Earlier this month, the journal Philosophy & Public Affairs published an article largely written by an AI.
Now that the experiment has been discussed a fair amount here and elsewhere online, the journal has decided from now on to prohibit AI-authored content.
Seth Lazar (Johns Hopkins), an associate editor of Philosophy & Public Affairs, shared the journal’s new policy on X.com:
Academic journals serve at least two functions: the promulgation of new knowledge, and identification and credentialing of talented researchers. There are many other means available to share new knowledge, but few for credentialing talented researchers, where “talented” means roughly: able, on their own or together with other researchers, to make and communicate significant progress on fundamental problems. Submitting AI-authored essays makes the task of identifying talented researchers harder.
In addition, while it is at this time possible for a researcher to use AI to author a paper that is itself high quality, it is disproportionately unlikely that they will do so, and much more likely that the paper will have all the surface appearances of sophisticated work, but ultimately be insubstantial or incoherent. This makes it harder to peer review, and makes the task of peer review in general more onerous, at a time when it is already under significant strain.
Those who enjoy and value the products of AI writing have many other places where they can publish such work, notably online. There may also be a case for creating journals of AI-authored research. So there are very few benefits to publishing AI-authored work in existing academic journals, and, given the above, significant costs.
Our policy is therefore to prohibit the submission of AI-authored content to Philosophy & Public Affairs. We define AI-authored content as papers that are substantially written by AI. We distinguish this from the use of AI for light copy-editing (which we in general define as reorganising or condensing written work without rephrasing). We also emphasise that authors are entitled (indeed encouraged) to use AI while conducting their research, in whatever ways they choose.
To implement this policy, we require authors to declare that they have not used AI to author their paper, and to explicitly describe how they have used AI in the process of conducting the research that led to the paper. We will use the leading AI detection software at the time to verify the truthfulness of the former declaration. We note that AI detection software, including updated models, may be used throughout the review process, up to and including after publication. If at any point we discover that the author’s declaration that AI was not used to author the paper’s content was false, the paper will be rejected or retracted, and the author will be banned from further submission to Philosophy & Public Affairs.
In a further comment, Lazar emphasized the “stick” part of the policy at the end:
I think it’s particularly important that we can retract papers and punish authors even when the paper has been published (going forward). No doubt at some points the “humanizers” will be in the ascendancy. If an author profits from that to get one over on us, they can be sure that if the AI detection algorithms improve and they subsequently get caught, they won’t benefit from a statute of limitations.
Discussion welcome.
By the way, I would not interpret this move as “the experiment failed”.
Sorry, my comment got cut off early.
I wanted to add that I conducted this experiment because I thought it would force several implicit questions out into the open and generate excellent discussion. It did. We learned some things from the resulting discussion and that informed what policy we decided to ultimately adopt.
If you know anything about my pedagogy, you know I have a strong preference for turning mere thought experiments into lived choices when possible, because people often discover that their considered abstract principles don’t capture their real judgments when the trade-offs become real.
How will you apply the policy moving forward? Will you run all the papers published recently through a Pangram test?
Glad to see this decision.
I’ll be curious to see how other journals respond, especially with upcoming watermarking and watermark decoding API’s being made available.
It should be easier to screen for AI content and human content, thus easing the burden on journals dealing with spikes in submissions.
“ We will use the leading AI detection software at the time to verify the truthfulness of the former declaration.”
This is highly troubling. The current “leading AI detection software” right now is allegedly Pangram, and it’s trash. Go ahead and feed it some of your own work from the before times and it will often tell you that more than half of it was written by AI.
Pangram itself will tell you that to use its detector effectively you have to feed in substantially long passages–a few sentences won’t do.
Just one perspective here, but in instructional contexts I have found it to be effective at identifying AI-derived work. YMMV, but it’s fair for assessors to use detection software as one tool.
When I said “Go ahead and feed it some of your own work from the before times”, I meant entire drafts.
Accusations of AI-based plagiarism are serious. It’s reckless to base them on Pangram in light of various considerations – e.g. lack of transparency regarding how their technology works, the company’s profit-based motivations to mislead, etc. – but most especially that the technology just isn’t sufficiently reliable.
I have uploaded some of my papers to Pangram and it said all of them were 100% human-written. Do you have an example of a published article that Pangram claims is AI?
I just tested a paper that I published in 2019. Apparently 70% AI-generated content!
Uploaded all my published papers to Pangram and it said they were 100% human.
Various independent studies show a false positive rate near or at 0%, such as:
https://bfi.uchicago.edu/working-papers/artificial-writing-and-automated-detection/
https://link.springer.com/article/10.1007/s40979-026-00226-w
Looking at these, they do seem to have methodological problems. The Chicago study used publicly available human-written texts, which Pangram might have used for training their classifier. The second study used papers written by non-native English-speaking students, which might be much easier to classify as human than professional writing.
If anyone has evidence of false positives, it would be great if they could share it.
(Note: unsurprisingly, false negative rates are bad.)
+1 I understand the desire to use AI detection software, but in my experience, the false negatives and false positives are so noisy, it’s at best, good for helping you identify which papers to give a closer look.
