“The key question is not whether AI is used, but how it is used.” That’s from Springer Nature’s policy on using AI when peer reviewing for its journals.
Wiley tells its journal referees: “Use AI responsibly: Apply best practices and properly disclose AI use in manuscripts and peer review.”
That’s right: two of the largest academic publishers allow referees to use AI when reviewing submissions to their journals.
Springer has created a “risk-assessment framework for AI use in peer review” to provide some further guidance:
Wiley provides fewer details, saying only that ” Manuscripts under review are confidential and must not be uploaded to AI Technologies, whether in full or in part.” But it seems have a workaround for this in development—at least that’s how I interpret them following a brief list of what editors and reviewers can “do safely” with AI with the announcement that they’re “developing new approaches that will expand these possibilities while maintaining the confidentiality expected in review. “
Oxford University Press says: “OUP selects peer reviewers for their expertise in the field and requires them to evaluate content based on their expert judgement alone. It is prohibited to upload project proposals and manuscripts, in part or in whole, into a Gen AI tool for any purpose.”
Meanwhile, the policy of University of Chicago Press on AI use by referees is more restrictive than the others above. It says:
In peer review, we require unassisted evaluations and expressly prohibit the use of AI tools in writing readers’ reports. Out of respect for the author’s intellectual property, peer reviewers should not copy any portion of a manuscript or a proposal, nor the reader’s report itself, into an AI tool, even one that promises not to retain data.
Some journals have their own policies regarding the use of AI by referees. Ergo, for instance, says:
We require our referees not to use AI in any form for reviewing of articles. In particular, referees must not enter any article text into AI systems, including AI checking tools such as Pangram. If referees have concerns about the use of AI in submitted articles, they should instead raise these concerns directly with the area editor.
But many journals lack such a policy, and don’t say anything explicit in their own guidelines or description of their peer review processes about the use of AI by referees. This includes some of the journals widely viewed as among the best or most prestigious in philosophy, from which you’re expecting to hear from fellow humans.
Such gaps—be they in policy or the communication of policy—should be fixed immediately.
Of course, one challenge with any policy that restricts the use of AI in this context is enforcement. Will editors have the capacity to catch reviewers who violate their journal’s or their publisher’s AI policies, or exploit vague policies?
And what happens when an author suspects their referee report was drafted by an AI? Many journals have a policy about the finality of editorial decisions. Mind, for instance, says: “The decision of the Editors on the acceptability of any manuscript is final. The Editors will not enter into correspondence with authors on any submission that is not accepted for publication.” Are reasonably suspicious authors just out of luck? Should they be?
So there are questions about what journal and publisher policies about the use of AI in peer review should be. There are questions about the enforceability of restrictive rules on AI use. There are questions about what recourse authors should have if they believe a referee report was written by an AI. And there are questions about what we should do in, and how we should be preparing for, the future rushing our way in which writing by AIs is indiscernible from writing by humans.
Here is ETHICS’ policy on this: “Guidelines for Editors and Reviewers
Submissions to Ethics are confidential. Thus, editors and reviewers must be careful to maintain confidentiality and protect the author’s intellectual property. Just as sharing a submission with a colleague would violate confidentiality, so too would uploading it (or any part of it) to external services that lack the access controls needed to preserve confidentiality. The editors of Ethics have agreed not to upload submissions to such external servers nor to use AI tools to assess them.
Those who agree to review for Ethics do so under the implicit understanding that they will provide the editors with their own independent assessment of the submission, an opinion that is based solely on their own reading of the manuscript. Thus, it would be inappropriate for a reviewer to base their assessment of a submission, even in part, on an AI’s summary of it. And it would certainly be inappropriate to use an AI tool to generate any part of their report or assessment.”
See here: https://www.journals.uchicago.edu/journals/et/ai-policy
Thanks, Doug. This is the way.
I think this is a sound policy.
AI & Society discourages use of AI for anything more than grammar and because we are an international journal, with many authors and readers for whom English is a second and/or foreign language, working from works sometimes published in languages which are not English, limited translation. I do increasingly often receive peer reviews which APPEAR to be (bearing characteristic marks of) AI generated. Some reviewers will acknowledge this use, in some cases for the entirety of the review, with statements such as ‘the reviewer approves of the evaluation’ or similar. IN such cases, I must advise the authors that the review is AI generated and must be discounted in favor of human reviewers, as we publish for human readers, works from human authors, not for machines by machines. I take it as the editor’s duty to safeguard a social contract between author and reviewer, to develop and to practice skills for scholarship, research and communication. Any use of AI in this process nullifies this contract, and is therefore not appropriate.
That’s a laudable editorial practice.
