- papers are read, summarized and digested by AI, because there are just so many papers at leading AI conferences that nobody has time to eyeball them all
We are very rapidly automating humans out of the academic publication loop here.
> We are very rapidly automating humans out of the academic publication loop here.
It more looks like the breakdown of the current academic publication system, which was rotten to the core pre-AI and which internal contradictions are just accelerated by AI to the point of breakdown now.
and don't forget, techniques described in papers are now implemented by AI's. For now, a human might point an AI at a paper and ask it to implement and benchmark the technique described there, but the human probably isn't writing the code anymore.
Human reviewers were not what they were made out to be. They would approve you if you cited them, they would disapprove you if your work questioned the validity of their work.
The entire concept of publishing as a validation step for actual science has been declining for decades, fully co-opted by pay to play and virtue signaling for academic hiring.
AI is simply going to bury it as a useful system. What will grow from the ashes will be something much more dynamic, where verifiable data is the gold and the conclusions and associated details will persist only as the human-level translation.
> because there are just so many papers at leading AI conferences
Ironically a big reason there's so many papers to review is because so many are rejected.
A low acceptance rate is unhealthy, especially in conferences (1 round of review). Papers just get recycled to the next conference, which, as is easy to model, creates an exponential feedback loop. It doesn't explain all the papers submitted, but it sure can explain a lot. Too much rejection is like shooting yourself in the foot.
Not to mention that it's just easy to reject works. All works are flawed, especially works that are in less mature domains. I see plenty a paper get rejected for lack of money. "Not enough experiments" is an common critique that's used inappropriately (along with the highly subjective "not novel enough" one) because it's fine to always want more but no lab has infinite funding. It is used lazily. The question shouldn't be about if your favorite benchmark is used, it should be if there isn't enough evidence to support the hypothesis or not. A mature domain where thousands of people work in it, yeah, that needs stronger evidence. A niche domain where dozens of people work in? Not as many required. Rejecting them ultimately slows down the progress of science because you require any new idea to outperform mature ideas. Ironically killing novelty as no one is going to, or even could (publish or perish), spend all the time and money to mature a niche all on their own.
> I can imagine in 2027 academia looking like Moltbook.
This is certainly “directionally correct”, but keep in mind this is all highly uneven across fields and subfields. The people generating slop articles are mostly trying to publish big flashy things, and naturally ML research has it much worse than most other fields. There are many topics that are super important and interesting, but niche or obscure enough that no slop authors is trying to publish on them yet. So plenty of topics are still dominated by real earnest researchers doing their best, but they’re niche enough that you wouldn’t know about them unless you study that field.
> Both papers were accepted for oral presentations with the condition that they simply fix the hallucinated references.
I do wonder what truthfully could be on ai verification, if even one paper with such an error is accepted it sets the precedent you hopefully get lucky to not get caught (then again verifying for basic tells isn't the same verifying is this genuinely a worthwhile publication, but that's a separate matter)
"A lot of the content of this blog was initially drafted by an agent of some sort"
What? I mean who does this. My voice is my voice and it's literally never occurred to me to have an LLM do a first draft. I thought that was college kid stuff.
For anyone that wants to evade the kind of people who want to figure out if an AI wrote your review or not, we wrote a whole paper (ICLR 2026!) on how to do that!
I consider all types of "I liked this output, but don't the moment I learned it was AI generated" to be externalizations of "carbon chauvinism" (https://en.wikipedia.org/wiki/Carbon_chauvinism) and basically bigotry.
And BTW, the term "meritocracy" was coined in a book that was extremely critical of the idea and which argued that a real meritocracy is actually dystopian. We consider our work "harming meritocracy" to be a good outcome: (https://en.wikipedia.org/wiki/The_Rise_of_the_Meritocracy)
> This all is very annoying from inside the review queue. Peer review is unpaid work that we do
The genie is out of the bottle. We need to figure out a way to contain it. I think a solution is to use LLMs for peer reviews also; fight fire with fire?
I'm concerned that slop authors (or their agents) will use bib-audit in the loop, and hence have perfect references, thereby denying a clear signal of low quality research.
This is foolish for several reasons. It’s an undue burden on authors to make them download and upload potentially a gigabyte of files (or more) from many different sources. It is also likely a copyright violation for many or the sources. This also makes it arbitrarily difficult to cite things like conference talks, which are not published texts.