But in this case, if you’re reviewing the submission carefully, you’re going to be much better at identifying AI written papers than any AI detection software will.
“if you’re reviewing the submission carefully”
Isn’t part of the problem, as noted in the justification for PPA’s new policy, the shortage of reviewer (and, presumably, editor) time? If there were much more labor available, it might be possible to thoroughly vet every paper, but everyone knows the big-name journals are inundated with submissions these days. AI threatens to exacerbate that problem.
Also, I’m ok at identifying AI writing, but I’d still bet that Pangram is more accurate than I am, probably by a significant margin, even though it’s quite fallible.
Perhaps, but my guess is that the journals and publishers in question are avoiding solving the real problem of reviewer and editor labor by claiming there is a shortage where there is none (if the right economic levers were pulled) so as to incorporate AI as the solution.
This is a common labor fallacy happening across the US at the moment.
There is no labor shortage, just a shortage of people willing to work without commensurate compensation.
Let me give a concrete example of how this works at the journal I edit (Philosophy of Physics). Philosophy of science journals have always received many low-quality submissions, usually by “independent scholars” who can’t get their work into mainstream science journals. Most of it is obvious nonsense that can be swiftly desk-rejected, but some is borderline, and it’s a judgement call whether you risk wasting a reviewer’s time, or risk rejecting genuinely interesting work because of its unusual style or origin.
AI makes it much, much easier to write papers that have the superficial form of an academic paper and avoid the obvious howlers that an editor can sieze on to make the case for desk-rejection clear. That both increases the volume of dubious submissions (all the philosophy of science journals I know, including Philosophy of Physics, are seeing big increases in these submissions) and makes the desk-rejection judgement harder. (In addition, reviewers understandably object to wasting their time on a rejection letter for a paper they’re confident is AI written.)
In that situation, “this paper is mostly AI-written” is a pretty valuable heuristic to exclude low-quality work, even leaving aside whatever else one might think about AI work. And because it’s a heuristic more than a strict criterion, using something like automated detection to help judge whether a paper is AI-written has a lot to be said for it. It does carry some risks, but all desk-rejection heuristics do, and the whole point of desk-rejection is to make a decision without needing to carefully read and assess the paper at the level that a reviewer would. Frankly, the obvious rival heuristic at present is the one archives are increasingly using: no submissions from independent scholars at all unless some community member vouches for their seriousness. That too carries risks, risks I’d prefer to avoid.
I agree with most of this. On using the heuristic as its intended, not as a strict criterion, another factor to weigh is fairness, on two counts: avoiding false positives and distributing the burden of enforcement.
Disclosure alone is not only ineffective without enforcement but unfair: the conscientious ones will comply; cheaters will lie. The former may be penalized while the latter get away with it. This seems unfair (distributively).
Add detection for enforcement and you probably reduce the risk of false positives (assuming, as seems likely to me, that the risk posed by the best detectors is lower than that of human judgment). Is a corresponding higher risk of false negatives bad? For enforcement purposes, yes. For fairness reasons, less so. But I see no way to avoid the tradeoff between avoiding false positives and effective detection.
All of is to say, journals should be free to experiment with different heuristics given the increasing scale of the problem (AI slop flooding editorial desks), but editors should be aware of and maybe explicit about the tradeoffs. As you note, all heuristics carry some risk, and our default is not no heuristic anyway.
You are assuming that AI detection software is effective at the purpose you need it for.
The reality is the technology cannot do what you wish it did.
We’re not in fact using it at present (we’re in the process of working through an AI policy) but as it happens I think the evidence is fairly strong that the current version of Pangram has quite a low false-positive rate when assessing long-form writing (and so it’s less that I’m assuming that, more that I’ve done some homework and inferred that). If you have evidence to the contrary I would be interested to see it.
Ah well specifying the AI detector helps. I’ve mostly worked with other AI detection software. The users above seem to have more experience with Pangram than I do and had their own concerns with it.
Out of curiosity regarding your point, I took a brief review of the GPTZero technical paper and the Nature article on Pangram, I’m still not convinced it is suitable for the desired purpose and I think attempting to use it before finding alternative solutions (or without at least experimenting with alternatives) seems to predetermine the outcome.
Regarding Pangram’s accuracy, if the 92.8% number in the GPTZero paper is accurate, and run against 1.25MM abstracts, than you still get 90k abstracts where it is inaccurate. I know it seems like I’m being a bit pedantic, but in a past life, we usually did not accept such low rates as being production grade. The gold standard was five 9’s (99.999%), with admittedly lower rates for less critical work. The use cases were different but the stakes seem similarly critical or at least worth debating before implementing.
At the very least, it seems we need to be asking when is it ok to be both 1) wrong about a paper and 2) not give it the dignity of being reviewed by a person?
https://arxiv.org/pdf/2602.13042
https://www.nature.com/articles/d41586-026-02569-3
Regarding alternative solutions, there are extra considerations for the field as a whole. From what I understand, philosophy has a problem with an oversupply of PhD graduates, all of whom are presumably qualified to review papers. And while I don’t doubt that there are many submissions which are poor, I find it hard to believe that the pool of submitting authors has been optimally utilized as a peer review pipeline.