But AI & Society and a social contract between readers and journals: well, I turn to Jill;
https://jilltxt.net/misaligned-citations-and-llm-generated-scientific-fraud/
I have doubts about the ability of LLMs to produce genuinely original philosophy. I have no doubts about the ability of LLMs to scoop philosophers whose work has been submitted to those LLMs (with or without their consent). I have serious doubts about the trustworthiness of “AI” companies’ promises not to use submitted work as training data.
Giving an LLM a draft of a peer review could be enough to result in the manuscript’s ideas getting scooped.
If you don’t have time to write a review without help from an LLM, you don’t have time to write the review. If you need proofreading help, limit yourself to software that runs locally on your own computer. Do not use proofreading software that runs in “the cloud.”
“This includes some of the journals widely viewed as among the best or most prestigious in philosophy, from which you’re expecting to hear from fellow humans.”
Indeed.
Philosophers, especially junior philosophers, should know that enough of us in the profession have now had enough bad experiences with referees using AI for us to discount recent publications in some of these formerly top journals. Play it safe and submit your work to journals like Ethics with AI refereeing policies that don’t look anything like Wiley’s and Springer’s.
Paul Bloom is the co-editor of Behavioral and Brain Sciences. He talked about this on his Substack, and he also just wrote an article for CHE. Here are the links:
https://substack.com/@smallpotatoes/note/c-339294574
https://www.chronicle.com/article/could-ai-be-better-at-peer-review-than-humans
Readers of DN might find what he says about using AI to referee of interest.
I’m not particularly exercised about intellectual property.
What concerns me, rather, is this: I have actively chosen not to use chatbots to “improve” my work, or to do it for me. Whatever you might think of that choice, it is entirely mine to make, and it matters a great deal to me. When you use a chatbot to “referee” my work, not only are you outsourcing the work, you are actively undermining the decision I made about my own work. Worse, you are hiding that fact from me, meaning that unless I guess what you are doing, I cannot even reject it.
That looks like a much worse consent violation to me.
For the reasons I give downthread, I’m very skeptical about AI reviewing in peer review, but I’m not sure I see the general point about consent violation (I do accept the point about disclosure). No-one is required to make any changes to their article in response to reviewer criticism; it’s just that they won’t get the article published if they don’t. I don’t “consent” to changing my article to conform to the mistaken views of reviewers who are confused on basic issues of physics, but that doesn’t convey on me a right to force the journal to find new reviewers.
Some thoughts:
1) Most AI systems licensed to universities will have enterprise-grade zero-data-retention policies, and most private packages can have ZDR activated if you want to, so I’m not hugely moved by concerns about data privacy. Yes, the AI companies could be lying, but so could Dropbox or the people who designed the OS on your computer. At some point, you either accept that legal and commercial pressures mean massive tech companies probably aren’t falsifying their data-privacy policies, or you go back to mailing photocopies of your article to the editor’s university address. (I do accept that there is a human-error concern in principle.)
2) There is a much better reason that reviewers shouldn’t feed the article they’re reviewing to AI, which is that if the editors wanted to know what Claude thought of the paper, they could ask Claude themselves. Presumably they want to know what *you* thought of the paper. (And the author could also have asked Claude, so AI-written reviewer comments aren’t useful to them either.) Maybe (but maybe not) editors should get an AI report at a preliminary assessment stage, but there’s no case for the reviewer doing it.
3) There’s more of a case for reviewers using AI to improve their own report (where the AI doesn’t get to see the paper, only the review). My own feelings here are broadly similar to my feelings about AI use to improve writing of papers: fine for copyediting and basic spelling/grammar; ok to take advice about structure and logic, provided you assess and implement that advice yourself; understandable to get advice on style, especially if you’re ESL, but I wouldn’t do it myself and I recommend extreme caution; absolutely don’t just give it a set of bullet points and get it to bulk it out into a review. (If you’re going to do that, just submit the bullet points as your review.)
4) I don’t have a principled objection to a reviewer writing a review and then giving the AI the review and the paper and asking what it thinks, but in practice I think a bright line is important, so journals should probably forbid that.
5) If you are an author and have access to a reasonably recent AI model, you are missing a trick if you don’t get its feedback before submission. Even if you restrict yourself to copyediting, it’s extremely helpful, and it often has good thoughts about structure. That doesn’t mean you need to accept all its advice, of course, but that’s also true for human advice. (But I don’t recommend actually letting it implement changes itself -it can get carried away, and you don’t want your paper to show up as AI-written in Pangram if it isn’t.)
FWIW 2-4 basically comprise Philosophy of Physics’ AI review policy, though we’re still finalizing the wording.
