Academia is in turmoil regarding partially or wholly AI-written research. AI-written work is being submitted to journals, but it isn’t clear how much. Publisher, journal, university, and department policies are emerging and fluctuating. Technology develops, rendering recently enacted policies ambiguous, vague, or subject to unintended loopholes.
With everything in flux, it’s impossible for norms to develop.
There is no general sense of whether and what kind and what degree of AI use by authors is acceptable. Different policies across different institutions means that what may be problematic AI use in one place is acceptable AI use in another, and vice versa.
Into this chaos steps forward the vigilantes. The vigilantes have always been there, but now they are newly armed and emboldened. In the everchanging and never ending race between the technology of deployment and the technology of detection, the latter seems to have a temporary lead in the form of Pangram. You can feed a text into Pangram and it will tell you its analysis of whether it is AI generated, and if so, what percentage of it is AI generated.
Pangram claims to have a remarkably high detection and low false-positive rates, and cites studies by third parties to support these claims. Yet I have seen plenty of posts by friends and colleagues across social media showing false positives with some of their own work. So user beware.
That said, even if Pangram’s claims about its accuracy are correct, it won’t always be helpful to simply know that a text is “AI generated.” This is because there are two further problems.
The first problem is ambiguity about what “N% AI generated” means. If someone asks Claude to write them a novel article about a philosophical problem, not even giving it a thesis to work with but asking it to come up with one itself, and that’s the extent of the human’s contribution, the resultant paper will get a “100% AI generated” evaluation. And if someone spends a year researching, writing, workshopping, and revising a paper, and then asks an AI to improve the writing style throughout the paper, the resultant paper will get a “100% AI generated” evaluation, too. Despite their same scores, though, these seem like cases that it might be plausible to treat quite differently. And of course, there are countless variations of AI-use between these extremes.
The issue remains even for “0% AI generated” findings, for such a score might be given to both an article that was written entirely by a human who never used any AI tools, and an article written entirely by an AI after minimal prompting by a human, but then rewritten by the human in their own words and style.
The second problem concerns the variation in policies across academic publishers and journals. Different kinds of AI use are allowed or prohibited in different venues. Learning that Pangram reports a published article as, say, “50% AI generated”, even if that figure is in some sense correct, cannot tell us whether the AI use it’s detecting was illicit, except if the article is published somewhere that prohibits any and all AI use. Not only does that percentage not tell us how the AI was used, but it doesn’t tell us whether the AI use violated any of the journal’s or publisher’s policies.
Now it is still the case, today, that most philosophers disapprove of other philosophers using AI to write their articles. Whether you agree or not, many philosophers view the use of AI to write for you as wrong—as a form of plagiarism or cheating, or a misunderstanding of what it is to do philosophy. In short, “AI generated” is, to most ears in our discipline, a kind of condemnation.
What this means is that when someone feeds an article into Pangram and publicly reports that it was “100% AI generated”, they are accusing the named author(s) of professional misconduct.
That is a serious accusation! It can be quite damaging. Given how human psychology works (and how internet searches work), even if the author is shown to not have acted wrongly, negative connotations may linger, damaging the author’s professional prospects for years to come.
Given the possible damage, and given the two problems described above, it seems we have strong reasons, at least for now, to refrain from publicly posting the “scores” of articles fed into AI detection services like Pangram.
What’s an alternative? Well, you could initially privately contact the author, the editor, or the publisher with your concerns.
* * *
As some readers will have figured out, the foregoing was motivated by a recent actual case in which one philosopher posted on social media that another philosopher’s recently published article was, according to Pangram, 100% AI generated, despite the article lacking an acknowledgment of AI use. This led to an investigation by the publisher (Open Library of Humanities), which ended up clearing the author of any wrongdoing. The publisher wrote:
The investigation found no issues with the content of the article that would suggest any academic malpractice was committed, nor does the OLH have any reason to believe that generative AI was used to the extent that the complainant suggested, or with the intent to commit academic malpractice… In line with the OLH’s ‘Artificial Intelligence (AI) Policy’ on providing an ‘AI Declaration Statement’, the author did not meet the criteria set out within the policy to provide a public declaration statement, namely: ‘the use of generative AI that forms the basis of a specific research purpose’, ‘the use of generative AI that forms the basis of a controlled simulation that produces outputs, such as synthetic data to supplement partial datasets, which informs the article’s argument(s)’, or ‘the use of generative AI for specific demonstrative purposes, with generated output(s) quoted and cited clearly with the prompts used, the model version and the URL of the service as part of its citation or caption’…
[W]e have concluded that there are no verifiable issues found that provide or support accusations of AI misuse or malpractice.
You can learn more about this case here. Note: I ask that readers not use or mention the name of the author or the title of the article in that case in the comments here. The article was published in Political Philosophy, so if need be you can call it the PP case.
By the way, the philosopher who had posted on social media about the article’s Pangram score has now kindly removed his original post about it.
>[Edited to add:] I thought this would be clear from the post but in light of some of the comments (not all of which have been approved) I guess it wasn’t, so I’ll make it explicit: vigilantism in the comments regarding the PP case is not welcome, and any further comments relitigating it or speculating about the nature and extent of the AI use in that case will not be approved.<
* * *
Relatedly, a friend recently wrote in with a tip for referees concerned about refereeing AI-written work: start with the references.
They write:
Usually, before reading a paper, I read the references to get a sense of the context of the paper. I immediately noticed a reference to me that was fabricated: I have no paper with that title and have never published on that topic in the journal it was supposedly in. When I put the fabricated reference into Google AI and then Chatgpt (free versions), both responded by saying it was a paper in [journal name redacted]… Only after I responded, “this paper does not exist!” did they both reply with their usual, “sorry, I made a mistake,” etc.
I then checked the rest of the references in the paper and found 3 other errors: 2 relatively minor and another fabrication.
Other suggestions welcome. Use them responsibly.
A quick thought: “vigilante” here is a very biased framing. Why not whistle-blower or sleuth? The framing assumes that the journal, publisher, university, etc. are the legitimate authorities in adjudicating the case. Obviously such legitimacy can be lost if such authorities are frequently found to close rank and perform cover ups (hence the need for retractionwatch et al). This is evidently the case from the other case which will not be discussed. We might also consider an analogy to public campaigns against serial sexual harassers that have been shielded by departments and universities.
I have no firm view about to what extent AI should be allowed in publication, but it’s worth noting that the ambiguity claim goes both ways. The PP case statement, as I understand it, is perfectly consistent with the paper being largely AI generated and the publisher simply as a rule not holding any detection software as reliable. That is to say, besides hallucinated references (maybe?), a confession, or a watermark, it seems that the publisher is not prepared to accept any of the currently possible evidence that a paper is largely AI generated.
I do think “vigilante” is more appropriate than “whistleblower”. I think that a “whistleblower” is usually an insider using their privileged access to information to publicize a private violation of policy, while a “vigilante” is usually an outsider who is keeping vigilant about apparent violations that aren’t being prosecuted by authorities. These terms have negative and positive connotations because of the historical track record of people in the appropriate positions, but I think they’re fairly descriptive, and appropriating one title for the other in order to change the connotation seems problematic. (A “whistleblower” here would be a co-author or journal editor raising the red flag.)
Whether or not the paper actually used AI, it absolutely reads as if it did—so much so that I’m skeptical of the investigation’s conclusion. But that’s not really the important upshot, imo. The real story is that this is the ecosystem AI has given us: one in which, among many other lamentable things, anyone who happens to write (justifiably) in the style of an AI will have aspersions (justifiably) cast upon them.
Whether or not the paper actually used AI, it absolutely reads as if it did—so much so that I’m skeptical of the investigation’s conclusion.
One of the points of the original post is that an author may have published a paper that reads as if it was written with AI, and in fact was written with AI, yet that that does not itself tell us how, exactly, the author used AI, nor whether its use was illicit, given the journal’s or publisher’s policy (unless the policy is one of absolute prohibition).
[Edited to add:] Vigilantism in the comments regarding the PP case is not welcome, and any further comments relitigating it or speculating about the nature and extent of the AI use in that case will not be approved.
Perhaps related to Meme’s point: I wouldn’t rely on any given AI detector as the “smoking gun” for AI misuse, but it can help identify work that smells like AI.
That’s still valuable because few people want to read AI. This is not in dispute. It’s soulless, tedious, and of questionable quality, unless there’s lots of expert human editing.
So, it’s reasonable to triage one’s workload — whether we’re talking about an individual scholar or a journal editor — by de-prioritizing or even ignoring stuff that even reads like AI. This lessens the risk of a journal publishing AI slop, which would turn off readers, would-be authors, etc.
Back to AI vigilantism: it’s not a good idea to publicly accuse someone of something as serious as AI misconduct in academia, without strong evidence.
But what if they’re just merely called out for writing in the style of an AI? It could then be relevant if the more-reputable AI detectors say the work sounds like AI to their digital ears.
As a possible problem: yes, that would still catch “innocent” writers who happen to write that way already, or maybe appropriately uses AI to smooth out language issues. But those charges aren’t false, that their writing sounds like AI, right? The identified authors can always offer an explanation, e.g., that they’re non-native speakers.
I’m not necessarily proposing or advocating for this, but it may be possible to preserve the deterrence value of AI vigilantism (i.e., no one wants to be caught up in that, so they’ll behave better w/rt AI use) without the serious liability of falsely accusing someone of AI misconduct.
At best, these more-responsible vigilantes would just be building lists of people whose writing styles they don’t like, specifically for sounding like AI. Maybe that’s still problematic for different reasons, but it’s not as bad as accusing people of something worse based on thin evidence, which seems to be the main complaint of the original post.
And it strongly seems we need to have a triage process or otherwise do something, given the unwelcome flood of AI slop that’s gumming up our workflows and processes. That’s also not in dispute.
That’s fair. But the use of AI which I have in mind—separate now from any suspicions about this particular case, so I’m not relitigating anything—strikes me as bad for reasons independent of whether it conforms with any journal’s policies.
Preach!
I’m on the side of those who think Pangram is quite reliable at detecting obviously AI-generated text, in the sense of having an actually extremely low false-positive rate, despite the oddly contrived counterexamples people come up with. But, and this is more important, what you can infer from a Pangram score is very limited when it comes to adjudicating whether a submission violates guidelines. You rightly note all of this, Justin. I think it’s fine to use Pangram in Substack to choose how you allocate your reading time, because there’s a lot of ‘slop’ there. And I think editorial desks could use Pangram *anonymously* as a first pass to identify potential slop — it need not be dispositive measure, and they can use their discretion to desk reject on that basis or not. But I totally agree that public accusations of identifiable people, whether supported by Pangram or not, are both epistemically risky and ethically unwarranted.
i think the problem of ambiguity is often missed and is crucial. llm ais are not simply language polishers, although they can do that if properly prompted. they are geared to, in a word, “generate” and many people use them to do just that.
text that contains obvious markers of llm use is (imo rightly) subject to suspicion precisely because the reader can’t know the degree to which ai was used to generate content rather than fine-tune human content. a given person might not have a problem with whether or not the content is human or machine generated on the philosophical level, but given the issues llms continue to have with misrepresentations of different stripes (see reference discussion above, or alex guerro’s earlier comments), i think a reader is more than justified to take clearly ai-inflected work less seriously at least initially.
the relatively easy solution here is to avoid using ai to write your work for you, and to be extremely judicious when using it to fine-tune, especially if you are a working in a language in which you aren’t fluent.*
*i mention the case of non-native writers because i’m sympathetic to that use case of ai. at the same time, relying on ai to adjudicate nuances in diction, syntax, and grammar across languages in research/professional work is fraught at best.
and then asks an AI to improve the writing style throughout the paper, the resultant paper will get a “100% AI generated” evaluation, too.
