“The fact that you, an expert human, generated the text is evidence that the view is worth thinking about—more so than if the text were generated by an LLM.”
That’s Eric Schwitzgebel (UCR), writing at * The Splintered Mind *about why he doesn’t want to receive AI-written emails about philosophy, and why philosophy journals should reject AI-written submissions.
Here’s more context:
I’m talking about evidence about evidence: meta-epistemology. Your email or your article presents evidence for a particular philosophical view (alternatively, evidence that you support a particular philosophical view). In a simple world, I could evaluate this evidence entirely on its face: How good is the proposed view? But in the actual, complex world, it helps to have evidence about the quality of the evidence. The fact that you, an expert human, generated the text is evidence that the view is worth thinking about—more so than if the text were generated by an LLM. This holds even if the text is exactly the same, which of course it wouldn’t be.
*An increasingly large part of the function of journals is to provide evidence about evidence—the value of their imprimatur. The fact that an article appears in *Nous *or *Ethics is evidence that it has been through rigorous review and was judged worthy by several expert humans applying unusually demanding standards of quality and importance. Its appearance in those journals is thus evidence (imperfect of course!) that the reasoning is of high quality and the arguments worth taking seriously.
Similarly, if I know that an email or an article was written by a respected colleague, reflecting their positive creative exertion in trying to choose the right words, guided by their intuitive expertise in how to phrase things, I have better reason to take it seriously than if I know that it was generated by an LLM and reflects only their passive after-the-fact assent.
Philosophers sometimes suggest that we shouldn’t care if an argument was human-generated or AI-generated—that insisting on human-generated prose is fetishizing personal human interaction rather than facts and argument quality. In a way, that’s true: A sound argument is a sound argument. Similarly, we shouldn’t care if an article was written by David Chalmers and published in Philosophical Review or whether it was written by someone with no institutional affiliation and published on an obscure blog. If the argument is good, it’s good—of course, of course!
But at the same time, we have limited attention, limited time, limited ability to understand the nuances when matters drift even a little from our tightest foci of expertise, and in these cases it’s helpful to have meta-evidence. What should I read? How far should I trust the author has the details right, versus how much should I critically and chase down independent sources? How much should I let their way of phrasing things, their habitual patterns of thinking, the presuppositions hidden in their word choices and sentence structures, slip gently into my brain, silently strengthening my own associations and predilections?
Why does “knowing this text was written by a human philosopher” provide some evidence that the text is worth a certain degree of attention?
*Human experts think *differently *and *better *than LLMs. Their word choices, even subtle ones, reflect sensitivities that they might not themselves be aware of. Typically, an expert’s prose will be more sensitive to the matters on which they are expert than the output of a language model. *
And checking an LLM-written text to make sure it reflects your own thoughts is risky:
*There’s a huge cognitive difference between nodding along while reading something and *actually productively generating a text. Two reasons: First, once the text is on the page, it’s easy to passively let the approximate word suffice, rather than thinking about word choice in the same effortful, active way we do when generating prose de novo. Second… I doubt that human beings, even experts, have a good sense of all the factors that shape word choice—everything they’re being sensitive to. You would have phrased it slightly differently, and even if you don’t know that, or why, a different signal is sent and received.
See his post for a good example of what he’s getting at, along with an explanation for why his argument “is not a call for blanket rejection of the use of LLMs in philosophical (or other) writing.”
Related:
– *Ethics *Announces AI Policy
– The Ethics of Using AI in Philosophical Research Eric Schwitzgebel makes valuable points about AI-written papers ‘nodded-through’ by their human instigator; the value of one’s de novo, even clumsy, own wording; and (in the underlying post) about careful use of AI. However, there is a slightly concerning note of anthropocentric self-satisfaction. The assertion, ‘Human experts think differently and better than LLMs’, appears open to challenge on both counts. It seems that human brains may predict continuations subconsciously in a similar manner to LLMs (https://www.sciencedaily.com/releases/2026/06/260624025514.htm), while (from personal observation) higher-version LLMs, at least, seem rather good at grasping what we wanted to say but didn’t quite manage to. Humans, at least for the time being, have advantages, e.g. the passion to press deeper into an idea, the ability to say what matters. But a little more circumspection might be in order.
Some of these things are likely really useful!
