The deeper question is whether legitimate AI generated results are allowed or not? Test - In the extreme - think proof of Riemann Hypothesis autonomously generated (end to end) formally proven - is it allowed or not?
The thread specifically points out that if authors can’t be arsed to simply proofread their text the rest can not be trusted either.
It’s a simple heuristic against low quality submissions, not an anti-ai measure.
I expect arXiv will still have problems with slop submissions but, at least, their references should actually exist going forward.
Sorry to be rude, but this seems like a dumb question. I want science to progress. A primary purpose of these journals is to progress science. A full proof of the Riemann Hypothesis progresses science. I don't care how it was produced, if Hitler is coauthor, etc, I just care that it is correct. Whether the authors should be rewarded for whatever methods they used can be a separate question.
The short of it is he argues how first to correctness shouldn't be the only goal / isn't a great optimisation incentive. Presentation and digestibility of correct results is a missing 1/3 when you've finished generation and verification. I completely agree with him. You don't just need an AI generated proof of the Reimann Hypothesis. You would really like it to be intentional and structured for others to understand.
A really beautiful quote I learned of in the talk is this:
> "We are not trying to meet some abstract production quota of definitions, theorems, and proofs. The measure of our success is whether what we do enables people to understand and think more clearly and effectively about math." - William Thurston
This is incredibly good for science. arXiv is free, but it's a privilege not a right!
I'm not seeing this clearly listed on https://info.arxiv.org/help/policies/index.html so it's possible this is planned but not live yet - or perhaps I'm not digging deeply enough?
As a certain doctor once said: the whole point of the doomsday machine is lost if you keep it a secret!
This is good for reference checking, but I doubt this will do much for the most likely shoddy science that accompanies hallucinated references.
Catch-22 [1]:
You will need to provide an arXiv article id. This is necessary for your paper to be processed correctly, the submitted version must be the same as the arXiv version.
[1] https://jcap.sissa.it/jcap/help/helpLoader.jsp?pgType=author
Are all venues like this, however?
The penalty cannot be high enough.
ArXiv doesn't even check the submission closely, so how can they know?
They say "errors, mistakes"
They use an automated system to check if the basic requirements were met, and sometimes papers are flagged for further superficial human review, but there is no way they can possibly do this at scale or check every reference. This would be like trying to do peer review, but for a preprint archive that gets easily 100x more volume than any journal.
Second, there is such a huuuuge gap between publishing on arvix and peer review. I can attest personally that it's not even close. I've gotten probably dozen rejections from peer review and no problems publishing in arxiv math. This is because peer review checks not just for if something is new or correct, but also if it's of "interest to math community," which is inherently subjective, but also makes peer review many magnitudes harder than publishing on arxiv.
Even when a well-known professor in number theory praised the paper when I got an endorsement and a second emailed me and and encouraged me to publish it, it still got rejected 3 times and still waiting.
Being required to publish in a peer reviewed journal will close off arxiv for many researchers for good. It also defeats the point of it being a pre-print.
Nothing stops someone from putting a PDF on the internet. I'm fine with ArXiv holding a high standard.
They can be informed by people who read the papers and check the citations. A zero-tolerance policy provides an incentive to report sloppy papers (namely, that you can be confident something will be done about it), and each time a paper is removed or an author is banned, it incrementally increases the value of the arXiv as a whole.
> Being required to publish in a peer reviewed journal will close off arxiv for many researchers for good.
At the end of the day, demanding that people carefully proofread their LLM-generated papers before sharing them on the arXiv seems like a relative low bar to clear, and I sort of question whether it's reasonable to call individuals who find it too onerous "researchers" in the first place.
It's enough for them to place this policy and enforce it when they become aware of violations. Someone reading the slopped paper (or, here, trying to follow a reference) will notice sooner or later.
> Being required to publish in a peer reviewed journal will close off arxiv for many researchers for good. It also defeats the point of it being a pre-print.
You sound like it's impossible for researchers to write papers without slopped references, and inevitable to get hit by this policy.
I disagree. It's just one darn hallucinated citation for heaven's sake, not fraud or something. It doesn't account for the substance or quality of their work at all. A one-year ban seems plenty sufficient for a minor first time mistake like this. People make mistakes and a good fraction of them can learn from those mistakes. There's no need to permanently cripple someone's ability to progress their life or contribute to humanity just because an AI hallucinated a reference one time in their life. That's punitive instead of rehabilitative.