This is said a lot, and maybe it’s true. (I don’t have strong evidence to say it’s not true, and I’m not saying it’s not true.) But, I also haven’t seen anything more than bare assertions of it. Is this something that people are just taking to be obviously true, or is there some good evidence for it? (I guess that if the “improved” writing style just litters the paper with things that AI detection software is looking for, that might be so, but given that that style is not very good, it wouldn’t reall be “improving” the paper most of the time, and in some cases it’s not obvious to me why certain things would come back as “AI looking”.) If people have some strong, or even probative, evidence for this claim that’s more than just that it seems obvious to them, I’d be very interested to see it!
(On a somewhat different note, where I work students are required to submit “AI declarations” on their work. Many obviously lie on these things, but the point that interests me here is that many also say that they asked AI to fix grammar errors for them, and yet, their papers will often still be full of grammar errors. It makes me wonder what is happening, other than that they are not learning grammar on their own)
Good point. Happy to hear what people report. To repeat Matt’s question: how likely/often is an AI-polished paper flagged as AI-written?
It depends on what exactly goes on with the AI style editing. If it just changes a few sentences between passive and active voice, and replaces a few prepositions with more idiomatic ones, it’s unlikely to be able to set off any AI detector. But if every single sentence gets re-worded, with different grammar and vocabulary, then it seems very likely to set off an AI detector.
You can try the examples in the AI writing quiz I created for my students (https://www.kennyeaswaran.org/ai-writing-quiz/quiz.html).
For the “easy” ones, I asked it for something like “please write three paragraphs in the style of an AI executive, stating the AGI will be extremely powerful, that the only way to deal with this power is to share it broadly, and that democratic control will remain important”.
For the “hard” ones, I gave it a multi-paragraph prompt describing the argument made in each paragraph, and the name of the author (example below). This accidentally made it way too hard for humans to tell which is which, but I think the AI detector still identifies which is which if you just copy and paste them one by one. You can look at the prompt and the result and see if you think this is what someone imagines with giving a full paper and asking the AI to improve the writing style.
Hard prompt:
“I’m creating a quiz for my students in an AI literacy class, where they will try to tell human-written text from AI-written text. Ezra Klein recently published an opinion piece in the New York Times about why AI probably won’t lead to mass unemployment. I would like you to write a four paragraph passage in his style to include in this quiz. (I will reveal immediately afterwards which is correct).
The first paragraph mentions that economists are skeptical about predictions of mass unemployment. He cites a recent essay by Alex Imas, and quotes Imas as saying that any question about future economics begins with identifying what is scarce.
The second paragraph goes through the history of humanity, initially with scarcity of food, then after agriculture other goods become scarce, then after industrialization knowledge becomes scarce, but there is a fear that AI will end knowledge scarcity.
The third paragraph notes that there will still be scarcity – we shouldn’t ask what AI can do, but rather what we won’t want AI doing.
The fourth paragraph notes that in econometrics, we have discovered that rich people want more from other humans, not less. He quotes another sentence from Imas about how people shift their spending, and discusses that clothing, food, medical care, education, and physical training all shift towards more personal aspects.”
Thanks Kenny, these examples are helpful — and good idea for a quiz! (I scored 6/6, though I benefited from knowledge of the human author’s voice in some cases, as well as of Vaticanese.) Pangram flags three of the hard outputs as 100% AI-generated, except the Dario Amodei “Labor disruption” one, which it identifies as 100% AI-assisted. For what it is worth, that is the one with the longest human-written prompt, with the most detail about what should be included in each paragraph.
I don’t see these examples as examples of writing a full paper and then asking an AI to help you with the writing style. You have an original human text, but then you’ve summarized key aspects of that text that rather than given it directly to an AI to improve.
With that said, there is also significant human contribution in these examples, and these AI texts read to me as more epistemically valuable than AI writing with no significant human assistance. I myself would want to see a lot more human involvement for a philosophy paper, though (beyond just scaling this up to 8000 words or whatever). Even for OLH, which has a more lenient AI policy than most I’ve seen, I suspect that a paper written entirely in this manner would violate the prohibition on “substantial, major uses of generative AI to be passed off as the author’s own research”.
Thanks, Kenny. That’s useful and intersting..
But if every single sentence gets re-worded, with different grammar and vocabulary, then it seems very likely to set off an AI detector.
I guess that, if this is what’s meant, than I’m glad to have an AI detector pick it up, because I don’t really think it’s the original author’s work anymore – at least not their own work – and so it needs an explicit declaration saying how AI helped. In many of these cases, I find myself thinking that, if a human did what the AI did, would we think that they human needed to be clearly acknowledged, or maybe even included as a co-author? In this case I’d think they’d need to be included as a co-author, and so not making the use explicit is a form of academic or research dishonesty.
(Here’s an example: A long time ago I co-wrote a paper for the APA Newsletter on International Cooperation on “philosophy in Russia” with a Russian philosopher who was a friend of mine. What we did was that I came up with a bunch of questions about philosophy in Russia that I thought would interest APA members, we discussed those, and she came up with some more. She wrote most of the substantive answers, and then I re-wrote them in more normal sounding/grammatical English, and then we talked about that and revised it a bit more. We published it under both our names, but with her as the “lead” author, as I thought most of the substance was her work. I think you could do basically what we did with either one of us being replaced by an AI. But, if either one of us then published that as just our own work, I’d think it would be a case of academic dishonesty. That’s a bit different from the example of yours I note above, but I think it’s along the same lines.)
I did a series of tests, and found two ‘results’ that I think are interesting to share.
First, I gave ChatGPT a handout of a class I’ve been teaching to my graduate students. The material is pretty original, not on a standard topic on which a lot has been written. I asked ChatGPT to write this 2000-word handout into a 7000- to 8000-word academic paper. I then read through it to remove any indicators that the paper had been written out by GenAI, and gave it to Pangram. Result: 100% AI-generated.
Second, I gave ChatGPT the powerpoint of a keynote presentation on my own research, and asked it to write it up into a full paper. Fully my own research. I then ‘cleaned’ the AI-written-out-paper from a few weird indications that it was written up by ChatGPT. Pangram reported this as 100% AI-generated.
My conclusion from these two tests is that whatever explains a “100% AI-generated” score in Pangram, it can still be the case that the ideas, the reasoning, and the claims, are all still the author’s own.
Thanks. I guess I’m happy to have both of those things flagged as 100% AI generated, as that seems accurate to me. No one suggests that it tests if the _ideas_ are AI generated. (How could it do that?) But, it seems completely accurate to say that the _writing_ in those cases is AI generated, and I think it’s reasonable to flag that, and think that cases you consider are ones where, if it wasn’t disclosed, I’d think it was misconduct.
Justin writes:
“ If someone asks Claude to write them a novel article about a philosophical problem, not even giving it a thesis to work with but asking it to come up with one itself, and that’s the extent of the human’s contribution, the resultant paper will get a “100% AI generated” evaluation. And if someone spends a year researching, writing, workshopping, and revising a paper, and then asks an AI to improve the writing style throughout the paper, the resultant paper will get a “100% AI generated” evaluation, too”
I’m not sure that’s correct. I’ve been experimenting a fair amount recently both with Pangram, and with using Claude as a copyeditor. Some findings:
I take from this (tentatively) that you can use Claude as a copyeditor or for translation work and still not flag AI detection. (It is a very good copyeditor, substantially better than any human I’ve worked with.) If, on the other hand, you give it an article and just say, “make it better”, you’re going to get flagged.
Thanks for this, David. It’s the sort of stuff I was interested in in my question above. (If others have links to even more systematic studies, I’d be interested in that, too!)
My experience is consonant with David’s. For example, I just had Pangram analyze a translation I had ChatGPT do of Plato’s Apology last year when I was teaching it (I did this as an experiment to see how good it would be — I ended up using the Grube translation instead). It returns it as 94% human-written, and much of the AI portion it flags is ChatGPT’s introductory “Background on Reading” it inserted at the beginning (I told it to do the translation with the aim of being able to use it in an undergraduate classroom). It only flags four short passages as AI-written, and one as AI-assisted. You can see the report here: https://www.pangram.com/history/f7ce553d-015e-4fb5-a430-07eb38fbe8f5?ucc=pJRkjTslHhl I also had Gemini do a translation, and had Pangram do a report on that; the results were similar (88% human-written): https://www.pangram.com/history/958daae1-d6d6-4624-9466-31313d8faf6d?ucc=pJRkjTslHhl
If others have examples of Pangram flagging human-written essays as 100% AI-generated after one “asks an AI to improve the writing style throughout the paper”, I hope they can provide them (with links to the Pangram report). I do not think I have seen any examples of this.
Take a look at the prompts I gave for the “hard” questions in the AI writing quiz for my students: https://www.kennyeaswaran.org/ai-writing-quiz/quiz.html
These have at least the structure of the argument for each paragraph in the prompt, though there’s still room for the AI to word in how it wants. I’m pretty sure the results end up flagged as 100% AI, though I don’t know if I’ve tested them all. (I noticed that some of these are still short enough that if I copy both the human original and the AI imitation and paste, then it detects the entire thing as AI-written, or at least did before the most recent Pangram update.)
For what it’s worth, David’s experiments inspired me to play around with Pangram on a free trial. As others have, I fed it a (drafty!) work in progress. The 5% that it flagged as AI-written or -assisted had not, in fact, been touched by AI. However, the flagged sentences were all perfectly grammatical and either repetitive or a slightly conceptually confused/ otherwise amiss.
haha, great. I am actually nervous to do this experiment myself. one reason is that I don’t want to subject my piece to machine evaluation. another reason is that I enjoy the naive confidence that my style is fine, and don’t want to risk doubting everything if it turns out that it is flagged as AI haha
Thanks David! Very interesting observation! These cases surprisingly parallel my own experiment not about AI detection but about subject feeling of ownership. Everyone: let me know if your experience is the same!
Case 1: “Several times I’ve asked Claude to go through a document and identify issues of spelling, grammar, style and structure, and make recommendations. I’ve gone through those recommendations and implemented the ones I thought were sensible.” In the end, I feel 100% ownership of the resulting paper.
Case 2: I wrote a first draft by myself, and then went through several rounds of revision like normal. After settling down a final version, I asked Claude to rewrite according to its own stylistic judgment without changing the content. (I read the result, making changes if *necessary*.) My sense of ownership of the resulting paper is significantly reduced.
Case 3: I provide the big ideas including thesis and argument ideas section by section and asked Claude to write the whole paper. (I review the paper and make changes if necessary) I do not feel sole ownership of the resulting paper.
Surprisingly, the sense of ownership between case 2 and 3 are not significantly different.
A random thought: perhaps some journal should just experiment with a policy that articles can’t be flagged as >10% AI by Pangram, without prejudice to whether it’s perfectly reliable. It’s not as if people can’t check their own work before submitting!
The longer-term complication is that eventually, inevitably, all philosophy (like almost everything else) will be “100% AI generated.” The role of human intelligence will no doubt be shifting from its current creative model to more of a curatorial and connoisseurial one. We can lament this all we want, but no one continues to lament quill pens, cuneiform wedges, or bards memorizing 74 hours’ worth of poems. The idea that human intelligence is fundamentally about creating thoughts and words ex nihilo in the face of blank paper will, I think, prove to be a timebound model. Disclaimer: I DO NOT intend to be an early adopter of the new methods, because the first scholars to do so will be publicly crucified. But there are already incredibly useful things that can be done with chatbots without having them contribute to your actual writing.
“[B]ut no one continues to lament quill pens, cuneiform wedges, or bards memorizing 74 hours’ worth of poems.”
This—both literally and in terms of the less specific message it expresses—is wrong.
Give an argument.
Sure.
First, I personally know people who lament losing specifically these things—with the exception of the cuneiform (I guess ya got me there).
Second, you can easily find others online who lament losing them.
Third, it’s obvious that there are many things like these which people do lament losing: there are essays about the value of using a pencil rather than a word document, there are those who bemoan the decreased use of physical books, there are people who think phones (even dumb ones) make us more anxious than older forms of communication, that the convenience of planes and trains and automobiles were gained at the loss of emphasis on local community, there are primitivists who reject this sentiment that “no one laments losing” various relics of the past, and so on.
