Brandolini's World

Yikes! I don’t think anyone realized just how many academic norms rested on the fact that writing is rate-limiting. Writing used to be hard, so if you bothered to write, we all took it to mean you’d spent time developing the idea. That was a convenient fiction.

In the LLM era, I guess we’re all learning Brandolini’s law:

The amount of energy needed to refute bullshit is an order of magnitude bigger than that needed to produce it.

— Alberto Brandolini, January 11, 2013¹

Moving forward, I’ll be collecting examples of how the slop firehose is drowning academic scientists and institutions.

Grants and funding

A global uptick in new grants

Across 12 funders in six countries and the EU, applications rose 57% on average from 2022 to 2025.

When both proposals and reviews are mediated by agents that are trained on the same body of previously funded work, the system will no longer be evaluating the quality of ideas. It will be evaluating how well agents have learnt to simulate the ideas that funders have previously rewarded.

— Geraint Rees and James Wilsdon, Nature, April 27, 2026²

The EU’s prestigious Marie Skłodowska-Curie Postdoctoral Fellowships are absolutely drowning in slop!

This is the highest number of applications received for MSCA Postdoctoral Fellowships to date and the highest number of proposals ever for any funding action in the 40 years of EU research and innovation framework programmes.

— European Commission, September 11, 2026³

NIH caps applications

Since September 25, 2025, NIH accepts at most six applications per PI per calendar year, and will not consider applications “substantially developed by AI.”

There is evidence that the use of AI tools has enabled Principal Investigators to submit more than 40 distinct applications in a single application submission round.

— NIH, NOT-OD-25-132, July 17, 2025⁴

What does that mean? Nobody knows! The first rule of Academic Fight Club is “you don’t talk about … AI-assisted writing.”

(Let’s break that rule. I use LLMs to help with my writing, this post included.)

Center for Scientific Review (CSR) triages two thirds of applications

CSR’s mission is to review nearly all new NIH grant applications.⁵

More than 35,000 applications are already in review for the January 2027 council, past the record 33,640 for May 2026. Study sections will discuss a third of applications, down from the usual ~50%.

NIH finds it necessary to continue modified peer review practices to effectively review an unprecedented number of grant applications.

— NIH, NOT-OD-26-114, September 8, 2026⁶

European Research Council (ERC)

The ERC is the EU’s funder of investigator-driven frontier research.⁷ Starting Grant proposals rose 22% in the 2026 call, after 13% the year before.⁸

Recently the number of applications has been rising dramatically. Some panels now have more than 250 applications to assess and we expect this upward trend to persist.

— Maria Leptin, ERC President, April 16, 2026⁹

Wellcome

Likewise, most UK and European funders all reported 50 to 100% more applications in their most recent round (2026).

It is likely that the use of AI is making applying for grants easier and driving up numbers. […] A rise in applications places an unsustainable strain on our operations and the capacity of peer reviewers.

— Alyson Fox, Director of Research Funding, Wellcome, July 14, 2026¹⁰

Publications

arXiv

The entire arXiv post is worth reading for the astounding explosion in arXiv submissions. Clearly, something has to change.

A relatively small proportion of authors are submitting a large number of low-quality papers and consuming a disproportionate fraction of the moderators’ time. This is unfair to authors who continue to submit quality papers — their papers can be delayed for days or weeks as a result.

— Thomas Dietterich, chair of the arXiv Editorial Advisory Council, October 1, 2026¹¹

Incidentally, I would love to know whether bioRxiv is also undergoing the same type of slop attack.

Journal submissions

Journals on ScholarOne got 33% more submissions in Q1 2026 than a year earlier. This burden is falling disproportionately on smaller journals, which got 81% more submissions. Desk rejections grew 72% from 2022 to 2025. This is a massive strain on volunteer editors.

The editorial labor problem is more acute, and I think it is the least discussed consequence of this surge. The journals absorbing the most disproportionate growth are exactly the journals with the least infrastructure.

— Josh Dahl, The Scholarly Kitchen, May 13, 2026¹²

Letters to the editor

Clinical Orthopaedics and Related Research got more letters in the first half of 2025 than in any full year before. About half were likely written by an LLM.

We have too much respect for our readers to flood the zone with so much empty content.

— Samantha League Smith and Seth Leopold, CORR, August 4, 2025¹³

JAMA bans AI-written letters

JAMA went further. As of August 10, 2026, it no longer allows AI to draft opinion pieces, letters to the editor or online comments at all, after a surge of submissions in exactly those formats.

We have seen marked increases in submissions of these types of articles, occasionally from the same authors, who may not have expertise in the topics they are addressing. These submissions are often formulaic and lack useful insights.

— Annette Flanagin, Roy Perlis and Kirsten Bibbins-Domingo, JAMA, August 10, 2026¹⁴

AI conferences

AAAI received more than 30,000 submissions for 2026, double the ~15,000 for 2025.

The rising scale of submissions means we face increasingly overburdened reviewers with more papers assigned, the need to recruit an ever-wider pool of potentially less experienced reviewers, and increasingly compressed timelines.

— Joydeep Biswas et al., April 15, 2026¹⁵

AI-written peer review

Of course, scientists are also increasingly offloading peer review. For example, the International Conference on Learning Representations (ICLR), one of the top machine-learning venues, had 19,525 submissions for 2026 and 76,139 reviews from 18,054 reviewers. Pangram, which sells an AI-text detector, estimated that about 21% of the reviews (15,899) were fully AI-generated, and that more than half had some AI involvement.¹⁶

The large scale of reviews (over 75,000) made it impossible to rely on human judgement alone to consistently flag and discover problematic reviews.

— ICLR 2026 Program Chairs, March 31, 2026¹⁷

So ICLR started a game of cat-and-mouse, which I’ll be watching with great amusement. A poor man’s steganography,¹⁸ if you will. They embedded a hidden prompt in some PDFs that nudged LLMs to use telltale phrases. The results were, as you’d expect…discouraging.¹⁹

Mathematics

Erdős problems

Paul Erdős²⁰ left behind hundreds of open problems²¹ that are relatively easy to state but very hard to prove. Thomas Bloom has collected them at erdosproblems.com.

On October 6, 2026, Thomas Bloom froze problem comments and proof claims on his site.

In practice the vast majority of the comments the site receives now are people announcing AI-generated proofs. It is not sustainable to have a moderation policy which would reject almost all the comments that are submitted.

— Thomas Bloom, erdosproblems.com, October 6, 2026²²

What’s to be done?

Thomas Dietterich, who helps run arXiv’s computer science section, proposes a zero-tolerance policy. His proposal would ban offenders from submitting to arXiv for a year.²³ Others have called for “deterrence, not assistance; accountability, not coaching.”²⁴

For academic funding, Martin Bulla and Peter Mikula argue that lotteries among the top-scoring proposals are “no longer a radical suggestion.”²⁵

I’m wondering whether we need to charge for submitting slop (either money or credits). Perhaps a small per-submission fee, refunded on acceptance and waived for researchers in low-income countries. Better still, institute escalating fees for repeat offenders, with the money going towards bolstering our volunteer-based scientific institutions (I see you Elsevier!).

Interestingly, economists already tried this experiment. Johan Fourie surveyed the 500 top-ranked economics journals that take unsolicited papers and found that 23% charge a submission fee: 68% of the top 50, but only 2% of the bottom 50.²⁶ I imagine that this just pushes AI slop to lower-tier journals that had less restrictive submission barriers.

Will any of this work? I have no answers, but boy-o-boy, is bullshit expensive!

What a time to be alive.

Notes