It is fraud.
> It doesn't account for the substance or quality of their work at all.
References are part of the work. If you're making up the references, what else are you making up?
> People make mistakes and a good fraction of them can learn from those mistakes. There's no need to permanently cripple someone's ability to progress their life or contribute to humanity just because an AI hallucinated a reference one time in their life.
A one year ban is not permanent. Having a negative consequence for making poor decisions seems like an inducement to learn from the mistake?
In an ideal world, one would be keeping notes on references used while doing the research that lead to writing the paper. Choosing not to do that is one poor decision.
Having a positive outlook, if asking an AI to provide references that may have been missed, one should at least verify the references exist and are relevant. Choosing not to do that is also a poor decision, even if one did take notes on references used while researching.
Your standards are lower than what they would accept at my high-school. Seriously.
And generally, if you are generating papers with LLMs, let other LLMs read them. Why would we waste human hours considering something that was generated? At this point publish your prompt because that's the actual work you're doing.
I don't think you need to publish on arXive to contribute meaningfully to humanity.
> That's punitive instead of rehabilitative.
Unfortunately science is competitive. Yours is a race to the bottom where the people who can afford the most expensive models and who are least concerned with the truth can publish the most papers and benefit financially and professionally by doing so. This is not a zero sum arena, grant money and opportunities will possibly be rewarded to them, and not to another team who is producing more careful and genuine output.
Your being set behind is less important than the fact that your publishing is setting everyone else behind.
Such a banned person is being helped to "step out of the way", and someone more competent will assuredly step forward to consume the limited maintenance labour more thoughtfully
It's not even that they "don't like LLMs". They just don't like academic fraud! If references were fabricated with a Markov chain it would be just as bad!
Bonkers. At the same time peak hn
I'm here because I enjoy building things. And today this mostly happens with AI. I could do without the often thoughtless comments and conspiracy theories about "LLM hypers" posted by people who don't like LLMs.
If it’s not worth your time to check the output of your LLM carefully, it’s not worth my time to read it.
Ever pick a random one and really dive in?
And I’m not talking about good faith research that didn’t pan out, I mean research that is completely useless for any other purpose other than convincing a casual observer that the authors are doing research.
$ curl -L "https://doi.org/10.47397/tb/43-1/tb133chernoff-widows" -H 'Accept: application/x-bibtex'
@article{Chernoff_2022, title={Automatically removing widows and orphans with <tt>lua-widow-control</tt>}, volume={43}, ISSN={0896-3207}, url={http://dx.doi.org/10.47397/tb/43-1/tb133chernoff-widows}, DOI={10.47397/tb/43-1/tb133chernoff-widows}, number={1}, journal={TUGboat}, publisher={TeX Users Group}, author={Chernoff, Max}, year={2022}, pages={28–39} }
So there are absolutely a bunch of tasks that could be evaled/benchmarked, but "hallucination rate" isn't particularly applicable/interesting as a metric of how good the tool is
that said, we do use various LLMs (mostly local, fine-tuned, small, for things like NER/parsing/metadata comparison, etc.). and they can and do hallucinate, but we have very hard constraints on the validation, so any extraction results that don't match 1:1 back to the input text are discarded for example. so again, rather than hallucination risk we prefer hard constraints
To be a coauthor on a preprint that you have not submitted, you have to actively "claim" it (using a password given to the author who submitted). It's on you to double-check before claiming.
I surely hope that only "confirmed" coauthors will get the ban, it's only logical.
The authors should value the time of the reviewers higher than their own time. So, if you include AI nonsense in your paper, it is insulting.
If you can't validate that your bibliography is full of real articles, you shouldn't get published.
LLMs have just poured gasoline on the fire.
He's toast if SSRN were to adopt a similar policy.
I'm a screen reader user and usually read papers as raw TeX. I've seen everything: slurs, demeaning comments towards reviewers and professors, admissions of fraud, instructions to coauthors to commit further fraud before paper submission to mask the earlier fraud... it's all there. There's far less of it than I would think, definitely <1% of papers, but it's there.
I think it would be useful to run an LLM anti-fraud pass on the TeX source of all new arxiv papers. It wouldn't catch everything, but it would catch some of the dumbest fraudsters.