Fourth, the examples are obviously cherry-picked to sound ridiculous. Replace cuneiform with, say, postcards, or small towns, or old-fashioned phone calls (rather than texts), and the examples become both (a) more relevant to the AI debate—which concerns things we might lose *now*, rather than things from the distant past—and (b) far less clear-cut in their normative implications.
Fifth, and less literally, what people often lament losing in these cases are not the literal tools themselves, but the social and psychological realities that attach to them: the sense of intimacy, or community, or whatever, which the appeal to highly concrete examples like the quill pen clearly misses.
Sixth, and finally, I don’t see why I’m even required to offer these arguments in defense of my claim, given that it was a response to a similarly—hell, far more—contentious claim which was itself given with no support.
I think you may have missed Graham’s point. Sure, a vanishingly small minority of people will grieve forever. But in the cases he mentioned, culture has moved on.
I think you might be missing the intended implication of his point. His examples are preceded by the phrase “we can lament this all we want, but…”. What do you think is the function of this phrase?
But surely neither you nor Graham need to be told that “culture moving on” is completely irrelevant to rightness or wrongness, or any value judgment of anything at all.
The point I would make is that rightness and wrongness aren’t always what is at issue. Was it “wrong” when militaries turned to the tank after realizing that trenches plus machine guns led to nothing but immobilized slaughter in WWI? Analogously, I don’t see the Egyptian debate in Plato’s Phaedrus between oral culture and writing to be primarily moral, even if the participants do try to make moral claims. AI writing, once normalized, will lead to plenty of nostalgia for the specific writerly culture we’ve developed over the course of centuries/millennia, but is also likely to open up a new intellectual world whose contours we can just barely begin to see from where we are. There are all kinds of intellectually stimulating things we can do with LLMs, right now, that were simply unthinkable pre-2022.
“The role of human intelligence will no doubt be shifting from its current creative model to more of a curatorial and connoisseurial one.”
I suspect that may be true for some significant amount of academic philosophy, and for many other professional fields. But there are also people who are driven to create, and who are not going to cease creating if/when AI gets much much better. (I know; I’m one of those people.) A more accurate statement would go something like this: An economic system dominated by AI will only offer low or no pay for human creativity; however, that doesn’t mean that creative people are going to stop creating, it just means that creative people will get paid even less than they do now.
Really, though, this isn’t saying much beyond what was said long ago in W. J. Baumol and W. G. Bowen (1965), “On the Performing Arts: The Anatomy of Their Economic Problems,” The American Economic Review 55 (1/2), pp. 495–502. Indeed, a significant portion of the current kerfluffle around AI can be explained by the Baumol effect.
People will still be able to write everything from scratch if they want, just as they can still refuse to use typewriters and word processors. Some people do refuse it. But speaking only for myself (and I do love writing, by the way) I think it would be a better use of my time if I could have AI write multiple book drafts for me, then go through and correct and improve style, connection between thoughts, and so forth. I think we have come to overly romanticize one specific moment in the writing process: namely, the sometimes tortured brainstorming that turns a blank page into words authored solely by ourselves. But I think that’s not the only moment or even primary moment when thought occurs.
I don’t think the creative people will be paid *less* than they are now – the main observation Baumol makes is that the number of people being paid to create by the means that haven’t made use of productivity-enhancing technology goes down, while their pay goes *up* (though it often increases by less than the pay of the people who do use productivity-enhancing technology).
In the age of IKEA, the few remaining artisanal furniture makers make a lot more than the artisanal furniture makers of the past. But most people who insisted on remaining artisanal furniture makers lost their customers, and only the few who could command this higher pay have remained in business.
Two purposes of philosophical writing are (1) discovering ways of approach, arguments, and conclusions about the topics investigated, and (2) credentialing candidates for jobs, tenure, and promotion. AI used in any manner does not necessarily interfere with the former (and will likely enhance it), although it clearly creates challenges to the latter purpose. Perhaps, once AI content increases in reliability and its use becomes universal, differences in the quality of approaches and arguments will become a measure of human creativity in the use of AI. If this happens, no one will bother to accuse anyone else of using AI, since almost everyone will use it and the only question pertaining to the second purpose above will be how skillfully someone used it. One analogy: I enjoy doing arithmetic in my head, and sometimes my results are more accurate than someone who is sloppily using a calculator, but I don’t think using a calculator is “cheating” and I hope my accountant will use a calculator. Second analogy: the use of research assistants by a principal investigator. No one thinks a PI is “cheating” by using assistants. The difference is that while human assistants require credit, AI does not have that human need for recognition.
If philosophical writing wasn’t done by named authors who took credit for what they wrote, and if the field didn’t depend for its existence on its being so, then sure, “no one will bother to accuse anyone else of using AI” once it’s good enough. But it is, and it does.
Philosophers don’t consider spell-checking cheating, because what the author takes credit for isn’t the good spelling, it’s the good philosophy, just as a physicist doesn’t take credit for the long division and square rooting she does (via calculator) in a physics article. But she does take credit for the physics. If she used AI for that, and took credit for it, we’d have a name for this.
Generally we know what we take credit for in a philosophy article because when someone helps us with any of that stuff, we — or a lot of us — thank them in a footnote. As in, “thanks to So-and-So for this example,” or “thanks to such and such for reminding me of this objection,” or “that unique phrase comes from so-and-so, or was first coined by Sprigge,” etc. etc.
You may be underestimating the degree to which mathematics – not just long division* or square roots, but symbolic manipulation – is now automated in physics. Calculations in general relativity that used to take months – and used to take serious skill – now get done by Mathematica and the like in minutes. And this significantly precedes LLMs.
* I don’t think I’ve seen long division even once in a physics paper.
OK, point taken. But a little beside the point, right? Remember I was responding to Zapnu’s claim that his use of calculators doesn’t make him a cheat (hence LLM use doesn’t eithr). My reply was it wouldn’t make a physicist a cheat either, but not because it’s an accepted technology. It’s because the calculator isn’t doing the part of physics she’s credited with in the first place, the core physics of it. Substitute “calculations” for “long division” (or whatever) and “Mathematica and the like” for “calculator” and it’s the same point, more or less.
My point is that the ability to do long symbolic calculations without invalidating errors *was* taken as part of “core physics” until relatively recently; that sort of skill was a valuable and sought-after aspect of being a theoretical physicist in the recent past but to a large extent no longer is. Mathematica has changed what counts as core physics.
We know what we *currently* take credit for in a philosophy article, because we know what are the unique and difficult things that one’s participation in authorship contributed. But as David Wallace mentions, technology keeps changing exactly what that is.
Miryam Naddaf and Richard Van Noorden, “Is AI Detection Good Enough Now?” Nature v656 (27 Aug 2026) 808-811 doi: https://doi.org/10.1038/d41586-026-02569-3
https://www.nature.com/articles/d41586-026-02569-3?WT.ec_id=NATURE-20260827
Is, unfortunately, paywalled though it may be accessible via some libraries, and most of it is available at the URL above, along with its cited references. It contains a brief summary of some of the strengths and weaknesses of Pangram 4 (July 2026 version) and GPTZero (owned by Superhuman, formerly known as “Grammarly”). Some interesting information about recent improvements:
“Pangram 4, the software launched this July, has changed evaluations in part because it divides text into more fine-grained sections. Whereas the older version analysed segments of some 200–300 words — meaning that a 1,000-word post might end up being divided into only three sections — the new version can analyse chunks as small as 30–40 words. [Pangram developer] Spero says that Pangram 4 better distinguishes between a light AI polish — which usually retains a human-authored label — and heavy AI edits. “It takes major AI input like fully- generated sentences or major rewrites to trigger Pangram 4,” he says. GPTZero analyses texts slightly differently. Like Pangram, it divides them into segments, but at the end, it gives a score that represents its confidence about the whole text, rather than a breakdown of the results of each segment [fn5]. Paid users can see which sentences contributed most to GPTZero’s assessment of a work as probably human, AI or mixed.”
On the general issue of trade-offs between false positives and false negatives, there is a very helpful discussion in:
Tafari Mbadiwe, “Algorithmic Injustice” The New Atlantis (Winter 2018),
in its final sections, “Confusing Standards of Reliability” “The Fairness Paradox,” and “Don’t Blame the Algorithm.” https://www.thenewatlantis.com/publications/algorithmic-injustice
“That is a serious accusation! It can be quite damaging. Given how human psychology works (and how internet searches work), even if the author is shown to not have acted wrongly, negative connotations may linger, damaging the author’s professional prospects for years to come.”
The same applies to other accusations, which also have similar fallibility concerns, but which I’ve seen shared on blogs and social media fairly indiscriminately, often on one agent’s (not necessarily a machine agent’s) say-so.
Maybe it took the AI case to drive the point home, but I hope the lesson will be applied consistently going forward…
I think if journals adopted rules requiring any AI use in any capacity to be declared in the paper, this might avoid a lot of issues. I think a lot of people who run stuff through AI detectors do so because they read something that sounds like AI because it was in fact written in part by AI. That may be beside the point with respect to the journal’s norms etc., because use of AI for (e.g.) translating from another language or fixing grammar may not amount to what the journal dubs “AI misuse or malpractice.” But if the AI use is not disclosed, some people still want to uncover it for various reasons. If AI use were always disclosed, people would not have to run things through an AI detector. They could just read the disclosure statement.
To have a norm that everyone is disclosing AI use requires trust, namely that people won’t lie about it. And we might think that right now the norms about AI use are too inchoate, and some people might want to use it for purposes that they judge to be licit (and that the journal judges to be licit) but which others judge to be illicit, and they won’t want to be negatively judged for this use. So, they’ll want to hide AI use, rather than disclose it.
I don’t think there’s any solution to this problem, so the situation will always be non-ideal. But I think AI users are better off if the norm is “everyone discloses it and witch hunts are rarer” rather than “everyone hides it and if someone correctly thinks your writing sounds like AI they launch a witch hunt and now there’s a cloud of suspicion that follows you for the rest of your life unless you disclose it after the fact at which point you look guilty for having tried to hide it.”
Of course, the norm could also be “no witch hunts,” but (like the AI users) the witch hunters don’t think they’re doing anything wrong, so I don’t know how much luck we’ll have convincing people not to go on witch hunts. Asking AI users to disclose is just asking them to say publicly that they are doing something they think is acceptable, whereas asking witch hunters to stop is asking them to stop doing what they think is acceptable. Presumably it’s less of a pill to swallow to admit you are doing something you think is acceptable than to stop doing something you think is acceptable.
Also, with respect to the witch hunts, although I’m not about to launch any myself with respect to journal articles, I understand the sentiment behind them. I don’t know about other people, but when I read at least some AI writing it has a very uncanny effect on me, like when you call customer support and you talk to someone for a little while and then it dawns on you that you are talking to a robot. I think it’s like what psychologists call derealization, which Wikipedia calls “an alteration in the perception of the external world, causing those with the condition to perceive it as unreal, distant, distorted, or in other ways falsified.” When I come to believe I am reading a robot’s output (altered however much you want by human input at the beginning or the end or both) it does give me a feeling that this thing is unreal, distant, distorted, and (if the AI use has not been disclosed) in other ways falsified. It’s a very weird feeling and I can understand how someone experiencing it might be inclined to launch a witch hunt in an effort to cope with the feeling.
Of course, there’s always the possibility that someone just naturally writes like AI. I haven’t done much research into Pangram. This Substack post makes it sound like you have to make a Herculean effort to naturally write a “100% AI” Pangram text by hand, but maybe some people just naturally write in ways that seem AI-esque to Pangram. It’s hard to know! I would like to have more data. One way to get more data would be if everyone disclosed all AI use in their journal articles!