If we’re looking at the economic impact, using Pangram’s Developer AI pricing as a proxy (I realize this will not be 1:1 but I assume this is the most relevant tier below Enterprise / Education), then $0.05 = 100 words on a ~7k word article would run about $3.5/article. Is this a cost journals are prepared to bear over the existing volunteer model? And is it preferable to pay this to a company, rather than to human reviewers who comprise the scholarly community?
Finally, whether the price *stays* this low is another question. As we’ve seen with other startups (DoorDash, Uber, OpenAI) profitability eventually forces products to raise prices or be discontinued, and this may conveniently come years after the journals are locked in and the reviewer / publishing ecosystem has been become dependent on the product.
My perspective follows from years of experience evaluating software products (in my past life). If the product turns out to be more useful than alternatives, cost effective, and durably advantageous over time, then we should certainly take that seriously.
Right now I find it puzzling that we so quickly decry AI-assisted papers but seem so eager to use AI to replace reviewers.
Partially and in brief:
Freddie DeBoer,
“How to Produce a Pangram 4 False Positive”
(August 3, 2026) is a sensible discussion of Pangram’s utility.
Having been a journal editor, I am sympathetic with the plight of editors awash in the flotsam of independent “scholarship” and jetsam of AI slop. As a retired academic and now independent scholar, I am painfully aware of how the absence of affiliation has become an expedient algorithm for busy associate editors to desk reject without even bothering to read beyond the abstract. Even when a paper makes it through peer review and into print, the stigma of independence continues, as with a recent paper attacked after publication, not on its merit but because the author was not an affiliated credentialed physicist.
In my view, lack of credentials or affiliation are not in themselves legitimate criteria for desk rejection. Of course, AEs do not say that directly, but usually cite “out-of-scope” or “not suitable for this journal.”
If part of the function of scholarly journals is credentialing of scholars they publish, then use of AI is a legitimate factor to weigh in decisions to desk reject or send for review and the final decision whether to accept. Other things being roughly equal, scholars who works things out fully on their own evidence a finer mind and deeper scholarship than the ones who lean on AI for ideas, background research, design, etc.
I am broadly inclined to agree that journals should not use credentials or affiliation as filters. That’s even though it would be an extremely effective heuristic, as (in my admittedly anecdotal experience) the vast majority of work submitted by independent scholars would fail peer review. But on principle journals need to be open to all sources. Certainly Philosophy of Physics does not check affiliation prior to making decisions on desk rejection. (I feel somewhat differently about arxiv.org and similar.)
I do think it is getting increasingly difficult for journals to screen neither for AI use, nor for affiliation – the slop is just getting too overwhelming.
I don’t think this is as clean a call as it’s being made out to be. I use AI extensively in my writing, not to generate prose but to organize arguments across the corpus of my published work and advance ongoing arguments around a set of themes and principles that define the scope of my practice and goals. One such tool looks for opportunities to make connections between my different research programs and highlights possible connections. Another identifies gaps in my knowledge graph for me to consider exploring. By the editor’s statements, the results credit me for insights that were not mine—my insight was to build the tool that could do that work. I also use it to critique my writing based on my own published work for strength of argument, conciseness, etc as a quality control prior to submission. Is it my writing if the entire source of recommendations was my own? There is not a clean line to be drawn on creation, or a viable epistemically valid tool so long as it is reliant on probabilities. This AI exceptionalism may cut reviewers time, but let’s not claim it’s doing more than it can.
“if you’re reviewing the submission carefully, you’re going to be much better at identifying AI written papers than any AI detection software will.”
I don’t think this is true. I’ve spent a while talking to AI models for various hobbies (and for the AI literacy class I’ve been teaching frequently), so I’ve developed a good sense for AI-written text. But even so, it sometimes takes me a while to identify them, or I do so only on a second read. And most people who don’t have as much familiarity with AI-written text as I do are much worse at this. I don’t think I can out-perform Pangram on this.
I wonder if journals using Pangram and similar might in some sense be violating author confidentiality. It isn’t clear whether these papers might form training data for the relevant company. Pangram says it won’t share data with third parties, but even taking them at their word, this says nothing about them using it for training their own models. At the very least, authors ought to be warned that their papers will be used this way.
Pangram are completely explicit that they don’t do this: Privacy Policy | Pangram. Relevant quote: “Your submissions to Pangram are not used to train our AI or machine learning systems”.
They might be lying, of course. But so might the owners of Dropbox, or MS Office, or the backend content-management software that the journal runs on. At some level, you either trust IT companies not to be falsifying their privacy policies, or you go back to the days when articles were produced on manual typewriters and then mailed to journal editors in triplicate.
In some of the cases (for both security – confidentiality, for example, and privacy) you *might* be able to get independent review. This does create a regress of “who do you trust” (your campus cyber/IT security might have opinions, and some professional organizations do say to use them here), but that aside, that does create a difference between very well known products like M365 (Office) and academic tools like Pangram. Incidentally, the former is a lot more complicated securitywise than even us in the profession realize at first (I know I was shocked.)