This is all really good. And also, anyone who hasn’t tried it, should try 5, with one of the frontier models. If you don’t have institutional access or a $20 subscription to either Claude or ChatGPT, ask someone who does to run your paper through Opus 5.5 or Astra 6 to ask if any of the points made here are anticipated in other literature that wasn’t cited, or if there are ways some of the presentation could be improved, or anything else. You’ll probably reject 70+% of what it says, but some of the points it makes are likely to make you think of something that can improve your paper.
Before running their paper through an AI service, authors might want to check the AI policies of the journal(s) to which they plan to submit their paper. Some (like the Journal of the History of Philosophy) prohibit any use of AI in submitted papers.
JHP’s policy just says that papers can’t be the product of generative AI, in whole or part. (He says having just looked it up, since I don’t publish in JHP). I would not naturally have read that as ruling out AI proofreading and copyediting (I wouldn’t think that my paper is a product of any human copyeditor I use, or of the friend I ask to read it and give feedback, and while I don’t actually use human copyeditors, I gather it’s quite common for ESL authors to do so).
If what you (as JHP editor) intend is more like “You should not use AI in any way during the writing of this paper” then you might want to clarify it. You might also want to give your authors a bit of advice as to how to follow it, too, as it’s actually quite hard in 2026 – it means not using Google Scholar or most other mainstream search tools, which are now largely AI-powered, and it also means disabling MS Word’s default spelling and grammar options, which use AI.
My understanding (but perhaps I am wrong) is that the grammar- and spell-checkers integrated in Word do not use generative AI, which is what the JHP policy prohibits. But you make a good point about Google Scholar and other search tools using it. Our policy for authors probably needs to be updated to be clearer about this.
You’re probably right about MS Word, having done slightly more homework.
Come on, David. Dropbox isn’t being sued for copyright infringement.
I haven’t made a close study of the copyright-infringement cases, but my understanding is that they’re mostly about the legal status of something where the facts are broadly agreed upon, rather than about hiding activity that everyone agrees is illegal.
But also: Dropbox has AI threaded through it these days via Dropbox Dash, so you’re relying on their promise that they don’t train their AI systems on your data. And of course if you use Onedrive or Google Drive as well as / instead of Dropbox, you’re using products from companies that are very involved in AI training. Everything is far too integrated for “trust data privacy policies from the rest of the tech sector, but not from AI companies” to really be a coherent dividing line.
And if evidence of bad behavior from Dropbox comes to light, I’ll trust them less than I do at the moment. Many of us have used Dropbox for well over a decade and I don’t recall them being in the news for copyright infringement, acquiring rare books and destroying them, hacking other companies, or operating rather unscrupulously more generally.
The relevant principle is more like “continue to trust companies who have spent the past 20 or so years earning it, but not _those_ new companies like xAI, OpenAI, and Anthropic over there who have been in the news a lot lately for seemingly bad behavior”.
Let’s stipulate that Dropbox genuinely has no ties to AI companies. What about using Gmail or OneDrive? I don’t see how one could consistently (e.g.) abjure using Gemini but be happy to use Gmail.
There’s plenty of great products outside the Google-verse. Replace Gmail with FastMail. You won’t regret it.
Well, sure, but: “do not use any tech product outside a narrow whitelist that excludes all the major providers” is a lot more radical and difficult than “do not upload to ChatGPT or Claude.”
I suppose one other thing I should have said is: even if there are permissible ways for a reviewer to use AI to improve their review without seeing the paper itself (copyediting, feedback on structure), I really don’t think reviewers owe that level of care in the first place. It’s hard enough to find them as it is. Reviewers: read the paper, think about it, write down your thoughts, hit Send. The editors and the author can cope with your typos!
Just on point 2. I don’t really think that editors are asking for reviewers to know what *you* think of it in a way that excludes what you think of it being influenced by critical examination of what Claude thinks of it. Maybe there are confidentiality worries against doing so, but I don’t think that as an epistemic matter it would be incompatible with what the editor wants or should want, for me to mention the paper to a colleague and change my report if they think of something I didn’t. Even if I ought not to do that, I doubt that an editor would have an objection to my incorporating my colleague’s input once I had; more likely they would think it was weird if I had changed my mind based on what my colleague said, but decided to submit my original review instead of my revised one which I endorse. We know from this thread and countless ones like it that many people including many editors *won’t* ask Claude for feedback, for a lot of reasons, many bad and some understandable. So the fact that they *could* get Claude’s input, doesn’t seem to mean that I shouldn’t, provided I see my role as a referee to give the best assessment of the paper, which I do.
The bigger worry is that I will upload the manuscript to Claude before forming my own opinion on it, since it would be very hard to resist being primed by Claude’s response to it, and then my expertise is undermined. That would be a huge derogation and should be avoided for pragmatic reasons.
That’s my (4): no objection in principle but it is probably sensible to draw a bright line.
Sorry you’re right, not sure how i missed that but i did
I really appreciate this discussion.