There are already standard keys to index publications, like DOIs. Requiring a list of DOIs for citations would make a lot more sense and be somewhat feasible, but still doesn’t prevent errors in the author list given in the draft.
The advantage of "publish or perish" is that it is based on something concrete. The system is gamed, for sure, and even more so with AI, but still, a paper which is cited a lot tend to be useful and the authors get rewarded.
Remove that metric, and for the lack of a better idea, it will just turn into a game of who has the best connections or who talks the most convincingly, I mean, even more than it is now.
That misses the fact professors have plenty of other requirements beyond publications already. Teaching requirements, service requirements, chairing committees, taking on and graduating grad students. All of these are considered when one goes up for tenure.
It is in alpha (hoping to do Show HN in a few weeks), but for those interested I am working on an application to do this, https://veruscite-data.com/
Most of the folks on HN will be more familiar with genAI tools and can just use the skill Caleb and Isaac provided in this blog post. My tool is just likely more token efficient and has a GUI where you can review the extracted bib and edit it more easily.
I was wondering why they don’t just feed the papers first to an LLM to spot obvious slop, and I was answered later in the post:
> Submissions are confidential, bibliographies included, and the audit works by sending pieces of one to a hosted LLM – even though the LLM never writes a word of your review. ECCV 2026’s reviewing policies state that LLMs “are NOT allowed to be used to write reviews or meta-reviews, whether it is run locally or via an API,” and separately bar reviewers from sharing substantial excerpts of a submission with an LLM. WACV’s reviewer guidelines call LLM-generated reviews “highly irresponsible behavior,” sanctionable by desk rejection of the reviewer’s own papers, and their confidentiality rules forbid showing a submission’s material to anyone who is not a reviewer – which a hosted LLM is not. NeurIPS’s LLM policy restricts what reviewers can share with LLM services; its AI-assisted reviewing experiment is the sanctioned route.
So you can send some LLM generated crap to be published and even if you get caught, there is no consequences. Since your funding is likely connected to how much you publish, even if it’s toilet paper, so this system actually rewards one from spewing out shit papers no-one reads. But if you use LLM to review them, guess what, you will get punished harshly.
I think if you send in LLM crap with hallucinated citations you should get 5 year ban on even sending anything to that conference or publication. And perhaps we should create a local model -based application that filters out this crap. The one the authors had made is a good start, but surely you don’t need Claude to review a bibliography for errors? Surely Qwen with a SearXNG limited to arxiv etc. can do the job?
I didn't read carefully enough at the beginning of the article, and at first I thought Caleb and Issac must be two different algorithms for detecting AI in published papers. Eventually I realized they are the names of the two authors who wrote this article (with AI assistance). Anyways, I was wondering how the two of them managed to disagree? Were the papers each human flagged as concerning the same between the two humans? I should keep reading carefully, just thought I would point it out to other humans too, maybe save them from the same mistaken thought.
We should oust and shun anybody being caught producing or spreading fake science. This is some of the most egotistical, anti-social behavior imaginable.
38 comments
[ 0.27 ms ] story [ 51.6 ms ] thread- papers are written by AI (as pointed out in this article, and as obvious to anyone who spends a while actually reading recent AI research)
- papers are reviewed by AI (NeurIPS is doing an AI assisted review experiment - https://neurips.cc/Conferences/2026/ai-reviewing-experiment - and I feel the trend is moving towards AI reviewers whether we like it or not)
- papers are read, summarized and digested by AI, because there are just so many papers at leading AI conferences that nobody has time to eyeball them all
We are very rapidly automating humans out of the academic publication loop here.
It more looks like the breakdown of the current academic publication system, which was rotten to the core pre-AI and which internal contradictions are just accelerated by AI to the point of breakdown now.
AI is simply going to bury it as a useful system. What will grow from the ashes will be something much more dynamic, where verifiable data is the gold and the conclusions and associated details will persist only as the human-level translation.
A low acceptance rate is unhealthy, especially in conferences (1 round of review). Papers just get recycled to the next conference, which, as is easy to model, creates an exponential feedback loop. It doesn't explain all the papers submitted, but it sure can explain a lot. Too much rejection is like shooting yourself in the foot.