Would a disclosure placate you? I understand you’re not willing to launch the hunt yourself but you seem willing to support it — at least you didn’t say you don’t support them and you said you understand them. Forgive me for being skeptical that generalizing disclosures would suddenly make you guys leave people alone.
PS: you should read the OLH guidelines. They’re well written and quite sensible (IMO) and they explain carefully why they don’t require disclosures below a certain threshold.
Are you referring to these parts of the guidelines?
I am not very convinced that it is unfeasible for authors to know and declare every instance of generative AI use prior to the submission of their research. I think you know whether you use these tools or not. This does not happen while you are asleep. Even if you forgot, these tools typically keep logs of your use, don’t they? So you can review the logs of the tools you use to remind you if you used them. And declaration seems effortless once you have the knowledge.
I agree that there is a potentially discriminating factor with publicly available declarations. But I think the costs of hiding these things in order to hopefully stave of discrimination is higher than the gains. And I think people who tend to wrongly discriminate will often find all sorts of things to discriminate based on. We don’t hide information like who has published an article or where they work, even though people discriminate on these bases, and I’m not sure we should hide information like whether any parts of the article were written by AI either.
Sounds like you disagree with OLH journals’ policies. That’s fine! You don’t have to submit there or read their journals. And you’re also free to criticize the policies. Let a thousand flowers bloom.
I think I’ve lost the thread of the conversation a bit. Your previous comment asked me to read the OLH policies, with the implication being that this would be relevant to something I had earlier written. I perhaps would see the error of my ways etc.
Now you are saying “ah, you disagree with OLH? Fine, but everyone should get to do what they want to do.” I mean, yes, of course, but my point is that people ought to want to do what I think they should do. If OLH gets it wrong then so be it, but my earlier point is that they are getting it wrong, and your response here is just pointing out that I disagree with them. I know I disagree with them! That was my original point!!
Or, in other words, as I noted originally, the reason not to let a thousand flowers bloom is that it undermines collective trust in the discipline that AI use will be disclosed in ways that will prevent people who prefer not to read AI from feeling like they are being misled and deceived. The OLH policies not only fail to respond to that worry but actively promote it by carving out the possibility of AI use that is not disclosed (and they do so with the dubious justification that it’s not feasible to disclose all AI use, which as I noted above seems false on its face). But it’s no use to say “you don’t have to submit to OLH journals or read those articles.” I might as well say “you don’t have to use AI to write anything.” That’s true but my saying this would ignore how damaging it is not to use AI to write for people who (say) did not grow up speaking English, and your saying this ignores how damaging it is to (e.g.) be a political philosopher who neither publishes in nor even reads stuff in one of the premier political philosophy journals.
Now, merely pointing out difficulties is obviously not dispositive. Even in light of the above-mentioned difficulty I’ve still come down on the side of AI disclosure. And perhaps you come down on the side of bifurcating the profession into AI-disclosing and not-AI-disclosing journals and people and so on, in which case, fine, whatever. But again, my point is just that at this point in the conversation I’ve lost the plot with respect to how the OLH policies are supposed to relate to my original comment.
The OLH policies distinguish between cases that require disclosure and those that don’t. You want a blanket disclosure requirement for any AI use whatsoever. I asked you what you thought of their rationale. You don’t like it. That’s fine. Do you want me to publicly denounce policies that don’t align with your preferences? I’m confused how simply admitting we disagree about what set of policies should exist is losing the plot.
As noted above, I thought your “have you read the OLH policies?” post was meant to suggest there was some thing in my previous post that I would rethink had I been properly apprised of the considerations the policies adduce. For the reasons also noted above, it does not seem to me that there are any relevant considerations: the OLH policies seem to me to rest on obviously false considerations about the feasibility of disclosing all AI use, and they seem to give rise to precisely the problem I was pointing out to prior to your raising the OLH policies in the first place. So I didn’t find anything in the policies that suggest I ought to rethink anything I had said in my posts prior to your raising the policies. This is why I am not sure why you raised the policies.
It’s true, as you say here, they they “distinguish between cases that require disclosure and those that don’t.” But as I pointed out in my very first post, this distinguishment gives rise to various problems, and for this reason I don’t think we should distinguish cases like this. So your adverting to the fact that the OLH distinguishes cases doesn’t seem to me to give me a reason to rethink my approach: indeed it just seems to highlight why my approach is better, given that the OLH policies are the occasion for the problem we have in front of us in the first place. (I have said this above too.)
You ask “do you want me to publicly denounce policies that don’t align with your preferences?” With respect to the points I’ve been making in this thread, I’m just trying to explain my view. I hope my view is convincing, and so I hope to respond adequately to objections to my view. What you wrote seemed to suggest that OLH policies contain some sort of objection to my view. For the reasons noted above, I haven’t discovered anything like that in the policies. So I worry I’ve missed something, and I am hoping you can point out what that is.
But as of right now, I can’t find anything in those policies that seems like a compelling objection to my view. Indeed the most salient parts of the policy (almost the entire content of the policy) consist of two things: 1) the limitation of disclosed AI use to certain cases, leaving open the possibility of undisclosed AI use which therefore CAUSES the very problem I think would be solved by my policy, and 2) a dubious (indeed, it seems to me, obviously false) justification for the limitation. So at least at this point in the conversation it’s mysterious to my why someone would point to the OLH policies as potentially having something that would cause me to rethink my view. Hence my confusion, my feeling of having lost the plot, etc.
One thing that might help is if you tell me what you think about 1) my claim that the OLH policies rest on a facially incorrect claim about feasibility or 2) my suggestion that the OLH policies have caused the problem that exists in this present case by allowing for undisclosed AI use and thus giving people who think a paper is written by AI an incentive to launch a witch hunt (whereas if the OLH policies required full disclosure, AI use would already have been disclosed, or people would not have to launch a hunt because they’d know no AI was used). Responses to either or both of these might help me see how the OLH policies have anything that might amount to an objection to my original point.
Hi Daniel,
First you should step back. Our initial conversation ended a long time ago. This sub thread was a mere PS inviting you to read the policies if you had not—it wasn’t obvious from your comment if you had, despite its length. The rationale of OLH says more than you insist but because I don’t especially enjoy reading your “novels,” as you call them (they’re more like lectures if you ask me), I let it go when it became obvious you were not buying their justifications. Simple as that. Now you can keep on pretending I’m crazy or stupid but this is going to sound like a bit of gaslighting. Trust me, I’m following the plot. I just don’t think it’s as interesting as the length of your lectures presuppose.
Daniel has said repeatedly that he worries *he* is losing the plot, not you. And, frankly, I’m with him—your responses have been needlessly aggressive (“Forgive me for being skeptical that generalizing disclosures would suddenly make you guys leave people alone”), dismissive (“First you should take a step back”), and difficult to follow (you very much did seem to offer the guidelines as a counterpoint to Daniel’s original comment, only to then say it’s okay if he disagrees with them). His comments have been perfectly respectful and calm from the start, and the accusations of gaslighting seem unfair and disproportionate. If you don’t want to address his questions/objections, just don’t reply.
Meme,
This fixation on a mere point of disagreement — repeatedly clarified: I find the guidelines, on why a blanket disclosure policy is misguided, sensible; he does not — as something that somehow requires continuing glib lectures on his part and explanations on mine is getting frankly weird, as is this idea that my skepticism about what “you guys” would do is “agressive.” What’s with the language policing? How is an invitation to zoom out (or “step back”) dismissive, when the point is simply that Daniel was losing sight of the fact that I wrote a quick comment as a PS. Daniel’s claim he was losing the plot was to me just way to say I didn’t make sense. I didn’t find that particularly kind. I’m running out of ways to say it’s okay that Daniel hasn’t found in OLH guidelines what I did—a thoughtful rationale for not requiring blanket disclosure.
More importantly, and this is a point I make here consistently: I have very little patience for pseudonymous accounts. The asymmetry in accountability bothers me. If you want authors to take responsibility for what they write, as I do, then do so. Now, because I think we’ve hit diminishing returns on the point of this conversation, let me bow out.
I think that bowing out and taking some time away from this thread would probably be a wise move.
Yes, Meme!
Why should plagiarists by default be “left alone” to perpetrate their plagiarism? Academic misconduct is something we should all care about.
Why are you talking about plagiarism?
Because AI abuse is tantamount to plagiarism.
You’re smuggling a lot of assumptions here. For one thing the author in the PPA case was cleared of all misconduct, so you should probably stop insinuating to the contrary. You want to have a different conversation.
PP case not PPA (but frankly applies to both)
Let’s recap.
Here’s what you said:
“Would a disclosure placate you? I understand you’re not willing to launch the hunt yourself but you seem willing to support it — at least you didn’t say you don’t support them and you said you understand them. Forgive me for being skeptical that generalizing disclosures would suddenly make you guys leave people alone.”
So it was you, not me, who elected to make the conversation a general one about what conditions would make “you guys” (presumably those who invite scrutiny towards suspicious submissions) “suddenly leave people alone” (not the author of the PP piece specifically, but rather “people”).
Sorry, but you can’t frame the conversation in such combative, broad, and factionalistic terms and then just retreat back into the specifics of the situation when it suits you.
Here is my point: We on the “you guys” side are concerned about academic misconduct. AI abuse constitutes academic misconduct. We invite investigations in order to identify and eliminate academic misconduct. The author in this case was cleared, and that’s fantastic. But the fact that the investigation happened was a good thing, as would be a disclosure requirement.
You are the only one in this entire thread to be making it about plagiarism. The whole point of Justin’s post is that not only was the author cleared of misconduct but we should be careful when leveling those accusations. When I’m asking if people like you and Daniel would be placated by disclosures I’m asking if that’s sufficient for you. You seem to think not.
So you’re going to overlook my comments on your rhetoric. That of itself is not a problem, of course, but worth pointing out.
Do you deny that out-and-out AI abuse constitutes plagiarism? IOW, if the author had been found to have committed the offense of which they weres accused (representing as his own work a piece which was not his own) do you dispute that this would be tantamount to a conviction of plagiarism? And if you agree with that, then what’s the problem with the word plagiarism? What is under discussion meets the textbook definition of the word. Again, I am not insinuating that the AA author committed this act, but why is the terminology wrong?
And then there’s this: “I’m asking if people like you and Daniel would be placated by disclosures I’m asking if that’s sufficient for you. You seem to think not.”
And now you’re quite far out on a limb. I’ll refer you back to this line from my previous reply: “The author in this case was cleared, and that’s fantastic. But the fact that the investigation happened was a good thing, as would be a disclosure requirement.”
Let me make it explicit for you: Yes, I would be placated by a disclosure requirement. I also think it would be a good idea to investigate thoroughly, on the basis of that requirement, submissions which can be reasonably suspected of failing to include disclosures when they should have.
I don’t have to comment on what you perceive do to be my ‘rhetoric.’ You could answer a simple question, the only one pertinent to this conversation: in the PP case, and hypothetical similar cases, if the author had declared their AI use (and so presumably the journal was aware of and fine with it), would you still complain? You have now clarified you would not, that’s great. Again, you’re the one making it about plagiarism. It should be clear by now that I do not consider what the author did to constitute plagiarism.
No you don’t, as I recognized in my previous comment, but observers are free to make what they will of that omission.
The prerogative is also yours to split hairs on the “plagiarism” moniker, but, as I illustrated to David below, your perspective on this is far from irreproachable.
“Do you deny that out-and-out AI abuse constitutes plagiarism? IOW, if the author had been found to have committed the offense of which they weres [sic] accused (representing as his own work a piece which was not his own) do you dispute that this would be tantamount to a conviction of plagiarism?”
I’ll bite: no, this isn’t plagiarism. Plagiarism is presenting the ideas of another person as your own. It’s bad in substantial part because it dishonestly steals credit from that person. AIs aren’t (at present!) persons in the fairly morally thick sense relevant here. It also involves taking unchanged or lightly paraphrased material from a source that the supposed author had no hand in creating, which is a relatively black-and-white thing compared to the shades of gray in AI authorship, which range from “give the AI a prompt, publish the unaltered result” to much more extensively human-involving activity.