On the positive side, you can also find stronger claims that didn't survive review, additional explanations that didn't make the cut due to the conference's page limit, as well as experimental results that the authors felt weren't really worth including. Those need to be approached with an abundance of caution, but are genuinely useful sometimes.
I have standing instructions with my agents: do not commit anything without review by at least two sub agents.
The chance of two agents hallucinating at the same time in the same way is almost nil. Why do people still do it?
Now: „Generate a bibliography. Make no mistakes.“
read the whole tweet:
If generative AI tools generate inappropriate language, plagiarized content, biased content, errors, mistakes, incorrect references, or misleading content, and that output is included in scientific works, it is the responsibility of the author(s).
It's not hard to divine their intentions: you are entirely responsible for what you summit and if it's clearly slop(py) you get a ban. In a reply they state that they are seeking to apply this rule fairly and accurately and are mindful of unintended effects.
Let's say you read a paper on Arxiv but cite the version that was submitted to a journal or conference, without realizing that the authors made changes to the version they submitted and forgot to upload them to Arxiv.
References which don't accurately reflect the quoted material seem more common in other subjects.
Research and practice has shown that the strongest deterrent is certainty.
If the fine is high enough (risk) , but probability low, people will not do the thing because of the impact on them.
One hallucinated citation does not in any way imply anyone is being left behind. All it means nobody is checked that particular line of the manuscript after it was written. The rest of the paper could still be solid and treated accordingly. If you find evidence of the contrary, of course treat it accordingly, but this is so obviously not that.
The parent said “setting” others behind, which refers to lost time.
Being “left” behind implies a degraded trajectory, which is defined not by time lost, but by the final destination.
Different but related things (e.g. lost time can indeed affect your final destination, for instance, after growing old correcting a scourge of hallucinated citations - which should have been table stakes all along).
If all you're genuinely worried about is the collective human time spent on tracing down one stupid hallucinated citation in a paper, may I remind you of the ludicrous amounts of time and effort readers waste trying to wade through the sea fluff, jargon, and complexity frequently added to papers in a completely deliberate fashion. If wasting even a little bit of readers' time is what you see as the crime here, you have orders of magnitude bigger fish to fry.
The fact is that, for one hallucinated citation to be the noteworthy bit that "sets others behind" in any meaningful way, the actual substance of your paper has to be utterly worthless (or worse); otherwise, you're contributing far more than you're taking away, and thus your paper is very much not setting others behind. OTOH, if your paper really is worthless or harmful enough for this part of it to be a big deal, that would be the basis for punishment, not this. A single hallucinated citation is simply not a bleep on that metaphorical radar.
I can't see that in the code of conduct.
No, this is career ending high stakes. it requires old school "actually check a record of reality" type methods, like a database query or http get to one of the many services that hold this info.
I kind of hate the idea, but you probably could do a lazy LLM check of every paper and every citation and have it flag possible wrong (second sense) citations for human review
But you'd need a LOT of tokens and a LOT of human-hours
And then what, we're done? How have we avoided the need for the same exhaustive human review? It only saves human review time if you trust the LLM not to miss things.
The device-fixer started breaking devices instead of fixing them. Tell it to fix itself!
The amount of people who confidently tell on themselves in these discussions continues to bum me out.
sometimes the presently available solutions are subpar. people go with what's available. it's not ideal, but it is practical.
Seriously? You can't fathom an honest researcher asking for AI to find a citation they know exists, and the AI inserting or modifying a citation incorrectly without them realizing?
If you find evidence of fraud by all means lay down the hammer. Using a single hallucinated citation like it's some kind of ironclad proxy just because you think they must be committing fraud is insane.
yes there will be rare exceptions but in general i feel like this is a really good addition.
What SOTA models are not hallucination prone?
If someone cannot meet that bar, they have no business publishing research papers. I have written academic papers myself, and I find it astonishing that people are trying to justify this as if it were some understandable workflow mistake. At that point it is simply slop with academic formatting. Post it on a blog or somewhere else, but do not put it into the scientific record.
A one-year ban is not a lifetime ban. Maybe six months would also have been enough, but the author can use that time to think about whether they should verify references next time — and to manually check every other citation.
So have I. That doesn't place me on some pedestal.