AIs can now do philosophy. How will the field adapt? My vision is that researchers will explore a wide range of AI uses. Some will focus on AI for lit review. Others will use AIs to subject their arguments to withering scrutiny, generating objections and replies and so on. Others will use AIs to build formal models, and to aid in proofs. Others will use AIs to help with initial drafting. Others will take an initial idea, and let AIs flesh it out into a complete paper. In the limit, some will explore end to end philosophical research with AIs.
PPA’s new policy is simple: the final draft must be written by a human. This has some benefits. It ensures the human author has a better understanding of the paper. It removes AI prose, which is usually worse than human prose. It imposes a barrier against pure slop. The policy also has costs, limiting the range of experimentation. But overall, PPA’s policy is permissive, allowing AI use throughout the paper production process, except at the final stage.
What is the point of all this experimentation? The same as it is in every other industry. Philosophers cannot sit by and pretend that AI does not exist. Nor can AI research be confined only to niche journals, while most researchers carry on using the same methods as before. Our duty as researchers is to use all of the resources available to us to produce the best work possible.
Right now, nobody knows exactly how to use AI for philosophical research. And the use of AI will require changes to the basic norms and structure of our field. That is our challenge. In this environment, a culture of open-mindedness, experimentation, transparency, and toleration is essential.
“Our duty as researchers is to use all of the resources available to us to produce the best work possible.”
How do you square this duty with signing your name to what you yourself judged to be work that’s below your usual standard, and which you “intentionally avoided correcting…too much at the finest levels of detail”?
Good question. In general, I believe that the best way to figure out how to use AI tools is to engage in all sorts of experimentation. This will involve familiar exploitation-exploration trade offs. If I simply try to make each *individual* paper the best possible globally, I’ll lose out on information on how well a new method can do. My goal in the PPA paper was to see how far a particular approach could go. We researchers are all bound by a duty to produce the best work we can. But this duty has to be pursued using mixed strategies; different projects will involve different methods.
I agree wholeheartedly with the last paragraph of Goldstein’s comment. But I worry about his claim that we have a “duty” as researchers “to use all the resources available to us to produce the best work possible.”
1) Who is this duty owed to? I have many duties to my peers, colleagues, mentors, and students. Is this duty owed to them, or to anyone? Or is it an impersonal duty to maximize the number of philosophical truths known? I find the latter ethically implausible. I would just say we have reasons to produce sufficiently good work, not a duty to maximize the quality of our work.
2) Are there really no ethical limits at all on the resources we should help ourselves to? Trivially, we should use resources only if it’s permissible to do so.
So I propose instead: “As researchers, we have reasons to produce good work. As a result, we have reasons to help ourselves to the best available resources for research, if doing so is permissible.” For me this does not create obvious pressure to use AI as a co-author, because I am not sure how to do that permissibly. (And right now, I don’t feel pressure to do that because I don’t think AI is good enough yet at generating new ideas. I have yet to see an LLM draft an excellent philosophy paper, or subject a paper to “withering scrutiny.” But I believe they will get there eventually.)
Lastly, I might be misinterpreting the comment, “Nor can AI research be confined only to niche journals, while most researchers carry on using the same methods as before.” But if I understand correctly, Goldstein is saying we shouldn’t have separate journals for AI-generated research and research made by humans. Why not? What would be wrong with that, or with just putting these papers in a free, widely advertised online repository? Goldstein seems to believe that we have a duty to publish every publishable paper. But we have no duty to adore every adorable puppy, or to do laud every laudable act.
I think it would be great to have some journals focused on philosophical research by and with AIs. But it is important that research by and with AIs is not confined exclusively to those journals. Instead, the use of AI needs to gradually become a central part of the workflow of most researchers, in myriad ways.
Otherwise, human philosophers risk sitting on the sidelines. We need our best human researchers to be trying hard to figure out how to use AI. We need the great research that comes from this process to propagate through research networks as effectively as possible. This requires the best journals to participate, rather than waiting for someone else to start something from scratch. The problem is that a brand new AI-only journal is likely to be ignored by most philosophers.
One heuristic is to compare philosophy to most industries. In most industries, AI use is not going to be put in niche sidelines. The best firms in the world aren’t going to ban AI use, and wait for competitors to start up brand new AI-only companies from scratch.
PPA’s policy is consistent with this approach. It merely requires that the final draft of the paper be written by a human. This allows all sorts of flourishing experimentation.
I agree. The claim that “we have a duty to use all the resources available to us to produce the best work possible” is ludicrously strong.
What if a button was invented such that when pushed an excellent philosophy paper is generated, but somewhere, a philosopher suffers excruciating pain?
I don’t think I have a duty to press the button!
Duties can conflict.
Sure, but is that’s what’s happening in cases like this? I am sceptical: I find it more likely that a duty that expressed so strongly is simply not a duty.
“Philosophers cannot sit by and pretend that AI does not exist.”
Watch me.
This attitude will not age well – it’s called sticking your head in the sand like an emu. Better to take your head out of the sand and face the music head on
The emus are hiding their heads from music?
Hi everyone. I am currently workshopping a joke centered around the idea that to get from “emu” to “music” you have to add “sic,” which sounds like “sick,” and so it stands to reason that emus would want to insulate themselves from music. It’s taking a while and I expect to be done in 2028 or thereabouts, so I just wanted to reserve this spot in the comments section for when I finish. Thanks!