I haven’t yet read through the text of any AI-generated mathematical proofs to figure out if the AI-generated text is itself fine as a section of the mathematical publication, or if you want to rewrite it and refine it. I’ve heard claims that they’re not as good as humans at generating re-usable structures and ideas in the process of coming up with the proof, so that they might burn through our stock of conjectures without actually growing the mathematical idea pool – but again, I don’t know if that’s exactly accurate (or if it might change in a few months).
Vibe coding has been great for little side projects I’ve been interested in doing outside of academia, and also for putting together little activities for students to do in Zoom breakout rooms to illustrate some ideas in online classes. Now that I understand that I can generate code to do things systematically, I’m trying to think if there are things that would be nice to have for bits of my work (beyond TikZ code for diagrams in LaTeX).
100% with Eric on the epistemic value of expert word choice. My worry is the move from “human word choice is valuable as a guidance signal” to “journals should require it.” Without a public verification method, making human authorship a formal requirement for acceptance collapses into a mix of honor-system pledges and subjective sniff tests by editors and reviewers. To get published, either your word already needs to be trusted, or you need to know the private handshake a given group will read as human. Both take peer review from something that can at least pretend to be anonymous to something that can’t. Plagiarism norms work because the source is traceable, which is not the case here. I wrote about this issue and the threat it poses to public reasoning here.
There’s room for debating where the line should be in edge cases. Maybe some benefit of the doubt can be extended when it’s not totally clear. But when it is, rejecting the paper for that reason makes sense. (And, presumably, publicly stating you don’t accept AI submissions discourages some people from submitting in the first place.)
When I’m trying to decide whether to read a random unpublished paper that I come across on the Internet, the fact that it was written by a competent human philosopher does seem to be useful meta-evidence that it’s worth my attention. However, if it’s been published in Ethics, then that seems to be such strong evidence that it’s worth reading that whether a human wrote it is no longer so important. So the implications of this meta-evidence point for journal practice are at best unclear.
That’s a good point, but I don’t think the fact that the paper is published comes close to screening off author identity. In my experience published papers vary dramatically in quality, even at the top journals (at least some of them). In the end, usually the person who by far has thought most deeply about the topic is the author themselves, so their judgment has significant evidential value over and above that of reviewers and editors. For example, I’d be much more interested to read a paper by a philosopher whom I respect than a random paper on the same topic at the same journal.
Makes sense. And here’s the additional rationale I offered in the last AI post on Daily Nous. I wouldn’t call it “meta-evidence” but only a practical way to **triage **your reading load, esp. in discussion forums:
I’m not necessarily saying that your AI-assisted post was bad or wrong, just that it’s “tainted” with AI in the view of some or many readers. Here’s why that matters:
For many of us, we already have way too much to read and can’t possibly engage with everything worthwhile. So, one reasonable way to triage our workload is to avoid reading or engaging with anything that used AI. One reason for doing this is to avoid the
[Gish gallop], where AI-assisted interlocutors can “flood the zone.” With AI, they can engage in[to overwhelm the conversation with rapid-fire, low-effort, instant arguments that no human can respond to fast enough. If you’ve never been on the receiving end of that, you might not know.]argumentum verbosumThere can be other reasons, such as feeling disrespected by those who write with AI. That becomes clearer in more personal communications, such as
[emails]or[eulogies].But more generally, separating the message from the messenger has been a problem for literally thousands of years, so I don’t think anyone should be surprised much. And it’s not always wrong; sometimes, who the messenger is matters in a non-
ad hominemway…
I’ve tinkered with Claude by feeding it one of my half-written papers, just to see if what it produces will come close to what I had imagined. (Don’t worry, these aren’t for publication.) I occasionally find it difficult to understand a bit of reasoning in some of the generated text. The weird part is that, in these cases, I have no confidence that what Claude wrote actually does make sense, and that it’s therefore worth trying to understand. It made me realize that, before AI, I tended to be highly deferential to authors; I always just assumed (with exceptions for obscurantists and the like) that anything hard to follow would still be coherent, and thus that struggling through it would be rewarded. I find this experience odd and interesting, and maybe sort of related loosely to Schwitzgebel’s discussion. Have others experienced it, or thought about it, or whatever?
