I'm not familiar with the world of academic publishing, so I want to ask: how is the industry making sure that submissions aren't at least partially AI-generated?
Is it standard practice for authors to have to defend their submissions via interview like this? If not, why not?
Does the vetting process vary with the quality of the publisher?
As an outsider, it's extremely worrying that anyone would even attempt to submit an AI-generated paper for publication in an academic journal. At that level I would have assumed literally everybody should know better than to even try.
> Is it standard practice for authors to have to defend their submissions via interview like this? If not, why not?
It isn't, but maybe it should be. For post-grad qualifications oral defense is standard, and I didn't mind defending my central thesis then, and won't mind now.
Not in a direct interview style, but most us conferences can request additional information or feedback. If they conditionally accept or reject a paper, that conditional relies on feedback from the author(s).
Interviews like this are interesting, but in no way can scale to the infinite paper slop conferences are facing.
> Not in a direct interview style, but most us conferences can request additional information or feedback. If they conditionally accept or reject a paper, that conditional relies on feedback from the author(s).
Requesting feedback is useless, as the article points out - the "authors" could not answer basic questions during the interview, but after the interview were able to send full explanations to the interviewer.
If you have indirect feedback ("please answer these questions we have") the "author" will simply feed it into an LLM and send the results back. You need to get the author to do an oral defense to verify that they wrote the paper.
This is the main problem with AI generated output, whether it's a research paper, a blog, an email, a comment on a forum, similar: the value in knowing that a human wrote $X sends a signal - that the human understands what it is they wrote, even if they misunderstand the concepts.
When you get a message from someone who is a "I only used an LLM to clean it up, the thoughts are all mine"[1] person, you cannot engage with them, because they may not understand the message they transmitted, and so any human engaging with them is only burning their own time for no gain.
When you get a message from a real person, you get not only the message, you also get a signal about their understanding. That signal is missing in AI generated messages.
==========================
[1] Sure, buddy. We believe you /s.
“Requesting feedback is useless, as the article points out”, I would agree post llm availability. Before that (~pre 2023 I suppose), this wouldn’t have been as large of a concern.
“You need to get the author to do an oral defense to verify that they wrote the paper.” While I agree with this point, it can’t scale. As with the indirect response, vishing and likely video mimicry are only going to be easier with the next llm release. This leaves requiring each author who submitted a paper to fly to a place just to defend a paper in person (assuming all volunteer reviewers are in the same place).
> “You need to get the author to do an oral defense to verify that they wrote the paper.” While I agree with this point, it can’t scale.
Are we prepared to sacrifice everything on the altar of scalability?
At any rate, why can't this scale? Go to the nearest accredited university, sit at a terminal they give you and do you interview.
Universities already have the space to do this, for potentially hundreds of defenders per day.
“Are we prepared to sacrifice everything on the altar of scalability?” If the performance of verification isn’t great, it won’t get adopted. Just general acceptance timelines, look at how long it took for Python 2->3, https 1.2, us real-ids. You can’t adopt a significant change immediately and expect things to work immediately.
“Universities already have the space to do this, for potentially hundreds of defenders per day.” Universities have space for their own students while they’re still funded for the time, but conferences aren’t tied to universities. Even so if they were, from the author “The fraction of desk rejected papers at TMLR used to be about 6% in 2023 but is now at about 53%.” universities would have to raise funds for this massive increase of slop review. Us conferences uses unpaid volunteers (most part) from those areas for reviews.
If an independent conference were to hold in-person interviews (and maybe fly people out) for 1000s of submissions, where would they hold it and how would they pay for it?
> I only used an LLM to clean it up, the thoughts are all mine"
I mean, it seems to me like that should be nearly everybody at this point, right? There are at least some elements of using LLMs that are the equivalent of a spell check, like asking to make sure that links work and that references are to the thing they're supposed to be, and so on. I feel like the only difference at this point is between people who do that and say that they do, and people who do that and don't say that they do.
> I feel like the only difference at this point is between people who do that and say that they do, and people who do that and don't say that they do.
No. Did you "clean this up" with an LLM before posting it?
I hate to go there, but how is that argument different from a thief who argues "Everyone steals. The only difference is some of them claim they do, and some claim they don't"?
> I mean, it seems to me like that should be nearly everybody at this point, right? There are at least some elements of using LLMs that are the equivalent of a spell check, like asking to make sure that links work and that references are to the thing they're supposed to be, and so on.
Right. Maybe. The problem is that when material has all the tells of LLM generation, do you expect the reader to figure out if the thoughts were the authors or introduced by the LLM?
The value in a human directly communicating with you is that you understand what they are saying; if they misunderstand, you spot their misunderstanding. If they are talking at cross-purposes, you know you are arguing past each other. If they agree with you, you know they agree with you.
That signal is missing when the communication is LLM generated; you message said "Not X, Not Y, just Z", but I can't tell if you even know what X, Y and Z are. f I ask you to give me a definition for them so I can ensure we aren't talking about different things, the LLM response will be "X is $SOMETHING", but I still can't tell if you think that X is $SOMETHING or if you still think that X is $SOMETHING_ELSE while your LLM and I agree that X is $SOMETHING.
Communication from a person tells us something about that person, outside of the message being communicated. If that communication is relayed via an non-invested third party, the missing signal prevents any communication.
You've nailed my thoughts on that aspect pretty well. You demonstrate why when we choose communicate with a person, we definitely want a person at the other end of the line.
But also: When we want to ask a question of a bot and then get an answer from that bot, then we'd have already made a deliberate and rational decision to ask a bot ourselves.
We not need nor want anyone's help on that front; this kind of low-effort help results in an insulting waste of time.
Abject silence would be an improvement in communication quality over having an unwanted third-party conversation with a bot.
I’ve only published a few papers, but this interview sounds extremely unusual to me (I mean, it is clearly a special thing that the editor is doing, which is fine). I wouldn’t do something unethical, but if I had and the editor asked me for an interview like this, I’d know I’d probably been caught.
Why is it unusual: it sounds extremely time-consuming.
As to how worrying AI-generated papers are… it sounds more like a headache for the editors really.
In general, journals don’t have to be perfect; mostly researchers read research papers. You already have to read critically (publish-or-perish has been a thing for a while, so there are plenty of not-so-great papers out there). Peer review is just the “entry” barrier, science is a social process and papers become more or less influential based on a fuzzy process of citation, conference talks, and peer-to-peer suggestions.
> it sounds extremely time-consuming.
For whom? Surely the authors can find an hour after submitting the paper to a journal?
Beware that a reviewer easily spend a full week on reviewing a paper, and there are typically three of them. So if one hour of conversation can save three weeks work, it sounds worth it.
I was thinking for the reviewer. Also note that I was only answering as to why it wasn’t done in the past.
For the time comparison, I’m not sure, it doesn’t seem quite apples-to-apples:
1) It is scheduled time vs unscheduled.
2) There must be some base rate consideration… the papers being discussed here were planned to be desk-rejected.
Maybe it could be a good process for saving papers that were going to be desk-rejected, but I dunno, that seems like it’d just lower the quality standards.
But it's not the reviewers who do it, it's the handling editor. Who spends relatively little time per paper (compared to authors and reviewers).
Honestly, I was sloppily thinking of the general reviewer/editor system as one thing (and didn’t notice that I’d lumped them together until you brought it up here). Good point. Although the editor has a lot more papers to handle, so I think they are probably busy as well.
The rates of paper publishing show that everyone must be using ai now or the rates wouldn't have gone up