This is neutral as to whether AI use in the sense you have in mind is bad, or even whether it’s as bad as plagiarism. But I think we do better to give it a name of its own (“AI authorship”, would be my suggestion) and then discuss its moral status, rather than trying to preempt discussion by annexing it to a distinct form of activity that we all agree is bad.
Hi David,
Here’s how Oxford defines plagiarism:
“Presenting work or ideas from another source as your own, with or without consent of the original author, by incorporating it into your work without full acknowledgement. All published and unpublished material, whether in manuscript, printed or electronic form, is covered under this definition, as is the use of material generated wholly or in part through use of artificial intelligence (save when use of AI for assessment has received prior authorisation e.g. as a reasonable adjustment for a student’s disability). Plagiarism can also include re-using your own work without citation. Under the regulations for examinations, intentional or reckless plagiarism is a disciplinary offence.”
(https://www.ox.ac.uk/students/academic/guidance/skills/plagiarism)
I happen to agree with the folks at Oxford that the personhood, or lack thereof, of the source misrepresented as one’s own is utterly irrelevant to the wrongness of the deed.
Sure, but the definition you’re quoting explicitly refers to AI – which is just to say that whoever wrote that definition has followed your strategy of expanding the definition of plagiarism to include AI. I think your strategy is mistaken and I think whoever at Oxford wrote that definition is mistaken.
A disclosure would absolutely placate me, yes. It would allow me to either skip reading the AI stuff until I am in the proper mood, and much more importantly it would remove the weird uncertainty of not knowing if I’m reading human-generated output and of therefore mistrusting my own capacities tor recognize whether I am seeing something human-generated.
As for leaving people alone, I think there are different sorts of being left alone that different sorts of people care about. I and many others would prefer to be left alone rather than to have AI-generated stuff snuck in front of us with no warning, people who like generating AI stuff might prefer to be left alone to generate AI stuff, etc. I don’t know how much clearer I can make it that I don’t want to launch witch hunts against people who write academic articles with AI, since I already said that I am not going to launch witch hunts, and indeed I proposed a way to reduce the likelihood of witch hunts in part because I don’t think we should have witch hunts. But if it helps, I’m happy to reiterate the point that I don’t think people should go on witch hunts against writers of academic articles in contexts like this.
In any case, your skepticism about whether generalizing disclosures would help doesn’t bother me. Let’s just have the disclosures, and you and I both will get to see if your skepticism is borne out!
Glad to hear. Thanks for clarifying. As for being left alone (from reading AI), “I understand the sentiment”, as you put it, but you’re equivocating. That’s of course a completely different kind of being left alone. Refocus on the case at hand. Leaving the author alone and leaving you alone mean completely different things. Readers were never entitled to know the complete conditions of production of articles. That said, I really do share the sentiment. As “pro-AI” as I can be, I find AI prose extremely grating. But I can choose to eschew it on my own without disclosure. I just stop reading when it sucks. I used to do that before AI. A lot of writing is just bad!
I don’t really use AI to write, but if I did I would be really hesitant to disclose my AI use I’m not sure I would advise other people to either. The AI-phobia rampant on this website is enough to deter me from doing so. I think people would scorn my paper, and I would have trouble getting a job (given the numbers, there’s going to be one anti-AI tablethumper on every hiring committee who will hate it).
Also, I really don’t trust that whatever my AI disclosure disclosed would be interpreted accurately by people who read that disclosure. The modal philosopher seems not to have a grasp of what these models do. I see people on this blog confidently pontificating about their evils, and pronouncing never to have used them in the next sentence. Or saying that the models hallucinate references but by the way they’ve only ever tried the free versions. Etc.
I understand not wanting to disclose AI use. But you can’t always beat the tablethumpers by hiding it, so hiding it is not a perfect solution. Indeed, there is never going to be a perfect solution here. I think if you consider all the likely possibilities and run the numbers, so to speak, you’ll realize that AI users would be collectively better off if everyone disclosed everything (as would non-AI users).
This is easy for me to say, since I don’t use AI to write papers, so almost none of the costs fall on me in any of these scenarios. So I acknowledge I might be wrong about this. But we have in front of us a very good example of how your approach can backfire. If the paper in question was written with AI (about which I have no opinion, having not read it), I think the author is worse off for not having disclosed it ahead of time, which would have precluded the witch hunt. And if the paper in question was not written with AI, the way to have avoided this whole situation in the first place would be to have had a norm according to which people always disclose any AI use.
To illustrate a contrasting example of a case that fits what I think should be the norm, in an edited volume I recently published in, one of the editors wrote in the first endnote to their introduction “AI tools were used, under the author’s direction and following ethical research practices, to assist in drafting, editing, and organising content for this work.” The editor is not a native English speaker and anyone who reads this endnote and gets angry is being unreasonable, I think. If these sorts of endnotes (or, please, let them be footnotes) were the norm, I think AI use would lose a lot of the stigma.
I really disagree. The young scholar who (at least as a matter of public record) authored the publication in question would be much worse placed, in expectation, in their career if they had said they wrote the paper using AI. Doing so would have vitiated any advantage that this otherwise respectable journal might have conferred upon their CV, and might even have counted as an overall negative. Just because one person was (or wasn’t) caught doesn’t tell you anything without the numbers of people who don’t get caught / the expectation of getting caught and in any case as we can see this person was cleared of wrongdoing. The fact is, an AI- assisted publication is not much of a publication at all. And except perhaps in certain high profile cases or ones that are capturing the public interest as scapegoats, the penalty for getting ‘caught’ using AI without disclosing it will tend to be the same as not publishing the paper at all, particularly since there is almost always deniability.
If wrongdoing amounts to using any AI at all, it is false to say that this person has been “cleared of wrongdoing.” Some of the comments in this thread (including the very first one) point this out. The only thing this person has been cleared with respect to is a violation of the journal’s AI policy, which explicitly does not cover using AI for stuff like grammar fixing, to the point where the AI is rewriting everything.
So, this person was not protected by your “hide it” suggestion, and indeed they may have been left worse off, because if they didn’t use AI, now it seems to people like they might’ve. I guess they could put out an announcement that they didn’t, which might help, but if the norm is the one you want (everyone lies) then this will probably just be taken to be a lie. Seems bad to me!
This doesn’t seem to engage with the bulk of my post, which is about the expected, not the actual, value of disclosing your AI use.
Yes, to engage with the bulk of your post I would’ve had to write a much longer reply. I feel like my posts in this thread are already much much longer than I usually like them to be. The deeper down into replies we get, the less I think it’s good to write novels, especially because at least on my screen, replies to replies to replies get skinnier and skinnier, and so long posts get stretched out quite a bit. This far into the replies I usually just try to point out things that I think are less obvious. Most of the other stuff I think about this topic is stuff I think people can figure out on their own.
That’s totally fine. Just as an observation which you can feel free to ignore, when it’s in the context of disagreement I think it comes off as uncharitably cherry-picking a small point. So it might be good to mention, alongside your minor clarification, that you are not going to engage with the point the interlocutor has been making
This is gonna turn into one helluva an arms race. AIs are just going to race against each other. Iterate { Hey Claude, write me a paper that doesn’t get flagged by Pangram. Update Pangram. Hey Pangram, learn to flag the papers written by Claude. Update Claude.}. Or human-AI centaurs will race against each other. But maybe it won’t matter: AIs will surely soon get so good that their outputs will be indiscernible from human outputs. I bet this arms race collapses into chaos real fast. And what happens when an author who didn’t use AI gets rejected for allegedly using it? A lawsuit?
Face it: the old ways are dead. “In this paper I will argue that . . .”. Nah, that’s over. We need to think about new ways to do philosophy. Or at least new ways to write philosophy. We’re not gonna beat the robots by doing something that the robots will soon do better than us. If we’re gonna insist on doing human philosophy, then it’s gonna need to be human. It’ll have to get weird and dirty in all the ways that only humans can get weird and dirty. There are lots of amazing futures for philosophy. None of them look like the past.
I don’t see any reason for thinking that AI will soon be better than humans at philosophy. I’ve used all of the models a lot, and I honestly, genuinely don’t think they are very good at philosophy. I have seen almost no progress either in the past year, and even some regression (I think Claude Opus 5 is substantially worse at philosophy than previous models, for example).
I think people are fooled by the fact that AI has gotten extremely good at certain tasks—namely those that can be trained using RL with an automated verifier. Well-defined math and coding problems fall in this category, but philosophy does not. As far as I know, the labs are still training the models to get better at philosophy through RL with human feedback. That’s very inefficient. In order for the models to get substantially better at philosophy, the labs will probably need to develop training methods that remove humans from the loop, but that’s a very non-trivial thing research problem.
Maybe you’re right. But if AI wasn’t already a threat, we wouldn’t be having these conversations (including various previous posts on Daily Nous on this issue). Yet here we are. Three years ago AI was barely competent at doing philosophy. Now it’s (mostly) written at least one published paper (and probably many others). Where will we be in three more years? Again, maybe you’re right that AI has plateaued. But given its astonishing progress in the last several years, I wouldn’t bet on it.
I don’t know what will happen in the future, of course, but it seems to me that the pretraining stage makes the models half-competent/half reliable bullshitters. That’s where frontier models were with respect to math two years ago, and it’s where the models still are philosophically in my judgment. More parameters and reinforcement learning via human feedback might make the models a bit better, but it won’t make them genuinely good. All the evidence so far strongly suggests that training an LLM to get really good at something requires vast amounts of reinforcement learning—on a scale that is feasible only if the training/verification process can be more or less fully automated (perhaps with a minimal amount of human input).
Consider math, for example. Three years ago, LLMs could produce plausible-looking bullshit (plausible to someone who didn’t know much about math). About one or two years ago, the labs realized that they could make the models good at solving math problems that had *numerical* answers by training them intensely via reinforcement learning (RL). The models (and RL) are extremely data hungry, so this was only possible because it’s easy to automatically check whether a numerical output from a model matches the correct answer. However, a year ago the models still couldn’t do proofs, because it is harder to create a verifier that can automatically check whether a natural language proof is correct, and having human mathematicians grade the LLM outputs would be hopelessly labor intensive. In the later half of 2025, the labs managed to crack this problem as well. We don’t know what OpenAI and Anthropic did, but DeepSeek were very open about how they did it: https://arxiv.org/html/2511.22570v1 Creating an (essentially) automated verifier for natural language proofs is what enabled DeepSeek to do the large-scale RL training that finally made their model very good at constructing proofs.
DeepSeek’s method is pretty ingenious, but it clearly exploits several features that are specific to mathematical proofs (for example, to show that a math proof is invalid, it suffices to identify a single step that is invalid). At the same time, RL training doesn’t seem to give the models skills that generalize very far. Again, one or two years ago, the models were already superhuman at solving math problems with numerical solutions, but couldn’t do proofs very well at all. Now they are arguably superhuman at (certain kinds of) proofs, but they are still not good at formulating their own definition or theorems, and my guess is that this is because it’s much harder to create an automatic verifier for whether a theorem statement is good, and so it’s much harder to RL this part of mathematics research.
Research in (analytic) philosophy is often of the form “Find plausible premises P1, P2, …, Pn that jointly entail an interesting conclusion C” (obviously that’s not all there is to philosophical research). My guess is that it is very hard to create an automatic verifier for this sort of task, and unless the labs manage to do so, the models will not get really good at philosophical research. Obviously, I could be wrong and I say all this with a dose of humility. But what I don’t think is legitimate is the inference “LLMs have improved a lot and gotten really impressive at many things in the past three years. Therefore, they will probably get really good at philosophy soon as well.” LLMs have gotten really good at *certain* things for fairly well-understood reasons.
Those are some very interesting points about potential limits to LLM philosophizing ability. I tend to think they’re all solvable.