> A one-year ban is not a lifetime ban.
Nobody said it was. Nobody was even objecting to it.
Please (re-?)read past the first 8 words of [1], the first 3 sentences of [2], or the first 3 tweets of the original tweet thread. You appear to have imagined (shall we call this a hallucination?) and responded to something that was never claimed. Which is especially unfortunate, given the very topic of the lecture here.
if someone writes a paper and an entirely different person takes credit for it without even bothering to check if the actual writer just made shit up, they deserve a lifetime ban. seems like a year is a very light punishment.
Deadlines are not an excuse here. Checking whether a cited book, paper, or passage exists is the absolute minimum standard for scientific work, not an optional extra. I have written academic papers myself, and I find it astonishing that people are trying to justify this as if it were some understandable workflow mistake. At that point it is simply slop with academic formatting.
A one-year ban is not a lifetime ban. Maybe six months would also have been enough, but the point is that the author gets time to think about whether they should verify references next time. They can also use that time to manually check every other citation.
Assumptions:
1. The entire document is loaded into an AI editor
2. The researcher is asking an AI editor to work on his references
3. The researcher has not checked his own references.
This could be avoided at 1, 2 or 3. But even just 1 implies that the researcher knows that they have a hot potato and might critically fuck up and lose all credibility. Being in that scenario and committing to 2 and 3 is at least extreme negligence.
Indeed I cannot. If you do that, you are not, in fact, an honest researcher. You're a lazy hack.
I'm not seeing the problem here. The only problem is that your lab partner should be banned and not you. But being incentivised to check your co-author's work before submission isn't a bad thing.
Nobody was arguing against this.
The citation is part of the substance of the paper. If you YOLOed in a citation without checking it, seems justified to suspect that you may have YOLOed in some data, or some analysis, or maybe even the conclusion.
Do bear in mind the degree of the described scenario. There's quite a difference between having an LLM shit out your entire citation section (and possibly the rest of the paper as well) versus asking the tool to make a targeted edit and overlooking a small piece of nonsense that results.
Yeah, that's exactly what I've been upset about. You really nailed it!
An LLM could replace the random sampling. It doesn't need to be particularly good for the approach to provide value. I would worry about LLM bias though.
Another thing to consider is that readers can detect fake citations after publication, report to arXiv, and the author gets banned.
> In an ideal world, one would be keeping notes on references used
[0] References can be entire textbooks, which is potentially too high of a bar
I submitted a paper with a reference author as Elisio because I couldn’t read my own handwriting. After submitting, I double checked all the references through an LLM. It pointed out that their name was actually Enrique. Yes, you should probably double check your references before submitting, not after.
Point is, I didn’t even trust the LLM at first. But after verifying the mistake, I was embarrassed af. I resubmitted with the fixes before it went live, but ultimately, what’s the difference between “mistake” and “hallucination”?
With proper bibliography management tools, everything (that has one) is centered around the DOI.
In fact, if a DOI is present, it's trivial to verify authors, title, venue, year, pages etc.
Of course, some older and more obscure papers won't have a DOI, but the vast majority of research work has.
Perhaps the issue is that people aren't logged in or using xcancel so missing part of the tweet thread. Here's an important line
> If a submission contains incontrovertible evidence that the authors did not check the results of LLM generation, this means we can't trust anything in the paper.
Followed by
> Examples of incontrovertible evidence: hallucinated references, meta-comments from the LLM ("here is a 200 word summary; would you like me to make any changes?"; "the data in this table is illustrative, fill it in with the real numbers from your experiments")
I wouldn't look at your case and read that as "incontrovertible evidence". They are looking for the absolutely brain dead, no one at the wheel, type of errors. They're looking for things like your paper saying "As an AI language model". Which, there will be real papers with that exact phrase, but it should get flagged, not auto banned
I think we are talking semantics here.
While fraud does require intention to deceive, I get the sentiment that hallucinated citations shouldn't be dismissed as simply carelessness. It should be something stronger than that: gross negligence or something MUCH stronger! There should absolutely be repercussions for this.
But let's not call it fraud. That word is reserved for something specific.
EDIT: someone else said "reckless disregard" equals intent or something to that effect. So I looked it up.
It appears so that is the case. "Reckless Disregard Equals Intent" in legal language.
