## Does Knowing the Source Is an LLM Rationally Change Your Epistemic Situation? ### 1. The Standard Bayesian Case, and What It Presupposes The most natural position -- and the one LLMs themselves reliably reproduce when asked -- runs as follows. Knowing that a text was produced by an LLM is a piece of evidence about the reliability of that text's source. If LLMs have a non-trivial rate of producing plausible-seeming but subtly flawed philosophical arguments, then learning that a given argument came from an LLM should rationally lower your credence in its soundness, prior to checking it yourself. This is straightforward Bayesian updating: you have a prior probability that any given philosophical argument is sound; you update that prior based on source information; your posterior credence shifts accordingly. This position has genuine force. Floridi et al. provide the machinery for it. They argue that LLMs are "stochastic engines of text" whose outputs "lack explicit logical rules, deliberate hypothesis testing, or reference to an external world model." They document the phenomenon of "over-abduction": "a human reasoner might say, 'I'm not sure; more information is needed,' while the LLM often makes a guess regardless." The concern is that LLMs systematically produce outputs that mimic the form of good reasoning without reliably tracking its substance -- that "the model does not understand what an explanation is, but it produces text that follows the typical phrasing and structure of explanations." If this is right, then knowing you are reading an LLM output gives you reason to be more cautious, just as knowing you are reading the work of a first-year philosophy undergraduate gives you reason to expect more errors than if you were reading the work of a seasoned professional. But notice what this position presupposes. It presupposes that you are in the position of a testimony-receiver -- someone who is going to take the source's word for things without fully checking them. In the testimony model, source reliability is central because you are delegating epistemic labour. You believe your doctor about your blood test results partly because doctors are generally reliable about such things. If you learned that "Dr. Smith" was actually a large language model in a lab coat, your credence in the diagnosis should rationally change -- because you were relying on the source's reliability as a substitute for your own assessment. The question is whether philosophical evaluation works like testimony-reception or like something else. ### 2. The Case for Screening Off Consider a different model: direct assessment. You read a philosophical argument. You check whether the premises are true (or at least defensible). You check whether the inferences are valid. You check whether the key distinctions are well-drawn, whether costs are honestly accounted for, whether objections are fairly treated. If you do all this -- if you evaluate the argument on its own terms, by its own merits -- what additional epistemic work is source knowledge doing? The screening-off thesis holds: once you have done a thorough direct assessment of a philosophical text, source information is screened off by the evidence you have gathered from the text itself. Learning that a valid argument was produced by a random number generator does not make the argument less valid. Learning that an insightful distinction was drawn by an LLM does not make the distinction less insightful. The logical properties of an argument are intrinsic to the argument; they do not change when you learn something about its causal history. [[Timothy Williamson|Williamson]]'s own framework, I would argue, supports rather than undermines this position. He describes abductive evaluation as a matter of ranking theories by their "intrinsic virtues": "Apart from its relation to E, the more T has the intrinsic virtues of a good theory, the better (ceteris paribus). It should be elegant and unified, not arbitrary, gerrymandered, ad hoc, or messily complicated" (Williamson, p. 354). The word "intrinsic" is doing heavy work here. Theoretical virtues are properties of the theory, not of the theorist. The ranking is of theories, not of theorists: "Such theories rank low on the abductive scale" (p. 366). When Williamson diagnoses problems in the philosophical community's work -- "Strikingly, the philosophical community showed very little aversion to the multiplication of complication" (p. 369) -- he is diagnosing properties of papers, not properties of minds. If the standards that define good philosophy are text-internal -- precision, cost-accounting, non-ad hocness, defeater-sensitivity, the integration and simplicity Williamson champions -- then they are assessable by reading the text. You do not need to know who wrote it. This is, of course, the entire rationale for blind review. [[Luciano Floridi|Floridi]] himself raises the question without resolving it. After discussing how LLM outputs can be "similar or even identical" to human-produced hypotheses, he writes: "if an AI can generate the same explanatory hypothesis a human would, does it matter that the process was different? From an epistemological standpoint, perhaps yes -- justification is significant -- but regarding the content of the hypothesis and our interpretation of it, maybe not." The concession is striking. Floridi acknowledges that at the level of content, provenance may be irrelevant. His "perhaps yes" for the epistemological standpoint gestures at a justification-based concern, but does not develop it. The question is whether that gesture survives scrutiny. ### 3. The Screening-Off Question in Detail When does direct assessment screen off source information? I think the answer depends on how completely the text makes its reasoning accessible. In empirical domains, source information resists screening off because the text cannot contain all the evidence. A medical diagnosis depends partly on clinical experience, on having seen similar cases, on subtle pattern recognition that does not fully appear in the written report. If you cannot independently verify all the empirical claims, you are still partially relying on the source's reliability. Knowing the source is unreliable (or of unknown reliability) then remains relevant even after you have read the report carefully. But philosophy is different in a crucial respect. In philosophy -- at least in the analytic tradition -- the standard of evaluation is that the reasoning should be in the text. A philosophical argument is supposed to make its commitments explicit, lay out its inferences transparently, and meet