ChatGPT (5.6 terra):
Why do emus avoid concerts?
Because if you add “sic” to “emu,” you get “music” — and nobody wants to catch something that sick.
Well now we know. AI is better at philosophy than humour.
Normally I would reward a joke of this quality with points for effort, but AI can’t really claim to have put in the sort of effort worth rewarding, so sadly I must give this a 0 out of 10. The phrase “catch something that sick” doesn’t make sense, and even if it did, it wouldn’t work when the “something” is the sick itself, since noting that the sick is sick is not a sufficiently comedic intensification or subversion of expectations, it’s just a repetition of the sentiment. Also, the framework the question sets up doesn’t get paid off in the answer. Is it going to the concert that would add the sic to emu? Why? The concert should add music to the emu, not merely sic. If the idea is it that adding the sic to emu happens independently of attending the concert, but somehow avoiding the concert is still something the emu wants to do in light of the sic + emu problem, this is badly-elucidated by the claim that “nobody wants to catch something that sick,” because going to the concert would not cause the emu to catch it; they’ve already caught it in this hypothetical, so at that point going to the concert can’t hurt the emu. And it’s a bit too much for people to parse to tell them “if you add sic to emu you get music” – you want to order it in the other direction, telling them what happens if you merge emu and sic. That makes it less likely they’ll encounter friction in terms of seeing the connection quickly. Better luck next time ChatGPT!
Isn’t this reply a performative contradiction?
Interestingly, no.
“AIs can now do philosophy.”
I don’t think this is true. At most it can help you do philosophy, though I’m not even sure about that.
AIs cannot do philosophy, since AIs cannot think.
Doing philosophy is a public activity. If one received conclusive evidence that David Lewis was actually a p-zombie, it wouldn’t seem relevant to the philosophical significance of his published work.
(That if, if the idea of a p-zombie was coherent, which it isn’t. Dan: we miss you.)
The output of an AI or p-zombie, by itself, would not have significance. The same strings of words might have significance by being the non-accidental product of being understood, but there is no understanding in an AI or a p-zombie. So yeah, my point is not that an AI can’t possibly output something that a knower might, by thinking it, imbue with philosophical significance. But that doesn’t mean it is doing philosophy. That’s important, because it’s important to know that philosophy is not just the generation of strings of words. I’m actually not sure how many people know that anymore.
Leaving aside the fact that you’re relying here on quite controversial assumptions in philosophy of mind, the distinction doesn’t really matter from the point of view of contributing to a research community.
Jr Phil Prof, it it turned out that the work of David Lewis was authored by a bot or an alien, or a cat, then it means that any of those things can (apparently) do philosophy, indeed, very well. It doesn’t even marginally count against the thesis that Lewis’s oeuvre (whomever did it) is good philosophy.
Steelmanning a bit, this would make sense on a broadly Searlean conception of content: the thoughts and utterances of conscious beings have original meaning, the texts they write derive their meaning from their authors, so if a non-conscious being produces a text, that text has no meaning. If a human then reads it and endorses it, it acquires meaning.
And, for Searle, our grounds for attributing consciousness to entities are biological and have nothing to do with syntactic information processing, so the bot, at least, is not conscious.
(Contrast Chalmers, who rejects a functionalist/behaviorist approach to consciousness but – IIRC – also doesn’t think content depends on consciousness.)
I have extremely little sympathy for this kind of approach to content, or indeed to consciousness, so my exegesis might not be reliable, but that’s my understanding. I do think it’s fascinating how LLMs have suddenly made these classic conversations in philosophy of mind rather directly relevant to concrete concerns.
I thought p-zombies could think, only that they were not phenomenally conscious. The reply would work better if Dennett was a Chinese Room. Unfortunately that’s not a coherent thought experiment either.
I wasn’t clear whether Jr Phil Prof thought that Real Philosophy required some ineffable quality of phenomenal consciousness or some equally ineffable quality of underived intentionality. If the latter, then agreed: I should have gone with a Chinese Room puppeting Lewis remotely.
Sorry, what’s AI again? I’ve been pretending that it doesn’t exist and I’ve completely forgotten what it is.
To reject a philosophical work because it was written by AI seems like ad hominem.
More like ad algoritmi.
Except, of course, that it’s only a fallacy when you judge the conclusion of an argument to be false because of who made the argument. I’m not judging the conclusion of any AI paper to be false, I simply refuse to read the paper. There’s no ad hominem in rejecting a philosophical work in that way
That’s only feasible as a strategy insofar as AI-authored work doesn’t generally enter the publication ecosystem. If it does, the usual rules on citing relevant literature would apply (unless we as a community explicitly decide they shouldn’t, in which case presumably the AI-authored work *won’t* have entered the publication ecosystem).
