Leiter Reports: A Philosophy Blog

News and views about philosophy, the academic profession, academic freedom, intellectual culture, and other topics. The world’s most popular philosophy blog, since 2003.

  1. Brian Kogelmann's avatar
  2. RMD's avatar
  3. Textbook Author's avatar
  4. New Author's avatar
  5. J. McKenzie Alexander's avatar
  6. Terence Rajivan Edward's avatar
  7. JR Williams's avatar

A sensible policy about refereeing in the age of AI

Brian Leiter

Philosopher J.P. Messina (Purdue) kindly gave permission to share what he posted on FB recently, and it seems to me his policy is a good one:

I have refereed two papers this week that were pretty clearly written or heavily revised by LLMs. New policy: unless editors have an AI-use policy which requires AI-use disclosure and checks for obvious signs of non-disclosed use, and unless author’s disclosure is shared with reviewers, I will not be accepting their referee requests. For context, in each of the last three years, I’ve refereed 20-30 papers for a variety of journals. But if your journal can’t be bothered with doing basic work to check for academic integrity, I can’t be bothered with ensuring that work submitted to it meets whatever ‘standards’ remain. And for “authors”: except in rare cases, if you don’t have time to write it, I don’t have to time to read it.

What do readers (and journal editors) think? Are there other approaches that referees should consider?

,

Leave a Reply

Your email address will not be published. Required fields are marked *

8 responses to “A sensible policy about refereeing in the age of AI”

  1. Michel-Antoine Xhignesse

    I haven’t encountered this yet, as far as I can tell (though I did recently get a paper that was wholly plagiarized from a book chapter). Until I do, I’m still happy to just accept referee requests when they come in.

    That said, I’m entirely on board with the spirit of the thing. If you didn’t write it, I don’t want to read it. Moreover, if I start getting too many chatbot papers to referee at any one journal, I will simply stop refereeing for them altogether. This is not a big loss to any given journal on its own, but it would, in those circumstances, be a real boon to me.

    I would probably be more compromising if I wasn’t already deluged with chatbot slop in all other areas of my life, from emails to student work to social media feeds. But since that’s the way the world is, my tolerance is zero.

    Reply
  2. The Pacific

    I wonder how feasible and time-intensive it would be to make it a standard policy check for undisclosed AI usage. Blatant cases, ok. Borderline cases, it is unclear whether it is worth time of a back-and-forth over uncertain evidence. In principle, why not publish all submissions that meet the desired standard and space availability of a journal, and let others in the system handle the apportioning of credit to those who deserve it? The result would (ideally) be that the best papers are accepted. Someone who uses AI will still face significant or perhaps even greater consequences (e.g., retractions or failures to be hired) down the line if it turns out that their peers or interviewers do not agree with they way they used AI. Discerning whether an author has crossed the line of acceptable AI usage (whatever that may be) might fall to hiring and tenure committees and external reviewers—who at any rate might have the benefit of additional evidence (more papers, recommendations, interviews, and AI disclosures). This is merely a suggestion of a general attitude one might have in lieu of trying to police undisclosed AI use—not an argument against policing it. Whether it is technically feasible is a question I don’t know the answer to and I would be interested in hearing opinions on.

    Reply
    1. Messina

      @ The Pacific: Good questions.

      My point isn’t that journals can’t adopt that stance, only that I won’t referee for them if they do, which seems to be my prerogative, as the setting of editorial policies is the prerogative of editors. It is however worth noting that the costs of the discernment for unacceptable use depend on the editors’ policies about acceptable AI usage. It’s possible to say: “we don’t care at all; we judge submitted work by its quality, not its provenance.” (Of course, that would also (I take it) rule in high quality plagiarism, which would be quite a departure from the academic status quo.) In that case, since there is no unacceptable use, there is no cost of discernment. Referees can then choose to participate or not. (I would bow out.) And, in general, the looser the standard for acceptable use, the lower the costs for editors. (That seems intuitively, true, anyway – but worth thinking about whether it is.)

      As for the costs themselves: substack embeds pangram; I don’t see why editors couldn’t do the same*. “But, false positives, and adjudication costs when scholars disavow usage!” Well, just require authors to certify that they didn’t use it (in unacceptable ways) – i.e., force them t disclose, and trust them (unless their claim doesn’t pass the sniff test). The entire enterprise depends on a measure trust. On the other hand, if it doesn’t pass the sniff test, adopt an attitude toward false positives vs. false negatives, and remember that there is no right to publication in any particular venue. The entire enterprise depends on a measure of editorial discretion. “But pangram is expensive!” Most editors have pretty generous research accounts and many will already pay for a subscription; some can exert pressure on their publishing hosts. The entire enterprise depends on the wise deployment of resources.

      Those anyway are my thoughts for now… curious what you think.

      * = or embed the best cost-effective checker.