Not so long ago, LLMs still hallucinated a lot. And they also have a much higher rate of confidently asserting things they are not sure about than experts at comparable levels. The latter remains true. So I am kinda not surprised by your experience. We were a lot more skeptical not so long ago, no?
That’s fair, I should have been a bit more careful. I realize that hallucination is probably why I don’t trust a difficult AI-written passage. What I found interesting was (a) the strangeness of the experience (of reading a chain of reasoning and not knowing if it would actually reward the effort of trying to understand it) and (b) the realization of how much I relied on basic trust when reading human work. Maybe it reminded me of the claim that future superintelligence will—if it ever exists—so vastly outpace us that we won’t even understand how it verifies its factual claims. The claims themselves might be tractable, but we’d have to take the method which got to them on faith(?)
Maybe that’s not interesting, or I don’t know what I find interesting about it; for some reason I found it worth mentioning.
Yes, I’ve had this experience too. In the past, when I read a piece of difficult text, I tried to find a way to make sense of it, because I knew the author must’ve meant something by it. But nowadays, there’s some chance I later discover the text was just some garbled nonsense spat out by an LLM and feel annoyed that my time is wasted.
The thing is, we don’t know if an article is written with substantial help from AI. As the latest models become genuinely more capable in certain subfields, I am starting to suspect a lot of articles submitted are like that. As a referee, I do not know what to think of it (obviously I am not talking about AI slops, in which case there is no question). It feels like I should be using substantial help from AI so as to write reports more efficiently. AI is coming for us, I think, whether or not we are using it or not.
I fear that OP and readers are not fearing enough.
The paradoxical question is that when Large Language Models write well enough and are examined meticulously, it’s impossible for humans to know how the author accomplished it. In fact, this involves a question: Is the hostility towards AIGC out of the fear of losing the anthropocentric status or due to concerns about responsibility? I’ve noticed more of this hostility in the humanities. Currently, Internet companies have widely adopted AI technologies, including vibe coding, to generate content. As far as I know, in fields such as physics, biology, chemistry, computer science, and artificial intelligence, the concept of AI4XXX has emerged and been quickly recognized. However, in fields like literature, philosophy, and art, there are still some people who strongly resist. In fact, I don’t think humans have any unique advantages in these fields. In other words, I don’t think human agency endows purely human – made products with any mysterious properties.
You don’t have to speculate why people are critical of AI, and attribute to them a “fear of losing the anthropocentric status” or whatever, you can just read their actual criticisms (e.g., unreliability, encouraging bad epistemic habits in people, etc.).
There is an annoying tendency of (certain) AI boosters to just ignore the stated criticisms of critics of AI and conspiratorially speculate on hidden motives. No, I dislike it for the reasons I said.
Both Alice and John also just assume that AI writing is undetectable, which is a bit funny after coming from that other Daily Nous article, where the comment section immediately noticed the author used AI without it being disclosed. So I reject that premise.
In fact, I have of course read relevant criticisms, but regrettably, the vast majority of the arguments in these criticisms focus on the differences between human writing style and that of Large Language Models. And this difference can be addressed through deep engagement. This has led to a situation where people constantly speculate about whether others’ papers are completed by AI, causing some previously normal language usage to be regarded as incorrect behavior. This is an overextension, and it is limited to the humanities and social sciences. In disciplines such as physics, biology, computer science, and artificial intelligence, linguistic expression only needs to be accurate, with more attention paid to the accuracy of experimental data and the standardization of procedures.
We can draw a conclusion that the main factor in determining whether an article is worth reading is the accuracy of its viewpoints and arguments, rather than the identity of the author. If an AI proves a certain mathematical conjecture (and such a case already exists), should we refuse to read it and directly dismiss it simply because it was generated by AI? This is in fact another form of identity discrimination. Sometimes we can observe a phenomenon where, in single-blind journals, articles written by renowned scholars are more likely to pass review. Correspondingly, some authors, concerned that scholars holding your viewpoint might discriminate against their articles, choose not to disclose AI assistance in their papers (He, Y., & Bu, Y. (2025). Academic journals’ AI policies fail to curb the surge in AI-assisted academic writing. Proceedings of the National Academy of Sciences of the United States of America, 123 9, e2526734123). For example, in the article I have cited, after major publishers announced their AI policies, among 75,000 papers in 2023, only 76 chose to disclose AI assistance. In fact, AI assistance certainly far exceeds 76, but the other AI-assisted articles all underwent peer review and were published. This shows that if a paper involves AI participation and has been revised, even professionals would find it difficult to detect. Moreover, it is precisely because of discrimination against such assistance that relevant policies are difficult to implement.