The verification issue was a really sharp insight. And it might raise nasty problems for human philosophers – e.g. are there any measurable standards at all for what makes good philosophy? If the answer is “no”, then probably philosophy will really just turn into a kind of conceptual art. However, if there is any measure of philosophical quality, it’s plausible that AIs can be taught to measure it, and then to replicate it.
You’ve actually raised a point that we ought to be seriously interested in answering: can an AI be tuned to measure philosophical quality? That’s an open problem of a very deep kind. (AFAIK, nobody’s written about it. That would be an interesting article to write.)
Right now, mathematical AIs are shallow, they lack deep
creativity. They’re nowhere close to the creativity needed to envision something like the Langlands program. Recently there was some mathematical work in which the AI was trained to measure the difficulty of open questions, and then tasked with ranking them, and of coming up with new problems of high difficulty. So probably the creativity issue can be incrementally deepened to human then superhuman levels. The same should hold for philosophy.
And I’ll agree that current architectures are probably inadequate for great mathematical or philosophical work. I suspect there will be recursive forms of evolution that come into play. Probably something much more Darwinian than older recursive self-improvement schemes (which were mostly a kind of intelligent design approach).
But here’s the fallback clause: Suppose AIs do some cognitive tasks much better than humans, and vice versa. Now it’s the Age of the Centaurs. A centaur will beat a human or a robot anytime. I don’t like the idea of AIs alone doing philosophy; I do like the idea of philosophical centaurs (which was essentially the case with the recently published AI-human paper). I don’t think AIs are ready to partner with humans to do philosophy. The AI’s I’ve used are nowhere near good enough. The issue is when do they become good enough. (I haven’t used any of the super-expensive or custom AIs, so maybe in fact the centaurs are ready to go.)
I don’t think categorizing AI’s recent mathematical accomplishments as shallow is quite right. For instance, Terry Tao believes some of the insights involved in disproving the Jacobian conjecture were essentially unexplained miracles likely requiring deep insight; on another vector, the construction of the first-ever non-sophic group by Astra appears quite creative indeed.
It might be true that they have no handle on the Langlands problem, but that’s pointing to one of the largest mathematical programs on offer, one in which hardly any mathematician alive has a good handle on everything that is going on, and has long been viewed as a multidisciplinary research program.
Hey thanks for this! I’ve read about the Jacobian result, but I didn’t know Tao talked about it. Or the non-sophic group. I will check those out.
I’m very happy to change my opinions on these issues as new data comes in.
“And it might raise nasty problems for human philosophers – e.g. are there any measurable standards at all for what makes good philosophy? If the answer is “no”, then probably philosophy will really just turn into a kind of conceptual art. However, if there is any measure of philosophical quality, it’s plausible that AIs can be taught to measure it, and then to replicate it.”
I don’t quite think so. There is a huge gulf between a practice having (more or less) objective standards measurable/detectable by human experts, on the one hand, and its being possible to encode those standards in a precise numerical function that can be used in reinforcement learning. LLM training needs the latter, not just the former. And it’s not even sufficient that it be possible to write down a numerical reward function—it probably needs to be possible to train an LLM to learn that function without requiring a ridiculous amount of human feedback. Note that in the case of proofs, it’s easy to write down a reward function: r(proof) = 1 if the proof is valid and r(proof) = 0 if the proof is invalid. It was still a difficult research problem to get an LLM to learn this function.
To be clear, I don’t think it’s in principle impossible for LLMs to become great at philosophy (or at the aspects of mathematical research that are not so easy to encode in a reward function). The argument is basically as follows:
1-6 are true of solving well-specified math and coding problems, but all are significant challenges for many other domains (including philosophical research).
Machine learning does not require the prior construction of a numerically explicit utility function. That’s the whole point of ML.
Reinforcement learning absolutely requires specifying a utility function (or more precisely a reward function or reward model) to be optimized. It doesn’t have to be numerically explicit; its precise form may need to be learned (indeed, that was part of my point—to get the models to be good at constructing proofs, the labs needed to create a reward model that rewarded validity, but even though it was clear that that is what they wanted, actually learning such a reward function (without relying on too much human feedback) was still non-trivial). Without a reward function, there can be no reinforcement learning. And reinforcement learning is only successful insofar as the reward function is an adequate approximation/proxy of what we actually care about.
More generally, machine learning absolutely relies on utility functions (although they are more commonly called loss functions and typically do not represent any individual person’s personal preferences), such as the cross entropy loss (used in pretraining LLMs), F1 score (used to evaluate classifiers), Brier scores (used to evaluate probabilistic forecasts), and so on.
That’s incorrect. Current AIs derive their own utility functions from the data they’re given. There are several ways to do this. The UFs emerge during AI training. There’s a whole branch of AI engineering called “utility engineering” that studies UFs as they emerge and figures out how to constrain those emergent UFs. But the UFs self-organize within the AIs as the AI’s extract patterns in their training data.
Yes, after a model has been trained, you can ask its preferences and even elicit utilities and probabilities from it. That’s interesting, but it plays no role in the training process itself. The training very much relies on (human-defined) loss functions and reward models—how else would we be able to get the models to do stuff that’s useful to us?
I encourage you to read DeepSeek’s paper to see how they managed to get their model to be good at constructing mathematical proofs. It involved constructing a reward function and then finding a way of enabling their model to learn that function from minimal human input (this in turn required constructing not just a verifier, but a meta-verifier, and then having both in the overall reward function). The model certainly did not create the reward function itself “from the data”! (In fact, LLMs are still pretty bad at this sort of conceptual research, which is arguably the lifeblood of real AI progress.)
FWIW this is fascinating stuff and you should write a paper about this or if there’s something good already written that you can recommend please do.
One other comment: DeepSeek found that to get a verifier that was useful for the purposes of RL training a model to be good at constructing proofs, it was not sufficient that the verifier be able to tell that a proof is valid or invalid. The verifier also needed to be able to explain why a particular purported proof fails. Only then did they get a virtuous RL cycle where the generator could propose a proof, the verifier could point to a particular flaw, which the generator could then try to fix, etc. The great thing about math proofs is that it is checkable without human feedback whether a purported flaw in a proof is actually a flaw.
To get a similar virtuous RL cycle for philosophical research, the labs would presumably need to not only develop a verifier that could rank philosophical papers/ideas in terms of their quality, but that could also accurately explain why an idea is good/bad/promising/unpromising/novel/persuasive/etc. And training the verifier would (probably) need to happen with minimal human feedback, because human feedback is very inefficient. I’m not saying this is in principle impossible, and there are a lot of brilliant people working at OpenAI, Anthropic, DeepMind, etc., but it seems to me to be a significant hurdle.
Could we take off the heat a little? Why are scholars who point to potential academic malpractice called vigilantes? Sure, you should not jump to conclusions after submitting another person’s paper to an AI detector and share your assertion on social media before the accusation is properly assessed. But what is the alternative to policing ourselves as scholarly community? And was not this case resolved as it should? There was an investigation by the journal and it cleared the author. Isn’t that how the process should play out? What alternative do we have right now? Do want an official APA committee that replaces what you call “vigilantism”? Or should we keep putting our heads in the sand, ignoring the problem of academic practice by using generative AI?
To make a more substantive point: Note that the Statement from the Open Library of Humanities does not claim that generative AI was not used at all. It argues that generative AI was not used “to the extent that the complainant suggested, or with the intent to commit academic malpractice.” Moreover, an AI Declaration Statement was not necessary in this case but the standard for disclosure is set high by the journal’s AI policy. I believe that we should adopt a much broader disclosure requirement for all published worked in philosophy. Any use of generative AI that goes beyond checking for grammar and spelling should be disclosed. This is, imo, the only way how we can learn what kind of AI use is legitimate in certain contexts. It would also help rebuilding trust among our scholarly community.
My read of the situation is that two people in tenured or tenure-track positions (so made men in our world) made a public accusation against a post-doc knowing it could be nearly impossible for them to erase the stain that a mistaken academic misconduct accusation would do to someone looking for secure employment in academia. They did this using tools that they had limited understanding of. I doubt they thought hard about the harms or how they’d inform our thinking about what an adequate standard of proof would be. And I doubt they worried much about whether that standard of proof was met.
Why did they decide to go with a public and particularised accusation against someone in a non-permanent position? They could have privately contacted the journal and then this author wouldn’t have an asterisk next to their name despite being officially cleared. (And it’s significant having that asterisk next to your name in this job market where a rational committee who understands that they have limited ability to reliably identify the “best” candidate given the surplus of talent will just not take a chance on someone there are these sorts of doubts about.) They also didn’t need to particularise this. They could have instead tried to contribute to the kind of work that Marijke Van Vlasselaer and colleagues have done, trying to understand the epistemic value of the tools they use and studying collections of papers to look for signs of AI use. (And if they didn’t think they had the skill set to do that, isn’t that a reason not to play public prosecutor?) Two opportunities to minimise the risk of harm against an insecure person in our field ignored. For what reason? Because they’re bored? Because it’s a chance for an attention grabbing stunt?
https://link.springer.com/article/10.1007/s40979-026-00226-w
Here is a pragmatic suggestion. 1) Journals use validation software for checking hallucinated references and impose a hard rule: one hallucination, desk reject. 2) Journals use a AI disclosure policy, requiring only disclosure. Failure to disclose is automatic desk reject; false disclosure is automatic reject. 3) Referees can decide themselves if they want to referee a paper that has been disclosed as AI written in whole or in part. If no one wants to referee it for that reason: desk reject. In time there will be a learning effect of what is acceptable. And whatever gets published is immune to vigilantism.
The advantage of this procedure is that it leaves the door open to truly interesting AI contributions (if they exist), eg. the case in maths with the falsification of the Jacobian Conjecture. The decision to proceed with the paper is then a matter for the referees and the editor to weigh on merit and not dogma.
I recently had the opportunity to test it this way. I received a paper to referee. I noticed a lot of tell tale AI indications. After skimming the aims and results of the paper, I concluded there was nothing new in it. I informed the referee personally by email saying that I noticed undisclosed AI and that it did not appear to have anything new and informed them that since I did not know what the journal policy is, they had to take a decision before I proceed with any review. I disclosed that I was prepared to review an AI paper if it had something worthy to say (it did not). The Editor had a quick look and desk rejected. Total time on my side: 10 mins including email. For the editor, probably the same.
I may be really missing something: where in heaven is the problem?
I think there’s a question of what exactly needs to be disclosed, and how. What most referees primarily seem to want out of a disclosure is some information about whether they’ll be reading any AI-written text or an AI-structured paper. But if we ever get a publishable writeup of the Jacobian conjecture, I suspect it’s unlikely to have either of those features, even though the central idea of the paper is from AI! As far as I know, no one has figured out how to get an AI to write a mathematical proof that anyone wants to read, so humans are still doing that on the basis of these AI-generated proofs (or rather, everyone is holding off on publishing any of these results in journals until we figure out whether AI gets the skills to write it up in the next year or two, or some human finally gives in and writes it up).
I don’t know what sort of disclosure we should want if someone came up with the idea for their paper in the middle of some after-class conversations with Claude, but then came up with the proper statement of the idea and the argument the traditional way, and wrote the entire paper by hand. Maybe one of those footnotes (usually redacted in the version sent out for review) saying “thanks to [name redacted] for the conversation where this idea was first developed”?
If I take the case at hand, the paper was well written, but easy to see that it was AI. I’m open minded here: I just want minimal disclosure. So I asked the editor. They said “nah, desk reject for that reason”. Done and dusted. If they had said, “ok, I’ll ask” and on return I got a message “yeah, its AI: do you want to do it?” I have the choice according to my preferences and standards. If I’m an AI-phobe, I can say “pass”. If I am an AI-phile, I can say “great, I’ll do it” and then there are all the positions in between.