Not to mention that it's just easy to reject works. All works are flawed, especially works that are in less mature domains. I see plenty a paper get rejected for lack of money. "Not enough experiments" is an common critique that's used inappropriately (along with the highly subjective "not novel enough" one) because it's fine to always want more but no lab has infinite funding. It is used lazily. The question shouldn't be about if your favorite benchmark is used, it should be if there isn't enough evidence to support the hypothesis or not. A mature domain where thousands of people work in it, yeah, that needs stronger evidence. A niche domain where dozens of people work in? Not as many required. Rejecting them ultimately slows down the progress of science because you require any new idea to outperform mature ideas. Ironically killing novelty as no one is going to, or even could (publish or perish), spend all the time and money to mature a niche all on their own.
If all these papers were not gatekept by journals, it would be trivially easy to validate at least the existence of cited papers and quotes.
Alas, it's probably just wishful thinking on my part.
They should consider swapping this for a log plot.
I can imagine in 2027 academia looking like Moltbook.
This is certainly “directionally correct”, but keep in mind this is all highly uneven across fields and subfields. The people generating slop articles are mostly trying to publish big flashy things, and naturally ML research has it much worse than most other fields. There are many topics that are super important and interesting, but niche or obscure enough that no slop authors is trying to publish on them yet. So plenty of topics are still dominated by real earnest researchers doing their best, but they’re niche enough that you wouldn’t know about them unless you study that field.
I do wonder what truthfully could be on ai verification, if even one paper with such an error is accepted it sets the precedent you hopefully get lucky to not get caught (then again verifying for basic tells isn't the same verifying is this genuinely a worthwhile publication, but that's a separate matter)
What? I mean who does this. My voice is my voice and it's literally never occurred to me to have an LLM do a first draft. I thought that was college kid stuff.
https://arxiv.org/abs/2510.15061
I consider all types of "I liked this output, but don't the moment I learned it was AI generated" to be externalizations of "carbon chauvinism" (https://en.wikipedia.org/wiki/Carbon_chauvinism) and basically bigotry.
And BTW, the term "meritocracy" was coined in a book that was extremely critical of the idea and which argued that a real meritocracy is actually dystopian. We consider our work "harming meritocracy" to be a good outcome: (https://en.wikipedia.org/wiki/The_Rise_of_the_Meritocracy)
The genie is out of the bottle. We need to figure out a way to contain it. I think a solution is to use LLMs for peer reviews also; fight fire with fire?
Is that really a thing? So if I submit a legit paper it is being reviewed forcefully by a bunch of random people from god knows where?
There are already standard keys to index publications, like DOIs. Requiring a list of DOIs for citations would make a lot more sense and be somewhat feasible, but still doesn’t prevent errors in the author list given in the draft.
I know managers and administrators want a simple metric, but there is no simple metric for research.
The advantage of "publish or perish" is that it is based on something concrete. The system is gamed, for sure, and even more so with AI, but still, a paper which is cited a lot tend to be useful and the authors get rewarded.
Remove that metric, and for the lack of a better idea, it will just turn into a game of who has the best connections or who talks the most convincingly, I mean, even more than it is now.
Even with the technical debt. We have zillions of papers published, we know that most of them probably won't replicate, we don't know which ones.
Most of the folks on HN will be more familiar with genAI tools and can just use the skill Caleb and Isaac provided in this blog post. My tool is just likely more token efficient and has a GUI where you can review the extracted bib and edit it more easily.
> Submissions are confidential, bibliographies included, and the audit works by sending pieces of one to a hosted LLM – even though the LLM never writes a word of your review. ECCV 2026’s reviewing policies state that LLMs “are NOT allowed to be used to write reviews or meta-reviews, whether it is run locally or via an API,” and separately bar reviewers from sharing substantial excerpts of a submission with an LLM. WACV’s reviewer guidelines call LLM-generated reviews “highly irresponsible behavior,” sanctionable by desk rejection of the reviewer’s own papers, and their confidentiality rules forbid showing a submission’s material to anyone who is not a reviewer – which a hosted LLM is not. NeurIPS’s LLM policy restricts what reviewers can share with LLM services; its AI-assisted reviewing experiment is the sanctioned route.
So you can send some LLM generated crap to be published and even if you get caught, there is no consequences. Since your funding is likely connected to how much you publish, even if it’s toilet paper, so this system actually rewards one from spewing out shit papers no-one reads. But if you use LLM to review them, guess what, you will get punished harshly.
I think if you send in LLM crap with hallucinated citations you should get 5 year ban on even sending anything to that conference or publication. And perhaps we should create a local model -based application that filters out this crap. The one the authors had made is a good start, but surely you don’t need Claude to review a bibliography for errors? Surely Qwen with a SearXNG limited to arxiv etc. can do the job?