objections on the page. As [[Notes/Provenance is not the right kind of variable in philosophical evaluation|the provenance note]] captures: "Treating provenance as epistemically relevant to argument quality imports a model from other domains -- testimony-based belief formation -- that doesn't fit philosophy. In medicine or history, you might reasonably defer to a reliable source without personally verifying claims. But in philosophy, 'verifying' a claim just *is* grasping and assessing the reasons offered. There is no separate checking procedure you can skip." This is not to say that philosophical evaluation is infallible. A competent reader can miss a subtle equivocation or an illicit premise-shift. But if they miss it, they miss it regardless of who wrote the text. Learning that the text was LLM-generated does not help them find the equivocation. It might make them look harder -- but that is a point about epistemic motivation, not about the logical status of source knowledge. There is a useful distinction here between what we might call justificatory norms and triage norms. Justificatory norms concern whether the reasons on the page are good. Triage norms concern how you allocate your finite attention. Source knowledge is relevant to triage: given limited time, it is rational to prioritise reading texts from sources with strong track records. But triage is a practical matter, not a philosophical evaluation. When you actually sit down and read and assess, the justificatory norms take over, and they are text-internal. ### 4. The Asymmetry Question Source knowledge might matter differently for different kinds of philosophical evaluation. This is worth taking seriously rather than rushing to a single answer. For validity checking, source knowledge is almost certainly irrelevant. Validity is mechanically assessable. A valid argument is valid regardless of who produced it. If anything, knowing the source is an LLM might make you *more* careful about checking validity, but that is motivational, not epistemic. For assessing originality, source knowledge might seem more relevant. To judge whether a philosophical move is genuinely novel, you need to know the landscape -- what has been said before, what is already in the literature. If you suspect the source is an LLM, you might suspect that the argument is a sophisticated recombination of existing material rather than a genuinely new contribution. But notice: this is still not about the argument's quality. An argument that recombines existing material in a genuinely illuminating way is a genuine contribution. And whether the recombination is illuminating is assessable from the text. The question of whether it has been done before is a question about the literature, not about the source. For assessing depth or significance, the picture is genuinely unclear. Does knowing that a text was LLM-generated give you reason to think it lacks some quality that is not fully visible in the text? Perhaps. If you think that philosophical depth requires something beyond what is textually manifest -- a kind of hard-won understanding that shapes judgment in ways that do not fully appear on the page -- then you might think source knowledge is relevant because it tells you something about the likelihood of that extra quality being present. But this is precisely the contested question. Those who hold that philosophical evaluation is artefact-level will deny that there is any such extra quality; those who hold that philosophy requires inner understanding will affirm it. The dispute about source relevance reduces to a deeper dispute about the nature of philosophical quality. ### 5. Structural and Historical Analogies The four-colour theorem, proved by Appel and Haken in 1976 with essential computer assistance, is the most instructive analogy. The mathematical community genuinely resisted the proof. The resistance had several sources: it could not be verified by hand; it relied on a massive case-checking procedure that was not surveyable; mathematicians felt they lacked understanding of *why* the theorem was true. The question was whether a proof that could not be fully checked by human cognition counted as a proof at all. The resistance was not irrational, but it was also not ultimately vindicated. Over time, the proof was accepted -- partly because it was independently verified by other computer-assisted methods, and partly because the community came to terms with the fact that the logical structure of the proof was sound, even if it exceeded human surveyability. What mattered in the end was whether the proof met the standards of mathematical validity, not whether a human could hold the whole thing in their head. This suggests a general lesson: process-knowledge is relevant when it bears on whether the standards have been met, and irrelevant when it does not. In the four-colour case, the relevant concern was surveyability -- whether anyone could actually verify all the cases. That is a concern about whether the proof meets mathematical standards, not an objection to its causal history per se. Once the verification concern was addressed, the process-resistance dissolved. Anonymous and pseudonymous philosophical work provides another angle. Kierkegaard published under pseudonyms; many significant philosophical contributions have been published anonymously. These works are evaluated on their merits. We do not typically think that learning the identity of a previously anonymous philosopher changes the quality of the arguments they published. We might learn something about their likely biases or blind spots, but that is triage information, not justificatory information. Ghost-written work raises a slightly different issue. If a prominent philosopher publishes work that was actually written by a graduate student, we might feel deceived -- but about what? Not about the quality of the argument, which remains what it is. The deception concerns attribution and credit, which are social and institutional matters, not philosophical ones. ### 6. Priors, Evidence, and the Right Description Here is a position that attempts to reconcile the standard Bayesian case with the screening-off thesis: source knowledge rationally sets your prior expectations, but can be overridden by evidence from the text itself. On this view, learning that a text was LLM-generated gives you reason to expect a higher error rate -- more subtle equivocations, more plausible-seeming but ultimately hollow arguments. You approach the text with appropriately calibrated caution. But as you read, you gather evidence. If you check the argument and find it sound, if