Here’s a case. Someone submits something new on Russell’s paradox. Some real formal wizardry. It’s all disclosed AI. But I recognize there is something to it. I can either say: publish as is. Or I can say: great stuff, now rewrite it yourself if you think its the bees knees and make a recommendation on that. Now suppose the author has a super Claude skill that makes it impossible to distinguish AI from human writing. Well, so long as the result is significant, does it matter. After all, what we are looking for “knowledge” or “human virtue”.
In time, the we will aggregate all these responses and reactions and through trial and error of refereeing and editorial decisions we will find new norms and have an idea of what works and what does not, and what the community wants. Some journals will gravitate towards AI-phobes and others to AI-philes. We will see which produce the better philosophy and philosophers. At the moment, I really can’t judge.
What we need now is a fast and efficient AI-submission filter to help editors. Once we have that, they can decide according to their judgements (that is why they are editors). Of course we will have the false positive/negative trade-off. But that’s life I’m afraid and with time we sort that out as well.
I forgot to answer the disclosure question. Here I’m liberal: minimal standards. I really don’t care if “if someone came up with the idea for their paper in the middle of some after-class conversations with Claude, but then came up with the proper statement of the idea and the argument the traditional way, and wrote the entire paper by hand.” Whether it’s Claude or a fellow philosopher who the author had coffee with to discuss the idea is irrelevant. Why? Because the the “someone” recognized a good idea and refined it in discussion with an AI or a friend. But again, if a journal is AI-phobic and demands that disclosure, then fine. Only problem is, it won’t work. If the community becomes more and more AI-affine, then the jnl will have less submissions or just submissions with dishonest disclosures which they can not do anything about so the demand for detailed disclosure becomes moot.
Here’s a flipside. I have another paper on my desk for review.It’s clear to me that it is not AI — at all. Its so idiosyncratic at times I think, “why on earth did the author not use an AI to clean this up? Do I really have to read it?”
So maybe a checkbox system that includes your case. Would I referee such a paper? Sure I would depending on if it shows promise.
But then again, I may be an outlier in the philosophy community. Among natural scientists and the quantitative social scientists it appears I am not.
I have worked extensively with Claude for 18 months now. I have a Pro 5x account. It’s my research assistant, secretary, technical support for my entire teaching workflow including generating and grading class quizes. Given the cut-backs in my university, I would be dysfunctional without it. I instruct it to do tasks I would actually give humans. The best thing is that its available 24/7, infinitely cooperative, and technically smarter than any human I know. And while it performs tasks in the background for me, I can do philosophy.
It does all my latex for me, I have designed and developed an automated manuscript production pipeline for diamond open access journals with it that includes mechanical proofing procedures. I have used it to do technical proofing checks. You name it. But, one thing it does not provide me: the mere joy of doing philosophy and writing it down and refining it for myself.
Maybe because I am tech-affine and open-minded I simply do not worry about these things and believe that with sincere attention to detail and a passion for what we are doing we will fumble our way out of the fog and be wiser for doing so.
There has been some discussion of whether Justin’s choice of “vigilantism” is justified. We do not need to be purist about it. What he clearly meant, is “social media-based self-righteous pile-ons” and these can be highly damaging. In that sense the terminology is well-chosen.
Again, all sense of proportion got lost in the PP case. When I saw it posted on X, I asked myself immediately, “why would anyone do that?” Surely, the appropriate means is a confidential letter to the editors. And that is what I did. I know one or two on the PP editorial board and immediately sent them a tip off with a screen shot of what is in X.
Of course academia is a collective enterprise, and post publication discussion and further review is normal. But that can be done in all kinds of ways. For instance, contacting the author themselves, or finding someone who knows them. We are after all a small community. We generally know someone on or connected to our journals.
So my reading of Justin’s use of “vigilantism” was a polite way to say “Social Media Pile Ons? No Thanks”. It reminded me of a famous car sticker of the 1970s in Germany “Atomkraft? Nein Danke” (the slogan originated in Denmark).
Justin: I hope I did not misinterpret you!
Thanks, Matthew. To be honest, I take “vigilante” to have a more neutral meaning, that is, as an unauthorized person who uses means typically reserved for officials to try to enforce the law, or what they think is right. To my ears, without further context, it has a slight yet defeasible negative connotation (consider that most superheroes are vigilantes, e.g., Batman). Vigilantism is usually a bad idea but under certain unusual circumstances, it seems, it could be justified. That said, the particular vigilantes I was referring to were people who were indeed posting stuff on social media. Whether they were attempting to generate a pile-on or feeling self-righteous about it, I don’t know, and I’m not sure it’s helpful to speculate about that.
Maybe you should have posted a Batman instead of a superman. In any case, I was not speculating about the intentions but rather the consequences. Its just so easy to trigger the pile-on, even unintentionally. I can see that someone might think that the global town hall might be the right place in some idealist world of a global demos with deliberative norms. But we are not there yet … I am glad you raised the topic because its a baptism of fire as someone said to me about the SEP entry X-storm.
https://youtu.be/nhbClGpMdaE?si=mZbq-frJggGDE2ZO
“We are our own superheroes. Fighting for our rights. We are our own superheroes. All power to the people…Superman was a po-po and Batman was a snitch. Wonder Woman fought to keep the rich people rich”
Mic Crenshaw & Dead Prez — Superheroes
“And if someone spends a year researching, writing, workshopping, and revising a paper, and then asks an AI to improve the writing style throughout the paper, the resultant paper will get a “100% AI generated” evaluation, too. Despite their same scores, though, these seem like cases that it might be plausible to treat quite differently. ”
I’m not sure I agree with this. The assumption here is that (a) you can draw a simple and unambiguous distinction between “content” and “style,” and (b) that for some reason the latter has a less serious bearing on how authorship should be attributed to a publication.
But, insofar as one employs natural language, you will find the distinction between content and style is neither simple nor unambiguous. And, moreover, it does not seem to me clear that we ought to treat style as irrelevant to authorship attributions. If, for the sake of argument, I were to excogitate a set of ideas, but then have another human do everything required to work them out into a publishable paper, then it seems to me that I do not in fact deserve all the credit for the resulting piece. Or, at the very least, my use of a writer should be appropriately disclosed in the publication itself.
Beyond this, I would be concerned for anyone who calls themselves an academic philosopher and yet cannot express their ideas effectively without the aid of AI. Articulation of ideas is just as important a skill as the production of them, and ought to be a central part of disciplinary curricula. Philosophers should know how to write.
I’m not sure why “I would be concerned for anyone who calls themselves an academic philosopher and yet cannot express their ideas effectively without the aid of AI. Articulation of ideas is just as important a skill as the production of them, and ought to be a central part of disciplinary curricula. Philosophers should know how to write.”
This seems remarkably close to expressing concern about any second-language speaker who calls themselves a academic philosopher but doesn’t have fluent writing skills in any language that professional journals publish in, or expressing concern about anyone who can’t physically type but has dictation software do it for them.
Obviously, there are important differences between language skill, physical typing ability, and skill at phrasing things, but I don’t actually understand why one would think that two of these are not essential for being a professional philosopher while the third is.
I’m having trouble not falling headlong into incredulity when I hear you claim not to understand why the ability to express one’s ideas effectively is an important part of being a philosopher.
“This seems remarkably close to expressing concern about any second-language speaker who calls themselves a academic philosopher but doesn’t have fluent writing skills in any language that professional journals publish in, or expressing concern about anyone who can’t physically type but has dictation software do it for them.”
I didn’t actually say or imply any of that. One can recognize and sympathize with such difficulties (each of which is importantly different from the other in various respects), without conceding that skill of expression is somehow irrelevant to the practice of philosophy. “X is not equally easy for everybody” does not entail prima facie that X is not a duty.
But lacking in skill in expressing oneself in *English* (the language most international journals accept nowadays) is not the same as lacking in skill in expressing oneself. One may have great facility in expressing oneself in one’s mother tongue without having great facility in expressing oneself in English.
Certainly. And such a person would have satisfied my little condition for “philosopherhood,” although they would also face unfair, unequal, and lamentable hurdles compared to native English sspeakers. Do you see how one can be consistent in holding both positions?
“This seems remarkably close to expressing concern about any second-language speaker who calls themselves a academic philosopher but doesn’t have fluent writing skills in any language that professional journals publish in, or expressing concern about anyone who can’t physically type but has dictation software do it for them.”
Thank you, Kenny, for posting this. A couple of points:
(1) In many cases, non-native speakers in academia, especially those who do not work in an English-speaking environment, can indeed produce perfectly legible English, but at the expense of devoting an amount of time to it that native speakers would never dream of. In those cases, AI polishing causes a marginal increase in productivity for native speakers, and a dramatic increase in productivity for non-natives.
(2) Several brilliant, non-native-speaking colleagues of mine have recently told me anecdotes along the same lines: during the last couple of years, they have managed to publish various papers in top-tier journals for the first time, thanks to the help of AI. Was their use of AI improper? Why, exactly? The published outcome was their thinking, their ideas, their arguments. Was it *their prose*? Probably not quite. Would it be flagged by Pangram as AI-generated? No idea; probably yes.
Having said that, I do understand the point of the original paragraph that motivated Kenny’s reply. For example, all the teachers I know, in all languages and all parts of the world, are concerned about students becoming unable to write fluently in their own language because of chatbot dependency. We want our students to be able to produce coherent text unaided, in some language, and if someone never acquires that skill, one is somehow reluctant to say that they have mastered philosophical thought. The world of academic journal publishing, however, is somewhat a different animal, precisely because of its international nature.
Off the back of my earlier experiments, *possibly* the solution here for ESL philosophers isn’t to write the article yourself into English and then ask Claude to improve the English; it’s to write it in your first language, ask Claude for a fairly literal translation into English, and then edit the English. At any rate it might be worth someone trying it, both to see what the quality is like and to see if it trips Pangram.
I just had Claude translate a 1750-word piece from my mother tongue (Dutch) into English. I changed a few words that I found odd choices, but overall the quality of the translation was impressive. Then threw it into Pangram. Result: 100% human-written.
Can anyone share a Pangram false positive involving a full paper (say over 4000 words) that was written before 2022?
You might also want to send it to Max Spero (Pangram’s creator), who has been offering paid bounties for verified (pre-2022) longform false positives.
(Note also that the latest version distinguishes “AI generated” from “AI assisted”, contrary to what Justin claims in his “first problem”.)
I just did a test run of Pangram. I wrote a short paper of roughly 1000 words. When I uploaded the paper to Pangram, it returned a score of 100% human. When I instead copy and pasted the paper’s entire text into Pangram, it gave the very same paper a score of 81% AI and also read that “This appears to be AI text that has been paraphrased or rewritten”. Not sure what lesson to take from that test run. But it does lower my confidence in the program.
Things just got weirder. I wrote another text and decided to add hyperlinks into the text, to citations for certain claims in the paper. I did this using the hyperlink option in Microsoft Word. This was the same program I did the first test run with. When I copy and pasted (not uploaded) that version of the text to Pangram, it rated the text as AI by 14%. When I uploaded it, 100% human.
Next, I pasted that same text to Gmail. When I then copy and pasted the version without hyperlinks from the Gmail window, I got a reading of 100% human in Pangram. When I pasted a copy with the hyperlinks from the Gmail window, I also got a reading of 100% human.
When I then saved the Word doc (with hyperlinks) as a PDF and uploaded it to Pangram, the score was 100% human. When uploaded as a PDF without the hyperlinks, also a 100% human rating.
Make of this what you will. Not sure what to think of these differences (and some similarities) in results.
Hi! Founder of Pangram here. This sounds like really weird behavior — Pangram is a deterministic model, so the same text should produce the same output every time.
