"…there is no way reviewers and journals are going to be able to keep up. Reviewers are typically unpaid academics who are already stretched to their limits; tripling their workload would not be feasible. […] the total number of articles may radically spike, many of them dubious and a waste of reviewers’ time. Lots of bad stuff is going to sneak in."
In the end, it doesn't matter who or what makes a scholarly article unreliable. It's all bad.
It may well be that an article of non-native speakers who have used LLMs for their introduction is more reliable than the article of native English speaker who have faked their data or whose sample size was too low or p-hacked their data. Even without LLMs, humans are perfectly capable of flooding the literature with unreliable articles:
@brembs@ivan_herman Sure, but the amount matters. If people can all of a sudden produce ten times more shitty papers using LLMs, that's a problem in and of itself.
Or maybe I missed your point?
If you look at the estimation I shared, it seems as if one actually could increase the number of automatically generated papers ten times and still be far below the number of unreliable papers humans write.
@brembs@ivan_herman Okay, got you! Don't you think those people who tend to publish unreliable papers will use LLMs to create even more unreliable papers? I feel like you're making a too simple dichotomy between paper mills (with LLMs) and researchers creating unreliable papers.
Just like a person with a shovel can dig only so many holes but someone with a digger can dig many, many more holes
Very valid argument.
Part of the point is that the vast majority of authors (as far as we know) writes unreliable articles not because they actively cheat. So if that is correct, there would be no reason to assume they now all start using LLMs.
That being said, eventually, all else remaining equal, LLM-papers will outnumber all other papers. All I'm saying is that even a factor of ten would still be less than what we already have to deal with.
In other words: people who are shocked at the scale at which unreliable papers get published via LLMs maybe haven't fully internalized the scope of he problem we already had before the advent of LLMs?
@brembs@matthias_aulbach@ivan_herman That's exactly my response to it. Did people at some point lose all sense of scientific scepticism? Back as an undergraduate it was 100% the default assumption you didn't automatically just... believe... stuff you read in a paper.
@TEG@brembs@ivan_herman I don't know who you mean with "people" but ever since first hearing about the replication crisis, my standard reading of any paper is similar to reading an opinion piece. Not even that people might have cheated or done things wrong but also sampling error and tons of other things just make results of a single paper way less relevant than I'd like them to be
@matthias_aulbach@brembs@ivan_herman My version of "people" here - it's the subset of researchers who seem so shocked about the unreliability of individual papers, while it was always like this.
It's maybe like there's now a stronger tendency to treat a paper as an isolated separate object. As opposed to always considering the professional meta-information as a matter of course; e.g. you'd ideally know who wrote the paper and have met them or have some connection with them, even if indirect.
@matthias_aulbach@brembs@ivan_herman So I wonder whether there's an element of "death of subject matter expertise" going on. Like a random person should expect to be able to pick up a paper and acquire knowledge, without being part of a whole complex field of relationships and caveats. I suspect that trying to force that reality into existence might not be the best way forward, as opposed to fixing the ecosystem.
Add comment