I’ll exercise my professional judgment about what work’s worth engaging with and what work’s worth citing so as to direct other people to it. That’s the same way I do things now. There’s no such norm as “cite everything that’s ever been published related to this topic,” that is, no such norm as “cite everything that’s entered the publication ecosystem”
But there is a norm of “cite prior originators of ideas you make use of, irrespective of whether you learned of the idea from the source or independently” – you can’t discuss Gettier cases without citing Gettier, even if you happened to come up with the idea independently as an undergraduate. And there is also a norm of “if paper A discusses an idea drawn from paper B, and you build on paper A, you also need to cite paper B”. Both of those norms are going to require citing AI papers if they’re in the ecosystem. Citation isn’t just an assessment of what you think is worth people reading: it’s also a discharging of concrete scholarly obligations.
Ah, good. So I’ll neither read papers written by AI nor read papers about papers written by AI
Even if an AI writes an influential refutation of your work?
This is really just a replay of the discussion about whether it’s acceptable not to cite Problematic Philosophers. What’s old is new again.
I’m a fan of ad hominem arguments. It seems relevant to know that Exxon made lots of money from selling oil when they tell us about climate change. My take on most of the “informal fallacies” is they aren’t fallacies at all. Just reasoning with background premises or experience.
That’s because when they “tell us about climate change” they would likely be making various unsupported claims that they simply expect us to trust them on. Its not an ad hominem fallacy to point out that we can’t trust them on such claims given their vested interest. However, suppose that Exxon didn’t ask you to trust them and instead started their argument by pointing to various claims that you already accept for independent reasons. They then presented you with a series of seemingly strong ampliative inferences and valid deductions that infer some conclusion about climate change from the claims you already accept. In that case don’t you see that it would be a mistake to simply dismiss this argument on the grounds that Exxon has a vested interest in supporting its conclusion?
I’ll bite here.
There is no puzzle if we distinguish between “argument” as a part of an exchange/dialectic and “argument” as a formal construction that is semantically interpreted. (Compare: “proof” in mathematical practice versus “proof” in proof theory.)
The former, but not the latter, have all manner of pragmatic coloring, implicature, etc., which might just as well be very nearly out of the control of the agent who makes the utterance. Exxon may, for its social place (etc.), simply be “the wrong person to make that argument”—a colloquial expression that highlights that, in discourse, the mouthpiece does sometimes matter as much as the contents of the utterance.
According this new policy, journals serve two purposes (1) promulgation of new knowledge and (2) identifying and credentialing talented researchers. If the journal accepts papers written largely by AI, then it would cease to be able to reliably identify talented researchers. Therefore, they should not accept papers written substantially by AI.
Here’s the reason why this is a bad policy. We need to distinguish two stages of producing a paper. There’s the research part and the writing part (I’m not assuming there’s no overlap here). This policy prohibits papers “substantially written” by AI. But what about the research part? The policy explicitly mentions this:
“We also emphasise that authors are entitled (indeed encouraged) to use AI while conducting their research, in whatever ways they choose.”
So presumably, an author could prompt AI in the following way as part of their research: “Hey Claude, here’s theory X (read the file), find counter-examples to points B and C. Illustrate them with examples. Then come up with the strongest counter-points. And then give responses to these. Write it out in outline form. Do not write a paper. I do not want to violate PPA’s journal policies. Outline form only Claude.”
Claude would comply with an outline of this research, which is the bulk of the intellectual work. Imagine it is of high quality (this is not what is at issue here). The author would then just write out this research in their own words. It would never be detected by Pangram or have the “watermarks” that will soon be made available by the AI companies. And this would be perfectly acceptable according to the new PPA policy. Here’s the problem. PPA would credential the person who submitted this paper as a talented philosopher, even though he or she is not (or need not be).
By making this distinction between writing and research, the new PPA policy fails to do what it was designed to do–maintain the function of the journal to identify and credential talented philosophers.
Of course, they could change their policy to prohibit AI assisted research (you can’t use AI to generate counter-examples, for instance). But since this research work with AI is undetectable, there would be little incentive for anyone to disclose this. This will just be an empty policy that will only end up harming the honest philosophers who faithfully follow the policy. Similarly, PPA could decide to call processes like “making an outline” a type of “writing” which violates their policy. The same point applies since they will never be able to detect this when the philosopher just writes out the ideas in his or her own words.
It’s better to simply have the journals publish the best articles possible regardless of how they were written. We need to solve the credentialing problem in some other way. Ironically, the more journals are used by institutions for credentialing, the more incentive there will be for people to use AI in writing papers (to get the best possible content out) and not tell anyone about it, especially as LLMs improve over time.
It’s not obvious to me that if someone pursued your strategy and got a paper published, that wouldn’t be evidence of them being a good philosopher.
Consider: in the first place, Claude might be pretty uneven in the quality of the ideas it comes up with via the prompt, so that very considerable human judgement is needed in working out what to do with those ideas and how to filter and modify them. It might also have bad judgement as to what problem to work on in the first place, so that more human judgement is going to be needed in directing it what to reason about. Certainly the present AI state of the art is, at most, at this level: it won’t spontaneously produce, unaided and with a naive prompt, reliably publication-worthy article sketches.
But suppose it gets so much better that it can do that too. Even so, philosophy is not just about the outline structure of the paper’s argument; except in its most formal corners, it’s also about the written expression of that argument. Even people like me who have a pretty scientistic attitude to philosophy will acknowledge that if you strip a paper down to its outline you discard much, perhaps most, of its philosophical content. And so insofar as Claude can’t write good academic philosophy, that’s an actual deficiency it has as a philosopher.
