How can research platforms deal with the ‘slop’ deluge?

Scott Kominers, a professor at Harvard Business School, regularly made an estimated three or four AI-assisted submissions per month to the open-access preprint platform arXiv. That was until 20 July, when the platform’s support team asked him to limit himself to one per month.

An email to him explained that his submissions “have required significant moderator and administrator attention”, adding that the site is moderated by volunteers who have a “limited amount of time”.

Hosting almost 2.4 million papers, arXiv is the largest global open-access repository for sharing early-version research articles in disciplines including physics, mathematics, computer science and quantitative biology. It does not conduct traditional peer review but it does carry out basic checks to ensure that “submissions are relevant to the scientific community and follow arXiv policies”.

The server has been running since 1991 but a recent “exponential increase” in submissions has forced it to grapple with how it can maintain that oversight without overwhelming its “small team of staff and students [who] work with 170+ volunteer moderators to review the classification, metadata, and content of all submissions”. Its submissions have doubled in the past two years alone, and this September it received 40,363 submissions, compared with 9,869 in the same month in 2016. 

Hence, in October, it reacted to this “watershed moment” by announcing that each author will now be restricted to two submissions per month. 

“A relatively small proportion of authors are submitting a large number of low-quality papers and consuming a disproportionate fraction of the moderators’ time,” the chair of arXiv’s editorial advisory council, Thomas Dietterich, said in the blog post announcing the submission cap. “It is unfair to authors who continue to submit quality papers [which] can be delayed for days or weeks as a result.”

The blogpost added that “advanced AI tools are helping drive these increases by making certain aspects of research faster and easier”.

But Kominers is not the only academic who is less than impressed.

He told Times Higher Education that his concerns stem from both being a scholar of marketplaces and a longtime user of the platform – which requires authors to report any “significant use of sophisticated [AI] tools” and noted that community opinion on the use of such tools may be “varied and evolving”. 

When his arXiv use was restricted in July, he was “very hurt and distressed”, he said, since “the arXiv is my primary means of disseminating math papers”. And while he does use AI to “support” his work, he is “assiduous about it”.

But it is evident that not everyone is. And according to Steinn Sigurdsson, arXiv’s scientific director, unless the platform had acted to “slow down the submission rate…the whole thing was going to break”. And although he admitted to THE that the cap was “an expedient measure to buy us some breathing room”, he said the arXiv team had spent some time thinking about how to “minimise the impact on good faith authors”. 

“The most prolific author in mathematical history averaged precisely two papers per month for his entire lifetime – but he did so with 500 co-authors,” he said, referring to the celebrated 20th-century Hungarian mathematician Paul Erdős.  

“If you think you're a greater scientist than Erdős and deserve to be publishing at a higher rate, then come talk to us.” 

Increasing research volumes are a long-term trend, of course. For decades they have been driven by the expansion of the global research base, particularly in China and India. For instance, 20,000 of the 29,000 papers submitted to the 2026 AAAI Conference on Artificial Intelligence involved Chinese authors, according to Michael Wooldridge, Ashall professor of the foundations of artificial intelligence at the University of Oxford. “The first time that conference was held, in 1980, it had one paper submitted from Hong Kong,” and none from mainland China, he noted. 

But AI has exacerbated the problem, not merely by making it much easier for authors to submit  “tons” of such “junk submissions…in the hope that some of it is going to get past the reviewers” but by making it harder for moderators to spot such junk, thus taking up more of their time.

“One of the indicators of very poor-quality work used to be that it was just badly written,” he explained. But now, “Generative AI means that you can disguise poor quality with a veneer of being well written. So, you have to work much harder to figure out that [a paper] is junk.” 

Source: 

sorbetto/Getty Images

Surging submission volumes are not restricted to arXiv. Other preprint servers, as well as peer-reviewed journals and funders, are all grappling with the same problem, sparking widespread concern that the scientific literature’s quality control system could collapse unless remedial measures are adopted.

But the extent of the problem currently varies by field.

