and maintaining become achievements in their own right, achievements that can accumulate. Change the way people take part: let the public be more than readers and, as the revival of citizen science has shown, become part of the scene where research happens. None of this can be left to the journals to change on their own. I have said harsher things: that print journals should all go out of business, and that the research community should build open-source platforms along the lines of GitHub. To put it more gently: if journals want to stay, they must turn from containers into windows – for certification, curation and display – while the complete, traceable version of record is deposited on open platforms.

The new approach carries risks of its own. Layered works may cost more; a system of achievements may breed new ways of gaming the score; opening up the scene of research may bring noise; and no new unit of counting can escape Goodhart’s law. But the difference between dredging and damming is this: the first spreads the pressure across many channels, so that verification, explanation and digestion each have their place and each win recognition; the second piles all the pressure onto a single dyke, and the higher the dyke, the greater the cost when it bursts. The flood of papers is not going to recede. What we have to decide is whether to keep raising the dykes or to start digging new channels.