      Reply
      1. David Wallace

        Pangram costs $20/month for 300,000 words / month, so a small journal receiving 20-30 10,000 word articles per month would be paying about $1/article, which is pretty minor compared to the costs of typesetting and copyediting accepted articles. I am confident its false-positive rate on long-form academic writing is negligible – I actually don’t know a single example of an academic paper from before 2022 that has given a positive result on Pangram. What is less clear is how good Pangram is in distinguishing human-written text that has been polished by AI, from fully AI-written text. The recent versions claim to do this but I don’t know any good third-party evidence on the matter.

        Reply
        1. Steve Innsbruck

          “What is less clear is how good Pangram is in distinguishing human-written text that has been polished by AI, from fully AI-written text.”

          Doesn’t that seem like a significant problem? Especially since the latter doesn’t seem like something we should obviously want to prohibit or punish?

          I’ll be blunt. I don’t trust Pangram. At least, I don’t trust it as much as some commenters on this topic seem to (despite their also often hedging their claims with caveats like the one above about polishing). Some of the discussions I’ve seen about it in academic and other circles, and some of the evidence offered against its reliability, have made me wonder why so many academics seem to be so eager to adopt it as a filter. Add to it that one can apparently revise AI-generated text so that it doesn’t get marked as AI when uploaded into Pangram, and I’m not sure how worthwhile a tool it is compared to just reading a paper ourselves and making our own judgments about the paper’s quality. If one can, as some have suggested, later revise a fully AI-generated text so that it passes Pangram’s test as human, then Pangram isn’t actually a reliable filter for preventing AI-generated papers; it effectively just blocks or punishes anyone whose text might get marked as AI by Pangram, whether human or AI-generated or human written but AI polished.

          If I read a philosophy paper and I can’t tell it’s AI and I think the writing is good (or at least as good as some of the real philosophers whose writing is NOT so good), or I suspect it’s polished by AI to some extent or some part was but the paper is otherwise good and advances unique worthwhile arguments and ideas, I’m not sure why I’d reject or dismiss it just because of some Pangram score. I can see myself suggesting the author revise the writing in those seemingly AI-influenced sections (much like I’ve suggested authors whose writing was wholly human revise parts of their work because their writing style lacked much to be desired). However, to toss out a whole paper or look down on it as bunk that shouldn’t be seriously considered by a referee because part of it seems to be AI influenced or it seems to be polished by AI or it has a Pangram score of some threshold seems, I think, to go too far. I’m not saying we should all praise and embrace AI use, but I’m also not convinced AI use, or various versions or degrees of it, deserves the level of hostility and dismissiveness that’s been directed at it. I would be fine, however, not refereeing a paper that seems completely AI-generated and clearly so. The author should at least put in some effort to try to make the paper readable and enjoyable in its prose. Still, in that case I’d be dismissing the paper more so because the writing quality is bad, much like I might dismiss any poorly written “fully human” text.

        2. David Wallace

          Replying to Steve Innsbruck (just on the issue of Pangram reliability):
          “Some of the discussions I’ve seen about it in academic and other circles, and some of the evidence offered against its reliability, have made me wonder why so many academics seem to be so eager to adopt it as a filter.”

          Speaking personally, it’s because I have *not* seen any evidence against its reliability on long-form academic writing, despite looking quite hard, and I have seen a lot of evidence in favor of its reliability. If you have evidence to the contrary, please share links. What I have seen is that it is pretty reliable at avoiding false negatives and *extremely* reliable at avoiding false positives.

          What I have seen about its ability to distinguish AI-tweaked text from AI-written text has also been very impressive – I am less confident here just because the systematic evidence base looks weaker.

          I am also more skeptical than you are of the virtues of “AI polishing”, which tends to homogenize text towards the particular AI’s style and which changes meaning more often than people sometimes recognize.

        3. Steve Innsbruck

          Replying to DW’s reply to me:

          Let’s grant Pangram is reliable in the standard cases (I still don’t fully trust this to be true). That people can apparently have a paper written fully by AI and then later revise it so it passes the Pangram test makes me feel like a lot of this hope for policing AI use by implementing Pangram checks is just ridiculous.

          It seems like much of the discussion has been caught up on the Pangram reliability question. Others seem both convinced AI use is bad full stop and that Pangram is reliable, and so are now pushing Pangram as the obvious solution. That all seems to be jumping the gun to me. The more important question to answer before pushing so hard to implement a policing policy or method is what forms of AI use are acceptable or not. Again, I’m now leaning toward thinking maybe even most uses of AI are fine for the reasons I’ve already mentioned. If the paper is generally good (good arguments and ideas), reads well, and so on, I’m not sure why I should care if it gets scored on Pangram. Even if it’s all those things and passes the Pangram test, I still have no good way of knowing if it was wholly or partly written by AI first and then merely revised to pass the Pangram test later.

        4. David Wallace

          Conversations about Pangram frequently involve people who say they’ve seen evidence it doesn’t work, but can’t produce it, so I’ll ask again: please can you provide links to any evidence that Pangram is unreliable on long-form academic text? (e.g., evidence that a fully AI-written academic paper can be human-rewritten to pass Pangram.) I’m quite prepared to accept it exists, but I would need to see it.

Designed with WordPress