If AI philosophy plateaus at something not qualitatively beyond where it is now, writing good philosophy will require human skill both in the original interactions with AI that outline the paper, and with the actual crafting of the textual realization of that outline. And so if someone produces a paper that way, that will be evidence that they’re a good philosopher. That wouldn’t be a terrible way for things to go. (Without prejudice as to whether AI will get even this good, though it’s already at the foothills.
Of course, if Claude’s thinking and writing gets qualitatively better, to the degree that I can give it a prompt like “write a 10,000 word paper on philosophy of quantum field theory in the style of David Wallace that materially advances our understanding of the field”, and without any further intervention it writes something better than my work and does it faster, then it does become largely pointless for me to write my own work except as a hobby. I have no idea at all if this will happen (and I don’t think anyone else really does). I have mixed feelings about the possibility: on the one hand, it’s nice to be useful; on the other hand, if it can do that, it can probably do equivalent things if you prompt it to make a meaningful advance in, e.g., medicine or energy-generation technology.
I hope that philosophers are taking the right lesson from this episode.
Before using a large language model as a writing aid, read the comments in the thread reporting PPA’s publication of “an article largely written by Claude.” Unfortunately some of the later comments were lost as a result of a computer security incident, but you can still read Alex Guerrero’s criticisms of the article, which seriously misrepresented Guerrero’s work. Note both the content of the criticisms and the tone.
You cannot rely on LLMs to summarize sources accurately. You must read what you cite.
So take two cases. Let’s consider Linda and Linda*. (a) Linda uses AI to find new objections, explore argumentative structures, perhaps discover new philosophical lines of thought. Then she proceeds to do the writing herself. She is a native speaker and does a decent job. (b) Linda* works completely by herself on the main idea, the arguments, the objections, and writes everything in her mediocre English as best she can. She is not a native speaker, so she then turns to AI to rewrite the text in idiomatic academic English. Alas, Linda* is at fault. It is striking not to find a single line in this whole discussion addressing this very obvious point. Privileges be privileges.
Is it really “striking” that there’s no mention of this?
If you want to pivot to framing discussions about these technologies in terms of there being tools to potentially right (or else to exacerbate) historical injustices, then go for it. (There are, of course, plenty of discussions along these lines. I doubt the new tools categorically tilt in the favor of the oppressed, because when does that ever happen…)
But that framing is an entirely different framing than that of the present discussion: one of productivity, efficiency at the community level, harms to outputs, etc. Call this a privileged framing, sure, in that it is a privilege when one does not feel compelled into a social justice framing. But all the same: it surely can’t be all that striking that your point, however obvious it is, would not be aired in this discussion that is simply framed in other terms.
Frame it however you like. A policy that penalizes the use of AI to smooth one’s English disproportionately penalizes those who struggle with academic English.
So does a policy against uncredited human co-authors, but all journals have that.
I think machine translation is an interesting point and does put some pressure on a no-AI-writing policy, but “this would disproportionately harm non-native speakers” is not an automatic reason not to have the policy.
“So does a policy against uncredited human co-authors, but all journals have that.”
I think maybe Lindastar has a point if you consider that some journals (many of Springer’s for example) advertise copyediting services in their guidelines section and some authors definitely use those services, or professional copyeditors, or even just friends helping as copyeditors for their work. Often times these helpers go uncredited, or if credited at all are maybe simply thanked in the acknowledgments with something like “for their helpful comments on an earlier draft of this article”. I know plenty of non-native English writers who have done something like this, and I see nothing wrong about that. It makes sense for, for example, a Chinese speaker who is not perfectly fluent in English to ask a colleague to read their paper and help them make sure they’ve articulated their points correctly or offer suggestions for better wording, grammar, etc.
My point was that there is nothing striking about the omission. Rather, it’s a consequence of the framing. I’m not sure why your response was to repeat your point, which I had essentially already endorsed in passing.
“So take two cases.”
You can just say “take two cases.”
While I understand and sympathize with the rationale behind the policy, I have to sit back and laugh at the big-picture absurdity of identifying authorship with the final presentation. To repeat, I understand and sympathize, but isn’t it absurd in the big picture?
Well, I mean, the policy didn’t allow putting AI as an author. My understanding is that Simon would have been happy to and even preferred to have listed Claude as the co-author. But since it wasn’t allowed, that just left Simon.
This is a COPE guideline, at least partly on the grounds that authorship requires taking responsibility and AI isn’t (at least currently) the sort of thing that can take responsibility in that sense.
Authorship and AI tools | COPE: Committee on Publication Ethics
(FWIW I asked Claude what it thought about the guideline and it said it was fair enough.)