I believe that the importance of an article depends on its viewpoints and arguments, not on the author. As long as the quality of the paper is good enough, whether the author is a renowned scholar, an ordinary person, someone without a diploma, or with AI assistance, or even entirely completed by AI, it deserves to be taken seriously.
To “discriminate” just means to tell the difference between two classes of things and treat them differently. We in fact often think of discrimination as a good thing when it’s not about people – we talk about a “discriminating taste” or a “discriminating palate”.
Even when it’s about people, it’s sometimes good – math professors want to write hard tests so they can discriminate the students with strong skills from the students without them, rather than indiscriminately giving everyone A’s, or discriminating only on the basis of minor slip-ups.
Of course one can discriminate against AI – and probably sometimes one should, but sometimes one should be discriminating on some basis that is more relevant for what is actually important.
Right! Try this test:
https://lancasteruni.eu.qualtrics.com/jfe/form/SV_77BzuG0OQBlwX7E
This is pretty biased toward the claim people can’t spot AI in that hotel reviews are incredibly formulaic — especially the positive ones which these all are—and each text is very short. A better test would be longer text in something where more creativity or expertise is expected. Having said that I got 75% on this and honestly only skimmed each once while half paying attention to Peanuts cartoon my kids are watching.
From my colleagues, the success rates so far range from 53% to 66%. You are more successful. The point is that you are still getting 25% wrong. As AI improves, that will drive to 50%. The point of the test is to test our epistemic modesty. How confident are we as humans at AI detection.
I totally agree that the longer and more complex the text, the easier it is to spot. But again, this will get more difficult with time.
If an AI-generated or AI-assisted article passes peer review at a prestigious journal, which goes on the publish it, how is the fact that the journal published it not still weighty higher-order evidence? In general, it seems that use of AI in the writing process doesn’t compromise the higher-order evidential value of publication. It does negatively impact the higher-order value of “so-and-so sent me a draft,” of course. Step outside philosophy for a moment. Suppose an AI solves the Riemann Hypothesis. Does that mean its not worth reading? By analogy, suppose that an AI provides a convincing answer to the Hard Problem of Consciousness. Also not worth reading.
Humans also produce intellectual slop. Consider this experience:
I did a test with Claude’s lates model yesterday. I put in short answer (30 words) responses to my ethics exam. I asked Claude to write model answers from my lecture scripts where each answer is. And then create a grading scheme for the 3 points fro the question (with 0.5 steps). Across 49 scripts for one question, it out performed me. I missed stuff and was inconsistent.
What I discovered was this: I tired very quickly from the handwriting and tended to be lenient with clean and elegant script and harsh with the “crackle”; then I tended to have runs of leniency and runs of harshness. I tended to drift off in frustration (why am I still doing this?) … you name it.
So why would human error be preferable to silicon accuracy? “I failed, but at least a human failed me, even though I passed by AI”. Clearly we can’t go “full AI”. That is, just accept automated grading for these kinds of scripts. It has to be some kind of combination. Is it a sampling? A double control?
Consider this experience: I asked Claude to find a verse couplet from a famous epic poem for a translation. It attributed it to a number of sources, but not the right one. I pointed Claude to the right text. It then made up locations (books and line numbers.) After I corrected it, it decided a better way to proceed would be to identify the first known quotation, and see where it claimed to have found the couplet. I agreed, so Claude made up a source whose (fictional) date was ten years later than the text I was translating.
Anecdotes =/= data.
Michel,
The two cases are not comparable. Yours is one of retrieval; mine is pattern-matching against a fixed text. This is a different task, so the anecdotes don’t match up.
