Handling cases of "potential" ChatGPT-generated reviews in non-anonymous program committees (as a PC member)

Question

I am currently serving on the program committee (PC) of a top-tier venue in computer science. Every paper is reviewed by 3-4 PC members; then, the authors can then see the reviews and respond with a "rebuttal." All the PC members who have reviewed a certain paper know the identities of the other PC members who reviewed the same paper. The reviewing process is double blind (authors do not know the identities of the reviewers, and reviewers do not know the identities of the authors).

The issue. I have just submitted my review for one of the papers in my stack. Upon reading the other reviews, I found one review is written in a "ChatGPT-like" style. Suffice it to say that it is not "obvious", but it's very likely that this is the case.

As a PC member -- but also as an author of a paper that was reviewed by ChatGPT -- I would like to "do something." However, I do not see anything that I can do that would lead to a happy ending. This is because I am unable to ethically prove¹ that the review has been generated with ChatGPT.

As such, I am stuck. I have several options, but all of these do not lead to a happy ending:

• I can contact the PC chairs. However, I cannot provide any proof, aside from mentioning the elements that make me believe that the review was generated with ChatGPT. Of course, lack of proof would also lead them to do nothing (e.g., they cannot "point the finger" to the reviewer without some solid basis).

• I can ask the PC member who wrote the "suspicious" review. This is unlikely to end well, since they will most likely be defensive (if they truly used ChatGPT) or feel offended (if they did not).

• I can acknowledge that I cannot prove anything, and try to induce the PC member who wrote such a review to amend the review by, e.g., pointing out specific elements in their review (which I think are vague and not informative). This may take a long time though, and may very well end with no change whatsoever.

• I can just do nothing. And this would also leave a sour taste in my mouth: if the review was truly written by ChatGPT, then the authors would receive an "unfair" review, which would be quite disappointing.

If I knew the identity of the authors, I could ask for their permission to carry out the analysis myself. However, I cannot do so since the reviewing process is double blind.

Do you have any recommendation or alternative solution?

---

1: As I mentioned, I have been on the receiving end of a paper that (i) was rejected and (ii) for which one of the reviewers wrote their review with ChatGPT. After reading the reviews, I became suspicious and, since I was the author of the paper, I went to www.chatpdf.com, submitted the paper to it, and tried a few prompts until I found the one that produced (almost) the same review I received. However, I cannot do this "ethically" for this case: if I were to upload the paper on www.chatpdf.com, I would breach the confidentiality agreement since I would give a paper (currently under peer-review) to a third-party that is unrelated to the venue; potentially, I may compromise the novelty of the finding discussed in the paper, damaging the authors. Therefore, I am unable to prove that the review was generated with ChatGPT or similar LLM. [Somewhat related: Can I improve a paper peer-review using ChatGPT?]

Answer 1

I can contact the PC chairs. However, I cannot provide any proof, aside from mentioning the elements that make me believe that the review was generated with ChatGPT. Of course, lack of proof would also lead them to do nothing (e.g., they cannot "point the finger" to the reviewer without some solid basis).

I think this is the best approach. You can provide the elements that give you that belief, you can explain your inability to be certain, and then you can let them decide. I would not push it much further, though; once you've brought your suspicions to light your job is done.

As far as evidence, the chairs may have other evidence to bring to the table like other reviews by the same person that you are not party to.

As far as proof, this isn't a court of law, proof is not necessary; they do not need to pillory the accused in the town square, they could instead merely choose to discount their opinion in this one specific case. If sufficiently miffed, they could choose to not invite this reviewer to participate in the future. They can potentially do either of these things without making any accusation known, and they certainly don't need to prove involvement of any specific LLM, they must merely be dissatisfied with the quality of review.

Answer 2

At the very least, you should report this to the chair because the authors most likely did not agree to have their abstract harvested by an LLM, and probably don't want it to be.

Answer 3

Is only the style reminiscent of ChatGPT or is the content of the review also poor? These are different questions. For instance, a reviewer not very proficient in English could have used an LLM to improve their writing, and that may give the style of ChatGPT but human content, or the reviewer could have discussed their technical points with a machine and taken on board some in their own judgement good points the machine made, or conversely humans are entirely capable of writing very poor/misguided reviews all on their own.

Depending on the confidentiality policy of your conference, it might be still ethical to try to reproduce the review on ChatGPT using their "temporary chat" feature (which wipes the discussion from the chat history, limits storage on OpenAI servers to one month, and excludes the content from training on user data). You could also try to see if a similar review can be reproduced by one of the more powerful open-weights models by installing a copy of, say, Llama-3-70b on a sufficiently powerful server that is fully under your control and feeding that model the text. It's easy to do, if you have the required server. But I think either method will still likely give you inconclusive data. It is easy with such things to see patterns where none are there (Google Toiletgate to get a taste for how low this can go).

I would simply strongly back up the authors if they contest any factually wrong claims made in the review, and/or use reviewer discussion unprompted to contest points that I think are weak/wrong, and try to convince the other reviewers/program chairs to discount the review if it is sufficiently poor on purely technical terms. Given such feedback, if the author of the review used ChatGPT, I would hope that they learn from the experience and produce better reviews next time they are asked.

Answer 4

Since the program committee has some control over who the reviewers are, the alternate solution is to have a policy about AI generated reviews and to get reviewers to agree to abide by those policies. This should, at least, reduce the problem if it exists.

The policy should be clear and give reasons for the rules.

One possible feature of such a policy could be that the reviewer attest on submitting the review that no part of it was AI generated (or something similar).

In the short term, however, the committee should be made aware of any suspected AI generated reviews so that they can evaluate the problem.

Authors need input on their submissions and the input needs to be done by a mind with an ethical sense behind it. If it could be automated then there would be no need for reviewers. Reviewers need clear rules about many things including AI generation. Without that combination, the quality of presented and published work can and probably will decline.

And, this isn't limited to CS.

Answer 5

I would ignore whether the review was written with the assistance of a LLM. It isn't really important and it is hard to prove (And is there a point to proving this? Is there an explicit policy against this?) What you really have a problem with is that the review is of low quality. The same problem can easily occur without LLM assitance. Is there a procedure for when you notice a review of low quality? If so follow that. In case you can still edit your review perhaps you can discuss the points mentioned in the other review and rebut them? It also seems like you will be in the position to point out that this quality is low quality and should be ignored later on in the process, or are you not? If not you should write someone who is.

Answer 6

You should do absolutely nothing.

The premise makes an implication that is not shown to be true. More specifically that there's something wrong with using tools to write reviews. Is that explicitly forbidden (although it shouldn't be in my opinion)? If yes, address it to whoever issued the policy. If it's not explicitly forbidden, then it's allowed. The reviewer is nevertheless responsible for the reviews they sign. If they used ChatGPT and then agreed with what it says there, then it's all good. If what it says doesn't make sense, then they're responsible for what they signed nevertheless.

If they're plagiarizing, or if it's a low quality review, address that, but not that they used a glorified spell checker to improve their writing efficiency.

In the comments to this question it has been raised the issue of possibly breaking a confidentiality agreement, as some GPT tools may not be associated with the program and may involve submitting the data to a 3rd party. This doesn't necessarily seem to be the raised issue, but I find this to be a legitimate concern. In this case, the issue should be addressed directly and no proof is necessary. It should be as simple as "Hey, I noticed a certain style of writing that made me think that you may have used a tool that may require submitting the article to a 3rd party, which may be in breach of our confidentiality agreement. Please double check that this isn't the case, and if it is, please let us know so that we can take the further appropriate steps."