But I am not sure if this particular clause should apply here. Perhaps it depends on what kind of research is being published? For e.g., if it is related to medical science and has a real consequence on people's health, we can then apply this?
In this book
https://news.ycombinator.com/item?id=44022957
there is this passage on p. 127:
"Any author citing another paper should be required to provide proof that they a) possess a copy of that paper, b) have read that paper, c) have read the paper carefully."
No, it is emphatically not. Fraud requires intent to deceive.
> A one year ban is not permanent.
...what text are you reading? Nobody was calling the one-year ban permanent, or even against it. I was literally in favor of it in my comment. I explicitly said it is already plenty sufficient. What I said is there's no need to go beyond that. My entire gripe was that they very much are going beyond that with a permanent penalty. Did you completely miss where they said "...followed by the requirement that subsequent arXiv submissions must first be accepted at a reputable peer-reviewed venue"?
In fact would be better if they just banned AI, so we could just get off the luddite platforms.
Automated research is the future, end of story. And really it couldn't have come out at a better time, given the increasingly diminishing returns on human powered research.
Which of those two does "hallucinated references" fit into?
It's a conscious decision to not take the time to check your AI output, and instead waste a whole bunch of other people's time letting them essentially do that for you in duplicate.
Feels like that should disqualify you from participation for a bit. Intent or no intent.
Doing your job poorly means giving more work to others and, consequently, stealing their time, their most precious asset.
Many here don't agree with this ban because they work in IT, where this immoral and antisocial behavior is normalized.
Exactly! For a bit!
Yet this is not for a bit! This is a lifetime disqualification, and that's been my entire grip the whole time! Is nobody reading this?
"The penalty is a 1-year ban from arXiv followed by the requirement that subsequent arXiv submissions must first be accepted at a reputable peer-reviewed venue."
The citation being incorrect is merely the proof of deception not the (relevant) deception itself.
Fraud is the correct description provided (and this is practically a guarantee) you intended to benefit from the submission of the paper (e.g. by bolstering your resume).
Fraud in a scientific context generally refers to fabricated research results. At least personally I agree with GP that hallucinated citations are generally something akin to laziness thus not fraud but rather some sort of professional negligence.
But what LPisGood was saying is that reckless disregard (as opposed to explicit intent) is sufficient to meet the legal bar for fraud.
I'm about as pro AI-as-a-research--and-writing-assistant and anti AI-witchhunt as they come, but I simply cannot parse what I've quoted here.
Posting slop to arxiv is blatant deception. Posting an article is an attestation that the article is a genuine engagement with the literature. If you're posting things to arxiv that are not sincere engagements with the literature, you are attempting to deceive.
Ditto. And its only 1 year. Like its about the most reasonable thing they could have done.
Regardless of terminology, I agree that it's certainly punishable and certainly a serious problem.
This part seemed reasonable too. I'm not in academia, but my understanding is most people writing papers intend for them to be accepted by reputable peer-reviewed venues, but post to arXiv because those venues don't always allow for simple distribution.
If your papers aren't going to be accepted at reputable venues and you posted slop to arXiv before (and they noticed it!), seems reasonable that they only want reputable stuff from you in the future?
No, it emphatically is not just a year! It's perpetual, and that's literally been my entire point this whole time. If it was just one year I would've had no complaints - and I made that clear from the very first comment!
What part of "...followed by the requirement that subsequent arXiv submissions must first be accepted at a reputable peer-reviewed venue..." is everyone here reading and still somehow interpreting to be limited to 1 year?
Well, "lifetime ban" means "you are not allowed back in". Their ban specifically allows you back in (after a specified period) subject to fulfilling a single constraint.
It's conditional acceptance back in, which is not the same as a lifetime ban which is unconditional.
I also haven't seen anything on how this works with multiple authors, which could go anywhere from draconian to weakening the entire thing.
Even if it had though, I'd be perfectly comfortable calling this fraud in this discussion based on the common meaning of the word. Just because we're talking about a scientific context does not mean we need to use the scientific-jargon versions of words - we're not in a scientific context ourselves.
---
And I'd disagree that this is just about the "letter of the ToS". While that is perhaps a necessary component in order to prove the deception, this is really about the cultural expectations of the community that merely happened to have been encoded in the ToS. The fraud would still occur without the ToS, it would merely be next to impossible to show you didn't simply misunderstand the cultural norms and what your actions would lead others to believe.