Feel free to drop any share links (share button is in the upper right corner) and I’m happy to take a look
This is a separate issue, but I have found that Pangram often flags my human-written text as AI if it is surrounded by AI text. For example, here is a human-written paragraph (some drivel about philosophy that I wrote) that Pangram correctly identifies as human-written with high confidence in isolation:
https://www.pangram.com/history/8b309b33-0a64-4e1b-a589-22559b5d4001?ucc=MQRT783oIvP
but which it claims is 100% AI-written with high confidence when it is sandwiched inside some AI-generated paragraphs:
https://www.pangram.com/history/63a0b85d-688f-48ec-9f84-2648459b2047?ucc=MQRT783oIvP
Pangram claims with high confidence that the entire 475-word essay in the second link is 100% AI; the correct answer is about 50%, as you can see from the word counts.
I realize this is a short text (though not that short— not enough for Pangram to give its “short text” qualifier), but even in longer texts I find that human-written text often gets “swallowed up” by surrounding AI text. This has led me to distrust Pangram’s percentages.
A general observation on this discussion, having reflected on it for a while:
Pangram exists; anyone can use it. It is widely believed (I think correctly believed, but that’s not actually essential to my argument here) that it does not generate a lot of false positives on long-form text. And while I don’t know for sure, my expectation is that in the near term AI detection is going to get better rather than worse (treat ‘Pangram’ in what follows as a shorthand for whatever is the leading AI detector in the future.)
So: if you publish something, sooner or later someone is going to put it into Pangram. Maybe the journal you send it to; maybe the person replying to you; maybe someone on a search committee for a job you’re applying to. And if your paper comes up as largely AI-written or AI-assisted, they are going to assume that AI was very substantially used in the writing of the paper. (Again: I think their assumption will be reasonable, but that’s not essential to my argument.) The issue that started this discussion – that people should be careful with accusations – isn’t ultimately that crucial given that everyone can check for themselves.
These facts rather cut through the ethical debate about AI disclosure:
What is the difference between having an AI polish a text to improve its style and clarity and having the same text edited by a heavy-handed copy editor? Non-native speakers work with copy editors not only, and not so much, to correct grammatical and typos, but also to refine the wording. And as many know, when working with a copy editor—perhaps one with a background in philosophy—over a long period of time, the feedback evolves from “this term is better than that one” to “this sentence needs to be rephrased” to “maybe you could write this paragraph this way.” No one has ever asked to credit a copy editor or to disclose that the text was edited by someone else.
“What is the difference between having an AI polish a text to improve its style and clarity and having the same text edited by a heavy-handed copy editor? ”
That the former, but not the latter, appears likely to detect as “AI-assisted” on Pangram, and that this may have reputational consequences. Whether it *should* have reputational consequences isn’t the point of my post.
(I also think that if your human copyeditor is *sufficiently* heavy-handed, they ought to get co-author status or at least be acknowledged. I appreciate there are gray areas.)
In my experience, most people who have friends help with copyediting might write in the acknowledgements something like “thank you to X for their helpful revision suggestions on an earlier draft” or something like that, but I haven’t seen much more than that. I’ve provided such copyediting advice to friends. I’ve never heard of someone being credited as a co-author for their help with copyediting, even somewhat heavy handed help, but maybe I’m just not as familiar with the practice and how others handle it. As another poster has mentioned in this discussion or an earlier one, journals like Springer and others also provide links to (and therefore encourage?) copyediting services for their authors. I highly doubt such professional copyeditors get credited as coauthors. That would seem rather absurd.
Copyeditors at major magazines or newspapers will also often provide the kind of copyediting the poster above has mentioned (“maybe you could rewrite this sentence or paragraph this way”) and will often even force those copyedits on the author. I’ve never seen such an editor get credited as a coauthor or even ask to be credited as such. Maybe there are examples of cases where that has happened, but they are certainly not common or standard practice.
People really seem to be reaching with their desire that AI copyediting be credited so profusely or treated so damningly. My feeling is that some of these posters are primarily motivated by wanting these authors to have an asterisk next to their name so they can be harmed reputationally or career-wise, perhaps for the poster’s own benefit. Something about it sits wrong with me.
The parenthetical point about “heavy-handedness” is fair but also seems to be being given with the implicit suggestion that most of the AI copyediting being done is at a sufficiently heavy-handed level to warrant being credited as co-authoring. That seems like a rather ungenerous assumption. Plenty who get flagged are probably using it at no more than a normal (not sufficiently heavy-handed) copyediting level. It would take a lot of guts to try to go all in on an AI paper at a major academic journal. I assume most who use it are just trying to improve their writing while being conservative, respectful, and staying within acceptable limits. There will always be bad actors, of course. However, there seems to be a lot of assuming the worst going on with this issue.
A bit of a rant on this:
First, this seems like a policy that would in practice have the effect of punishing non-wealthy professors, lecturers, postdocs, grad students, and undergraduates — that is, unless the same Pangram user package is made freely available to all.
Pangram has a limit on how often you can submit a paper for free and how many words can be checked at a time, correct? Presumably the paid package allows a user to submit more words and more often. This means those with money will be able to buy the more expensive Pangram package that allows users to submit longer written works and check more of their work more often. Those who can’t afford the paid version will be stuck with only being able to submit so much of their work at a time, having to wait until they can submit and check more of their work again. For graduate students especially this means the wealthy students could submit multiple course papers to Pangram at once compared to the non-wealthy students who could only submit one paper (and maybe not even a full paper) per day. Given many important papers are often due at around the same time (the middle or end of the term), that puts wealthy students at an advantage. Also, consider that if someone puts one of their papers into Pangram and it gets flagged, they’ll need to revise the paper and put it into Pangram again (and perhaps again and again) to make sure the revised version doesn’t get flagged too. If Pangram gives different results to the same paper when it is uploaded vs. copy-pasted, then, once again, that means even more required uses of Pangram.
Your suggestions would also plausibly have another potentially harmful effect. Anyone who didn’t use Pangram might be harmed career-wise or reputation-wise if their work just happened to get flagged as AI, even if wholly human written. Those (or some of those) who do use and can afford Pangram might now incorrectly assume everyone else must be using it too. So if someone gets flagged, they might think that person must have used AI and just not been smart (clever?) enough to check and revise their work with Pangram before submission.
Those who can afford AI *and* Pangram and use both effectively enough to have their writing not get flagged as AI will be at the greatest advantage. Such a situation would be a bit of a farce as far as preventing AI writing and publication goes. The recent note by Angel Pinillos (ASU) on this has changed my mind on this issue.
It also seems likely that the reliance on Pangram will have on odd effect on writing styles. It might make even non-AI users alter their writing just so they can make sure it won’t get flagged as AI by Pangram. It’s not clear the writing resulting writing style wouldn’t be worse if they had to do this.
I’m not sure what to think of all this, but this does make me think more and more that perhaps we should just give up on this idea of trying to push the use of Pangram and the like to try and police the discipline. It feels like it is just becoming a witch hunt where the worst is being assumed of everyone who gets flagged (guilty until proven innocent) — and to what end? We’ve also seen now that Pangram appears to get things wrong sometimes. This is compounded if it might not flag a paper as AI when it is uploaded but then flag the same paper as partly AI if it is copy-pasted into the program. If it’s also true that the line between AI copyediting and AI generation isn’t always clearly distinguished by Pangram and that both sometimes get flagged, this is even more troubling, especially since copyediting doesn’t seem to be problematic.
Suppose it’s true what a friend told me that if you use Pangram to look at the recent publications of some of the big names in philosophy these days, you’ll find that some of them have articles that get scored as AI at 20% or higher. Some of the names they mentioned are tenured (and some non tenured) professors at prestigious universities whose names you would probably recognize and who you likely respect.
If that’s true, it makes me wonder if what we really shouldn’t care about primarily is just whether the arguments are good and not whether the writing is wholly human or even mostly human. Otherwise, how can we decide not to hire or tenure someone if some of our own “best” colleagues are getting their own work marked by Pangram at such high percentages? It would be hypocritical to say the least. Or do we for sake of consistency first demand that those tenured and non-tenured professors who get flagged be fired? What percentage do we settle on as fireable/unhirable vs okay?
I think “purity” of human writing is a fair and good thing to value, to promote, and to practice. However, I’m not sure it’s necessary. A department might have the right to choose to be a “purist” department and only hire or tenure those who don’t use AI. I do wonder, however, how they could enforce that. They’d run the risk of turning down candidates who write their papers without AI but who still get flagged. How would such a department handle cases where it turns out new future Pangram-like technology flags one of their non-tenured or tenured members? It’s not obvious to me there is a good clear path for such a purity demand.
None of this is to say I love AI or think we should all go all in. However, I think some of our colleagues have gone a little overboard in their witch hunting and condemnations. It saddens me to see so many in the profession being so callous and downright aggressive on this matter. I feel for the person who apparently got accused at PP.
I’m not advancing a policy, just pointing out that in practice:
(a) if you put your work in a public-facing space, interested people are sooner or later going to put it into Pangram;
(b) Because Pangram is widely believed to have an extremely low false-positive rate for long-form academic writing, if your work is flagged by Pangram as AI-written or AI-assisted, people will draw the conclusion that you used AI to help you write the paper.
(c) Because many people have negative attitudes to AI writing, and even more have negative attitudes to undeclared AI assistance, this is likely to have consequences for how your paper is received.
These are just empirical facts, which I think people would be wise to be aware of.
Now, it is also true that the belief that Pangram has a low false positive rate in this context appears correct: despite fairly extensively looking, I’ve failed to find even one example of a long (say, 2,000 words or more) piece of academic writing from before 2022 that gets flagged as AI-written by Pangram. (If you know one, post it.) And so I think the inference from “this paper is flagged by Pangram as mostly AI-written” to “this paper was AI-written or AI-assisted” is highly reliable, and the concerns that you and others raise about false positives aren’t really an issue. (They seem to be based on either much shorter snippets of text, or on older and less effective AI detectors.)
I just spoke with two researchers who evaluated Pangram and GPTZero specifically and their work was very insightful to this point. You can read the paper here: Why AI Detection Fails for Academic Integrity (https://arxiv.org/abs/2608.11256).
Worth noting, though, that:
(i) this research dealt with quite short pieces of text (abstracts for papers – on average 350 words, I think) whereas AI detection gets a lot more reliable on longer texts. See Who wrote this? Evaluating the reliability of AI detection tools in higher education | International Journal for Educational Integrity | Springer Nature Link for how various tools did on 4,000-word-plus pieces.
(ii) the research doesn’t produce evidence of false positives: Pangram (and GPTZero) correctly assessed all the 2013-2015 abstracts as human written, and assessed 9%-15% of the 2023-2025 abstracts as AI-written or AI-assisted, which is correct for all I know.
(iii) this wasn’t the latest version of Pangram (not a criticism of the paper in itself, this is a fast-moving field.)
Fair enough, we’re getting somewhere now.
A few things to consider:
1. Given the cost, are journals going to run Pangram on the entire paper, or just the abstracts?
2. There is a note about how formulas, and other cross-domain markers lower accuracy, so how will that impact say, Philosophy of Science or Mathematics papers versus other types of papers?
3. I think it’s worth expanding and replicating this study on longer texts and philosophy papers specifically. While false positives may look low statistically, the person who is accused of the false positive faces a career penalty. In the least harmful sense, their paper is filtered out, in a more harmful sense, their reputation is called into question.
At least where I work now (and admittedly, we are not using Pangram), we never use AI scores as evidence of AI-use or integrity violations on their own. They are only useful for corroborating more obvious violations.
Narrowly: the cost is negligible, a dollar or less per article.
Brian Fry, a professor of law at Tulane and specialist in intellectual property has a radical stance on plagiarism and the use of AI. Basically, it’s “go easy”. Warning: before anyone gets up in arms about it, take a deep breath. Brian is also a filmmaker and therefore is looking at this from the perspectives of the arts, which historically live from copying and imitation. His approach is roughly “borrowing without asking” for both plagiarism and AI. Remember, for any Banksy fans, he famously said in his book Wall and Piece that “Copyright is for losers,”
If you have an X account, its worth a peek and some minutes of reflection:
https://x.com/brianlfrye