Richard Sever, a senior lecturer at the Cold Spring Harbor Laboratory, co-founded arXiv’s equivalents in biomedicine, bioRxiv and medRxiv, in 2015 and 2019 respectively, and he now serves as chief science and strategy officer at openRxiv, the non-profit organisation that operates them. He said bioRxiv and medRxiv have not so far experienced “the same deluge of slop” as their older cousin has. In September 2026, for instance, bioRxiv had 5,049 new preprints posted, compared with 4,770 in the same month last year. But “there are [sharper] upticks in certain areas” such as “entirely computational work”, and Sever fears that this could be an “indication of a surge to come”.

“One of the things we seek to avoid is cookie-cutter submissions, where people are producing dozens of papers of little value by changing one little variable,” he said – and in such cases, typically, “the AI detector shows that the text is 100 per cent AI-generated”.  

But the platforms’ exclusion of narrative review articles (also known as literature reviews), viewpoints and essays has helped keep volumes reasonable since these are more easily manipulated using AI, Sever said.  

“I have always argued that the activation energy for writing a review article is [merely] opening a word file,” he explained. By contrast, the production of biological studies is constrained by the “limiting factor” of the need to conduct physical experiments. And although AI makes it easier to “create an entire fake paper with fake data”, he does not think “there are that many frauds out there”. The “bigger challenge” for the scientific community is “people producing things that are technically correct, but of no value to anyone”, he said. 

Abbott, professor of molecular genetics at the University of Edinburgh and a volunteer affiliate reviewer for the platform, agreed that with the arrival of AI, “the [research] system is just so open to abuse that you have to have some kind of quality check”. And although both bioRxiv and medRxiv require authors to detail their use of AI and stress their accountability for “the accuracy, integrity, and originality of their work”, she noted that bioRxiv does permit systematic reviews, in which “you explain the methods that you have used for reviewing and you have synthesised all the evidence and come up with a conclusion” – and these can be done with GenAI “very quickly”. This could be an example where an author “may not have bad intentions, but it could swamp the whole system with slop”, she said. 

The use of the word “slop” to describe AI-assisted output has become commonplace. But Philip Moriarty, professor of physics at the University of Nottingham, urged colleagues to be more discerning.

“We can’t just say everything spat out by AI is slop,” he said, pointing to recent AI-assisted developments in coding and mathematics.

That said, he has no argument with arXiv’s submission restrictions. The idea that anyone could produce more than two studies a month was “absolutely bonkers” to him – at least in his nanoscale branch of condensed matter physics. 

And both Moriarty and Wooldridge agreed that the submissions surge arXiv is experiencing will travel to other scientific fields as researchers become more proficient at using AI tools – not only in science but also social science.

There, the main preprint server is the Social Science Research Network (SSRN), launched in 1994 by two economists but now owned by Elsevier. According to reports, SSRN recently removed 257 articles submitted by a single author – Nicholas Polson, junior professor of econometrics and statistics at the University of Chicago Booth School of Business. He had submitted 200 articles in the first nine months of 2026. 

A spokesperson for Elsevier said: “SSRN does not currently impose a cap on submissions. We continually monitor activity on the platform and maintain a range of policies, controls, and verification processes to support [its] quality and integrity…We regularly review these measures to ensure they remain appropriate and effective.”  

Source: 

elenabs/Getty Images

Effectiveness is another contested concept, of course.

Harvard’s Kominers admitted that arXiv’s policy will help people “optimise for the best selection of their papers” but he does not believe it will fully solve the volume problem since it won’t curb the number of new authors publishing on the platform. And he suggested that it is their submissions that moderators should focus on – perhaps aided by an “AI tool”, which “could be used to figure out how much moderation a paper needs”.

For authors with a “historical reputation for high-quality output”, meanwhile, scrutiny should be limited to an occasional “spot check”, Kominers believes, to use the “constrained resource” of moderation time “as efficiently as possible”.

For its part, arXiv, too, has recognised the particular threat to quality posed by new authors. That is why, at the beginning of this year, the platform updated its policy to require submissions by previously unpublished authors to be endorsed by “an established arXiv author in the same endorsement domain”. The move was billed as an attempt to stem “the flood of low-quality, non-scientific submissions”.