> deception in order to get a benefit.
The point being that reckless or negligent conduct is not commonly taken to constitute deception. There's a reason we have different terms for these things.
Sure, you can say "well he exhibited reckless disregard for his professional duties when he opted not to bother reading the citation section that the LLM shat out, and reckless disregard is sufficient to meet the legal bar for fraud, and also the ToS specifically says that you certify that you validated all references manually so bam! two counts of fraud legally speaking" and you wouldn't be wrong but the distinction between "legally fraud" and "fraud as is commonly meant when talking about scientific papers" is essential to effective conversation in this particular instance.
> Just because we're talking about a scientific context does not mean we need to use the scientific-jargon versions of words
The context is essential because (obviously) it affects how people interpret the meaning of your words. A fraudulent submission to a scientific journal has a specific and well understood meaning in common usage.
If you still disagree with me imagine polling a bunch of tenured career researchers about what they would think if they read the statement "X caught submitting a fraudulent paper to journal Y". I can just about guarantee you that none of them are imagining hallucinated citations.
That said even in a scientific context I still disagree and your example at the end is a fine starting point. By comparison imagine one of the profs said told the others that their house was burgled. The others would probably be thinking that things like TVs or computers or money was stolen, and not that the thief simply stole all their spoons. That doesn't make having all your spoons stolen not burglary. Likewise the profs expect that the results or authorships are where the fraud occurred because those are the best places to extract value with fraud, not by avoiding the simple act of writing the paper with correct citations. That doesn't mean fraudulently using an LLM to hallucinate a paper from your (we'll suppose for sake of argument) actual results is any less fraudulent though, it's just an unexpected form of fraud.
Edit: I want to be clear that this is not my argument: "well he exhibited reckless disregard for his professional duties when he opted not to bother reading the citation section". I see other people making that argument, and I'm not sure if they're right or wrong that that's another reason why it is fraud, but I'm certain that we don't even need to reach that question.
My argument is that it is fraud to represent the paper as a scholarly work when you don't know that it is correct. It is not that you are taking a risk it might be wrong, it is that you are actively representing that you know it is correct and if you do not know that you are committing fraud even if it happens to be so. This is a case of intentional deception, the deception being the representation that this is scholarly work, not reckless disregard for the truth as to the accuracy of the citations.
I concede that I was sloppy when I referred to what the researchers would be imagining. I should have phrased it as asking them if they thought that transgression X constituted fraud.
Regardless, hopefully you can see the idea that I was attempting to communicate? The burglary example isn't equivalent because while the spoons are unexpected the end result is still an event that most people would agree constituted burglary and resulted in noticeable harm to the victim.
I'm struggling to adjust your example on the fly but perhaps if it were the contents of the yard waste bin that had been pilfered? That's still technically burglary but I think most people would view it quite differently and might question the wisdom of prosecuting it.
I think the key difference here comes down to motivations as well as impact. Falsifying results (for example) is an active attempt to counterfeit the core value proposition of the endeavor and the end result of that is proportional - personal benefit directly as a result of the falsification and significant damages to anyone sufficiently bamboozled by the fiction long enough to base any decisions on it. Whereas no one using an LLM to generate just the bibliography is doing that to get ahead (at least not on its own) and any damages are limited to the reader wasting a few minutes trying to figure out the extent of the issue and who to contact about it.
I'm not sure you're right that they wouldn't think it's fraud at first glance, but I also think it's the wrong test. Words have meanings that result in an actual truth to the matter that is sometimes non-obvious.
A theorem isn't false just because it intuitively seems false until you spend decades trying to prove it and finally find the proof. Requiring people be cleanshaven isn't obviously discriminatory, until someone points out the existence and prevalence of pseudofolliculitis barbae and perhaps that an action being discriminatory doesn't actually require intent.
So I'd propose the correct test (and again, this would only be for a discussion in an academic context, which this is not) looks something like giving them access to the common dictionary definitions, and legal definitions, of fraud, and then asking them to define it. Then taking their definition and applying it to the facts at hand.
I don't think your average academic thinks academia has a special meaning of the word. I think they would simply acknowledge the dictionary definition as the correct one, and as a result I think this is plainly fraud even as the word is used in an academic context.
