Just when you thought peer review couldn’t get any worse…
“We examined some 50,000 peer reviews for computer-science articles published in conference proceedings in 2023 and 2024. We estimate that 7–17% of the sentences in the reviews were written by LLMs on the basis of the writing style and the frequency at which certain words occur.”
That’s James Zou (Stanford), summarizing in Nature a study he and his colleagues conducted
What are such reviews like? Zou says:
Reviews penned by AI tools stand out because of their formal tone and verbosity—traits commonly associated with the writing style of large language models (LLMs). For example, words such as commendable and meticulous are now ten times more common in peer reviews than they were before 2022. AI-generated reviews also tend to be superficial and generalized, often don’t mention specific sections of the submitted paper and lack references.
Unsurprisingly, “the rate of LLM-generated text is higher in reviews that were submitted close to the deadline.”
I haven’t heard of any cases of this happening in philosophy yet. Have you?
Perhaps some of the tell-tale clues to LLM-authored referee reports will be different in philosophy (e.g., “This report was suspiciously free of any obnoxiousness”).
Zou thinks it’s inevitable that, despite their limitations, such tools will be part of the writing and reviewing process. He says:
Ultimately, the best way to prevent AI from dominating peer review might be to foster more human interactions during the process. Platforms such as OpenReview encourage reviewers and authors to have anonymized interactions, resolving questions through several rounds of discussion.
Perhaps encouraging the use of such platforms in philosophy would be worthwhile.
This was inevitable given how overloaded the review system already was.
OpenReview sounds nice and all, but it also sounds like it adds to the workload.
Unequivocally not. If people have too much on their plates, they can say ‘no’, and decline to accept reviews.* Instead, people are choosing to respond to review requests with content that does not reflect their own thoughts, views, or professional opinions, and doing so in a way that passes that off as their own work. This isn’t a consequence solely of an overburdened reviewer pool, but of choices some would-be-reviewers are making.
*In some contexts, reviews might be being delegated to graduate students or other subordinates, who might not feel like they’re in a position to decline the ‘request’ on the part of their superior/supervisor. But until we have data proving that this 7-17% comes inordinately from people in such subordinate positions, it doesn’t make sense to deflect any criticism of this practice by impugning an imagined population.
As this study is about computer science (CS), it might be of value to note that in CS undertaking reviews for conferences, it is sometimes obligatory for anyone who submits to also undertake reviews. Thus, anyone who wants to publish at these conferences might be considered forced to undertake reviews.
“If people have too much on their plates, they can say ‘no’, and decline to accept reviews.”
If you can come up with a better way for journal editors to find more reviewers, then you can start encouraging would-be reviewers to decline requests. I can sometimes find a reviewer in the first request or two I send, but sometimes it takes 5 or 10 requests before I get someone willing to accept.
Quibble: they’re being used for review, not peer review.
Good quibble. LLMs aren’t our “peers.”
Shameful self-promotion, but we argue for a “Sponsorship Model” of peer-review in a world where AI is developed enough to both publish and/or review papers: https://philpapers.org/rec/HADPRI
Do they have any way of identifying whether the LLM is reading the paper and writing the review, or whether the LLM might just be taking the bullet points the reviewer made while reading and turning those into a review?
Recently, a co-author ran one of our papers through an LLM asking it to “peer review it”, to see what we would get. It seemed to have a guess that a standard review would be “revise and resubmit” and said a few things that would appear to justify that. Surprisingly though, at least a couple of the comments it made about the paper were useful and we are incorporating them.
If the comments are constructive, then it would be fun to try to figure out whether LLM reviewers are better than most reviewer 2s.
Yeah I try to generate reports on all of my drafts now. I find that the quality is at least 30th percentile of all the reports I’ve gotten, since many human referee reports I get are actively harmful or totally useless, although the feedback is more like a smart non expert than a peer
I was curious and tried this with two drafts I have. It recommended acceptance with minor revisions for both, and it was nice to read it praise the merits of my papers. I then asked it to review a few mediocre term papers from students, and it recommended acceptance with minor revisions for those too. Well, at least *my* papers were evaluated correctly.
Yeah this seems like a promising use case to me: a real expert reads the paper and bullet-points the review, then the LLM makes it diplomatic and grammatical.
A lot of journals will request that you do not share a paper you’ve been sent to review. And even in the absence of such a request, giving LLMs access to a paper without its author’s permission is a dubious move.