“But the danger is [that arXiv] becomes a clique” that disadvantages early career researchers, Oxford’s Wooldridge warned, adding that his own inbox had become inundated with endorsement requests. 

Another possible response to the submission deluge, he suggested, would be to charge a “modest” fee to submit a piece, he said. “The problem with that approach is that it…immediately advantages groups that are well funded and immediately disadvantages groups that are not,” he conceded. 

In short, “nobody has come up with a good solution” to the volume problem – but arXiv’s submissions cap “doesn’t seem an unrealistic limit, actually, for almost all working scientists”.  

Source: 

elenabs/Getty Images montage

But Kominers still isn’t convinced. Another negative unintended consequence of the policy, for him, is that it will end up “pushing people to other platforms”, he said, leading to the “fragmentation of all the new knowledge, [which would be] very costly for the field” of mathematics. 

One alternative venue is the newly established preprint server Hexagon, which also covers the mathematical sciences and theoretical computer science but deliberately has more permissive submission rules.  

On its 10 October launch, the platform’s founders said that given the need for “the proliferation of results proved with LLM [large-language model] assistance”, they “wanted to allow lower barriers for authorship and for community interest than the arXiv.”

arXiv’s Sigurdsson, who is a board member advising Hexagon, stressed that arXiv’s policies are “deliberately conservative” and added that there needs to also be a place for more liberal experimentation with AI. To that end, Hexagon “specifically allows submissions entirely generated by LLMs and even where the submitters claim no human understanding”. 

But that lack of human understanding is a matter that arXiv’s Dietterich, who is distinguished professor emeritus at Oregon State University, recently raised alarm about on X. And such a laissez-faire attitude to AI-assisted production isn’t a route that bioRxiv and medRxiv will be going down any time soon either, Sever emphasised. Neither platform currently has any plans to drop the human screening that it augments with automated checks for things like plagiarism and AI use.

“One can envision that in the future for some manuscripts automated screens might function like spam filters for certain categories of paper, but we are not at that point,” he said. “We would need to assess their effectiveness and create appropriate appeals processes.”

But although bioRxiv and medRxiv are not currently contemplating submissions caps, Sever argued that arXiv’s cap was necessary. “It's all very well for people to not like the rate limits, but the alternative is the entire system of arXiv collapses,” he said.  

Similar quota-based solutions are being adopted by computer science journals and funders in the field. But Sever described these moves as “a temporary solution…until we figure out what the long-term solution would be”. And that, he thinks, demands a shift in “the ratio between data, outputs and narrative papers”. 

For instance, in biomedical sciences you can sequence a thousand genomes. “Do you write a thousand papers, or do you have a thousand data deposits and write one paper?” he asked. 

Sigurdsson agreed with Sever that arXiv’s current solution is unlikely to be permanent. “We're not done yet. We are trying to be responsive and adapt our policies as necessary, trying to take into consideration both our internal [and external] needs,” he said.  

There are many “tensions” currently being considered, he said. And in terms of submissions policy evolution, “There’s much more coming in the near future. And I think, if anything, the teams are going to accelerate and be even more rapid,” he said. 

Sigurdsson also said that “AI reviewers are coming” and that there are already “claims that they are more consistent and thorough than humans”. He said there is “a little saying that the only thing that scales like AI is AI…and there’s truth there.” 

But, like Sever, he thinks the only long-term solution is “to change the incentive structure for academia”, so that quality of publication is valued over sheer productivity.

“The way things are now, counting papers and counting citations is meaningless,” Sigurdsson said. “Because generating a paper, and the associated citations to many other papers, is a trivial task. I could literally make a paper that would be acceptable in some sense while we’re talking.” He was speaking to THE on a Zoom call for 20 minutes. 

A shift in career incentives away from publication volume has long been advocated for, of course, and the principle is frequently endorsed by research institutions. But convincing busy appointment and promotion committees genuinely to put the principle into practice has proved much harder. And even in the age of AI, Sigurdsson is realistic about arXiv’s ability to move the dial to such an extent that its author caps would no longer be necessary. 

“That’s a big task for a little organisation,” he conceded.