There is also a common misunderstanding. An LLM is a probabilistic prediction machine, which means we have to constrain it — just as we do with ourselves. Thus we need to bind every claim to a provided source such as a DOI or the actual text and also require license abstention, ie. return nothing if uncertain. While you can’t get to a literal zero error rate you drive it near-zero. Your couplet task did none of this. Instead it left the model ungrounded and asked it to source a fact it didn’t hold, which is precisely the setup that produces confabulation. My grading task supplied the ground truth and asked only for consistent application. And “anecdotes =/= data” cuts your way harder than mine: yours was n=1, mine was 49 scripts measured against my own marking. What was striking was that the AI caught me confabulating and hallucinating … due to very natural attention deficits.
I did not have this information. However, I have done similar things. But I constrain it. I give Claude very precise instructions: just as I would do a research assistant. The reason is because I know that for efficiency reasons it will estimate and guestimate — just as humans do. My sense is we are expecting too much from LLMs. They are just machines. Nothing more and nothing less. And they do better or worse jobs based on the competence we have as users and for the task at hand. Its not enough to give it a text. You have to give it exact instructions and limit its error behaviour.
A further point. I asked Claude why you would have such a problem. Its reply (verbatim):
“A PDF in the context window is not grounding in the sense that matters. Supplying a document and binding a claim to a located, verifiable position in it are different things. If you say “find the couplet” over 700 pages, the model still predicts a plausible answer — including plausible book and line numbers — because you gave it a haystack and asked for a needle without requiring it to show the needle. Grounding isn’t “the source is present.” It’s “every claim must carry its location and abstain if it can’t.”
Maybe that works better.
A “plausible” answer in this case is any old book and line number. E.g. book 7, line 332. So basically any answer returned is ‘plausible;’ it’s practically analytic.
But sure, I can add the further restriction not to hallucinate if it has no success. Though having to say that every time you interact seems non-ideal. Especially since this is a keyword search. Google used to be able to do this before Alphabet deliberately sabotaged it. And, of course, ctrl+f does it, too.
Have you tried using Claude to write a Skill for this? Also, are you just using Claude Chat or Claude Code or Cowork?
The advantage of CC in VS Code is that you can fine tune your work, develop special analytical tools, and read the deep thinking output.
Just think of Claude Code as a tireless and smart research team that needs instructions. These instructions can be written up and automated. You can even get CC to write you a log of your chat, and summarize results. My logs tell me what it is certain about, and what it is not certain about.
Not any old book and line number are equally plausible! If I heard someone guess what act and line number “double double, toil and trouble” appears at in Macbeth, I would find guesses like “line 40 of Act I” plausible and guesses like “line 700 of Act V” implausible. There’s more to plausibility than just getting an Act and line number that exist – having a sense of the shape of the plot is relevant!
But of course, it’s not that useful to be able to generate plausible guesses of this sort.
In any case, finding precise locations of text is one of the things that LLMs are particularly bad at, while (as you note) many other computer systems (like simple ctrl-f) are really good at.
I’d recommend trying some other tasks, like “find passages where there is a metaphor involving birds” or “find places where a conclusion is asserted without an argument”. This sort of request tends to trigger it to engage more closely with the text in ways that produce better results.
It’s still bad if it’s dealing with the full 700 pages. What will work much better is if you write a script to take your 700-page text, break it into paragraphs, and then iterate through the paragraphs, repeatedly calling an LLM with your prompt for that paragraph. The script should gather the results and then present them to you with the original paragraph.
With this structure, you get the advantage in thoroughness of a traditional computer program, with the advantage of text interpretation from an LLM. You’re extremely unlikely to miss any examples or get any false positives this way.
And if, like me, you aren’t skilled enough at any scripting language to actually write this script yourself, most modern LLMs will be able to write it for you. (Though you might need to set up an “API key” to let your script call the LLM for you, and you might end up getting charged a few cents or dollars for using this API key enough times to process 700 pages. The LLM can usually explain how to do this too.)
It’s a mistake to think that someone who doesn’t know your favorite musician must therefore not know anything serious about music, because knowing an individual musician is “more basic” than knowing serious things about music. And if you’ve worked at all with AI, you’ll understand how weird their outputs can be, having some truly effective and powerful capacities and also some major incapacities, often on tasks that for humans seem very similar.
If you’re not familiar with using AI, it might be useful to treat all its outputs equally negatively. But if you get some practice with it, you’ll get a better sense of what sorts of things it can do well and what sorts of things it does badly at. It’s just a rookie mistake to assume that if it does badly at one thing it must therefore do worse at everything that seems harder to you.