I think Lazar’s description of the credentialing function of journals is a bit too coarse-grained here. Identifying talented researchers matters for allocating jobs and, perhaps, grant money. But journal publications also identify experts in a subject matter. When you write and publish a paper the old fashioned way, you learn a massive amount about the topic it’s on—that’s my experience anyway. You spend hundreds of hours reading, writing, revising and that builds subject matter expertise. This matters because it signals you’re relatively expert on the paper’s topic. This means editors call on you to review future papers in the area, grant agencies ask you to evaluate proposals, and colleagues, in the broadest sense, should take your opinion on the area more seriously than otherwise. I think this is how academic consensus in an area is forged and changed. If you use Claude to generate a paper over 6-7 hours I don’t think you build anything like this level of subject matter expertise. You are not a suitable reviewer in that area, you shouldn’t evaluate grants and your opinion shouldn’t be given much more weight than that of any other philosopher. So a system in which publications are quickly AI generated means we lose a very important way to identify subject matter experts. It’s not clear to me how to replace it.
At present, and unless the technology gets qualitatively better (which it might!) I don’t think you could use Claude to generate an excellent paper without being a subject matter expert in your sense.
Well, levels of expertise and quality of papers are both scalar. My claim is that, using Claude, you can publish a paper of a given quality at a lower level of expertise on its subject matter than when you had to write it yourself. So publication becomes a worse signal of expertise. And my further claim is that having a signal of expertise on fairly narrow subject matters is useful for our general system of knowledge production.
I’m not sure about that (genuinely not sure, that isn’t meant to be passive aggressive). It’s also easier and quicker to write papers if you have lots of chances to bounce ideas off colleagues and students who are also well versed in the subject, but I don’t think that means a paper from a philosopher who does that is (even ceteris paribus) less of a sign of expertise than one from someone who doesn’t.
So would you say that Simon Goldstein’s PPA paper is not excellent or that he is not a subject matter expert in political philosophy in Adam Lovett’s sense or both?
Not being a subject matter expert in political philosophy myself, I don’t know. But the logical structure of my claim only commits me to: if Simon Goldstein’s paper is excellent, then Goldstein is a subject matter expert. So I’m not quite sure what point you’re making.
Journals may have been used that way—as a proxy for expertise—but that doesn’t mean it is, let alone should forever be, their function. It is not clear to me either how to replace this proxy, which I agree is helpful if imperfect.
Hey Adam, thanks for the thoughtful analysis. In general, I think that people should defer to sources based on their expectation of the accuracy of that source. So imagine I can defer to one of two potential sources: a normal human using an AI, or an expert human. I want to defer to the source I expect to be more accurate.
How do I figure out how accurate each source is about a topic? One way is to check whether either source has produced research on the topic in the best journals for that topic.
In this way, I think that AI will change our practices of deference. We will need to shift a bit away from deferring to whoever has the best ideas on their own, and shift a bit towards deferring to the people who can use AI to produce good ideas.
Unfortunately, if the best journals all block AI use, we will lose this way of assessing the accuracy of sources. But to be clear, this is not what PPA’s policy is doing, since it allows AI throughout most of the research process.
Maybe this would work, but I wonder. On the one hand, it’s a bit hard to imagine how the norm of consulting human-AI hybrids would function. Today, when I ask a colleague what they think about an issue I think they know about, I ask them; I ask a human being. I don’t qualify my request with the demand they consult AI before answering. If I started doing that with some colleagues, I’d need to identify who to do it with (so we’d need a good way of identifying AI-assisted papers) and I’d probably need some way of telling whether they’d consulted AI. I’m not sure how I’d do that. And obviously all this would bring consulting colleagues very far out of line with ordinary human communication, which I’m unsure would really be sustainable.
On the other hand, it’s not obvious than centaurs who are good at writing papers are also good at evaluation. Obviously, capacities can be spiky. One concrete worry is sycophancy. The current models will do a lot to try and support my judgement about an issue. That might be fine when I’m writing a paper, if my goal is to come up with the best argument for some position. But it might really mess up evaluation. If I don’t know about something myself, and someone asks me to form a judgement together with an AI, I’ll probably form the judgement in my head and then the AI will do a lot to back up the judgement. If my initial judgement is not very accurate, this process won’t lead to accurate evaluations. So it seems to me unclear whether people who are good at writing papers with AI will be comparably good at using AI to evaluate things.
I don’t mean these issues to be decisive. But I don’t think the problem is solved by saying we can just consult human+AI rather than a colleague.
“In addition, while it is at this time possible for a researcher to use AI to author a paper that is itself high quality, it is disproportionately unlikely that they will do so, and much more likely that the paper will have all the surface appearances of sophisticated work, but ultimately be insubstantial or incoherent. This makes it harder to peer review, and makes the task of peer review in general more onerous, at a time when it is already under significant strain.”
I really don’t get this reasoning. Is the presumption that reviewers are bad at separate quality from fancy “surface appearances”? That may very well be the case, but then the whole function of identifying talented researchers fails anyway (because reviewers are unable to separate philosophical talent from superficial writing talents) and then the starting premise of the policy fails anyway.
You can’t submit a peer review that just says, “superficially this looks fine, but it’s really a confused mess, trust me bro”. The most annoying and time consuming reviews to write are the ones where it’s quickly obvious the paper shouldn’t be published but then it ‘s a tedious and painful drag to explicitly pin down why not.
Well, time consuming efforts for reviewers doesn’t seem to be the argument, right?