---
I don't think anyone defines fraud as only deception relating to the core value proposition of the instrument of the deception.
You didn't pick your co-author very well, but arXiv lacks investigative powers to determine which co-author did the bad, so they all get the consequence.
Again, I'm not in academia, but most of the papers I see have two to five authors; maybe I've seen one or two with ten authors.
Regardless, if I'm signing my name on something, I check it out.
A reference that doesn't exist is like a source release that doesn't compile. Any one of the 100 person team could have figured it out but no one did. In a 100 person team you get diffuse responsibility dynamics where checking citations is not assigned to specific people so no one does it. Or perhaps it's assigned to a single person who was also in charge of writing the citations and falsified them rather than doing the work.
You’re right that a single hallucinated line is not evidence of reckless disregard - because that could have happened on a final follow-up pass after you had performed due diligence. It’s happened to me. I know how challenging it can be to keep bad patterns out of LLM generated output, because human communication is full of bad patterns. It’s a constant battle, and sometimes I suspect that my hard-line posture actually encourages the LLM to regularly “vibe check” me! E.g. “Are you sure you’re really the guy you’re trying to be? Because if you are you wouldn’t miss this.” LLMs are devious, and that’s why I respect them so much. If you think they’re pumping the breaks then you should check again, because they probably just put the pedal to the metal.
That being said, I regularly insist on doing certain things myself. If I were publishing a paper intended to be taken seriously - citations would be one of the things I checked manually. But I can easily see myself doing a final follow-up pass after everything looks perfect, and missing a last minute change. I would hope that I would catch that, but when you’re approaching the finish line - that’s when you expect your team to come together. That’s when everything is “supposed to” fall into place. It’s the last place you would expect to be sabotaged, and in hindsight, probably the best place to be a saboteur.
It absolutely is.
> - because that could have happened on a final follow-up pass after you had performed due diligence.
A "final follow-up pass" that lets the LLM make whatever changes it deems appropriate completely negates all the due diligence you did before, unless you very carefully review the diffs. And a new or substantially changed citation should stand out in that diff so much that there's no possible excuse to missing it.
> It’s happened to me.
Then you were guilty of reckless disregard.
> I know how challenging it can be to keep bad patterns out of LLM generated output
If your research paper contains any LLM generated output you did not manually vet, you are a hack and should not get published.
You can only get in this situation if you let a bullshit generator write your paper, and the fraud is that you are generating bullshit and calling it a paper. No buts. It's impossible to trigger this accidentally, or without reckless disregard for the truth.
They're explicitly not writing papers. The fake citations are created and inserted by the LLM
If the bread is moldy in one spot, throw out the whole loaf, because the same invisible process that produced the mold spot is taking place throughout the loaf.
And flatly, if a person can't be bothered to check their damn work before uploading it, why should anyone else invest their time in reading it seriously?
The people I worry for are the junior researchers who are going to be splash damage for dishonest PIs. The PIs, though, deserve everything that’s coming for them.
However, we can have zero tolerance for certain techniques for "writing" a paper. Plagiarism and inventing data are already examples of this, if there is evidence for these techniques being used there is no excuse. We could say the same for AI references - any writing process that could produce these is by definition not a technique we want.
So the mistake isn't not checking a reference the AI gave. The mistake is letting the AI make references for you.
If we agree that academic research is important then I think we can impose certain standards on how you do it. We can dissalow certain tools if that means we can't trust the output. Just like an electrician can't use certain techniques, even if they're easy, because we don't trust the final result.
Responsibility really is diffuse. That's the nature of things. On a 100-author paper, 80 people are there because they worked on the upstream data source. 15 are there because they gave advice to the primary author. 3 are there because they wrote individual sections or appendices. 1 is there because they proof-read the paper and made general suggestions. 1 is there because they led the work and wrote the paper.
Now, do you actually think every one of those people is going to go through all 75 citations in the paper and carefully check that the bibliographic information is correct, and that every citation is correctly summarized? This is not a one-liner like checking if a project compiles. This is hours and hours (or more realistically, several days) of work.
You have an idealized view of how science works that is just not in line with reality.
I don't see anyone excusing hallucinated citations. I see people arguing that this is an overreaction by the arXiv. AI is going to become a major part of research, including in writing papers. It would be better if people would freak out a little less about that fact.