you test the distinctions and find them sharp, if you look for the characteristic failure modes and find none -- then the evidence from the text overrides the prior set by source knowledge. Your posterior credence in the argument's quality should be high, regardless of its provenance. Is this "provenance matters" or "provenance doesn't matter"? I think the right description is: provenance matters for attention allocation and initial calibration, but not for final evaluation. It is relevant to how you approach the text, but not to what you conclude once you have done the work. This is a coherent position, but note how much it concedes to the screening-off thesis. The provenance is doing no work in the final assessment. It is functioning as a triage heuristic, not as a justificatory consideration. One might object: even after thorough reading, you can never be fully certain you have caught everything. Source knowledge gives you reason to think there might be subtle errors you missed. But this applies equally to human-produced philosophy. Every philosophical text might contain errors you missed. The question is whether LLM provenance gives you *additional* reason to worry, beyond the general fallibility of your reading. If you have read carefully and found the argument sound, it is not obvious that it does -- unless you have a specific theory about the kinds of errors LLMs systematically make that are undetectable by careful reading. And such a theory would need to explain why these errors are undetectable: if they leave no trace in the text, how do you know they are there? ### 7. Is This Question Philosophical or Sociological? Some resistance to LLM-produced philosophy is probably rational caution; some is probably status anxiety or prejudice. How do you tell the difference? A principled test: resistance is rational when it can specify what text-internal failure it is guarding against. "I am cautious about LLM-produced philosophy because LLMs sometimes produce subtle equivocations that are hard to catch" -- this is a rational calibration of prior expectations. It predicts a specific kind of failure and can be checked by careful reading. Resistance is not rational -- or at least not *philosophically* rational -- when it cannot specify any such failure. "I am uncomfortable with LLM-produced philosophy because it just does not feel like real philosophy" -- this is either a placeholder for a more specific concern (which should be articulated) or it is a sociological attitude dressed as a philosophical judgment. [[Notes/The appearance-reality gap collapses for competent readers|The appearance-reality note]] is relevant here. For competent philosophical readers, "looks like good philosophy" in the strong sense -- not genre markers, but genuine constraint satisfaction -- just means "the standards are satisfied in the text." If the appearance-reality gap collapses for expert readers, then Floridi-style "mere appearance" rhetoric becomes question-begging when deployed without identifying specific text-internal failures. But I want to resist closing this off too quickly. There might be a legitimate concern that is not about specific text-internal failures but about a systematic property of LLM outputs. If LLMs produce arguments that are locally sound at every step but globally miss something -- if they satisfy every individual constraint but fail to cohere into genuine understanding -- that would be a real problem. It would also be very hard to detect, precisely because it is not localised in any particular passage. Whether this is a real phenomenon or a theoretical spectre is an empirical question that deserves investigation rather than assumption. ### 8. Where the Philosophical Action Is This question -- whether source knowledge matters for philosophical evaluation -- is not a methodological footnote. It connects to deep issues in epistemology about the nature of justification. If philosophical quality is fully determined by text-internal properties, then philosophy is a domain where justification is transparent -- where the reasons for accepting or rejecting an argument are accessible to any competent reader. This is a strong claim about the nature of philosophical justification. It entails that there is no philosophically relevant form of understanding that transcends what can be expressed and assessed in text. Tacit knowledge, hard-won intuition, the kind of judgment that comes from decades of immersion in a problem -- all of these must either be expressible in the text (and so assessable) or irrelevant to the quality of the philosophy (and so not part of evaluation). Conversely, if source knowledge genuinely matters for philosophical evaluation -- if knowing that a text was produced by an LLM rationally changes your assessment of its philosophical quality even after careful reading -- then there must be some dimension of philosophical quality that is not textually manifest. This would be a significant claim about the limits of transparency in philosophical justification. Williamson's framework, I would speculate, pulls toward the transparency view. His emphasis on theoretical virtues as "intrinsic" to the theory, his insistence that the criteria are publicly applicable even if their ultimate justification is unclear, his focus on properties of papers rather than properties of minds -- all of this suggests a picture on which evaluation is artefact-level. The standards are the standards. Meet them and you have done good philosophy. Fail them and you have not. Who you are does not enter into it. But Williamson also says something that complicates this. He argues that simplicity is not merely aesthetic but truth-conducive because it prevents over-fitting. Over-fitting is a property of the method by which a theory was produced -- fitting to noise rather than signal. If over-fitting is detectable only by knowing something about the production process, then production process is relevant to evaluation. However, Williamson's own examples suggest that over-fitting is detectable from the text: an over-fit theory is gerrymandered, ad hoc, messily complicated. These are textual properties. You can see over-fitting in the paper. The [[John Bengson|Bengson]] framework adds another dimension. Their Tri-Level Method requires accommodation (the theory handles the data), explanation, substantiation, and integration. These are all assessable from the text. Their account of understanding -- that inquirers "fully grasp" theories with properties including accuracy, reason-based support, robustness, illumination, orderliness, and coherence -- defines the "reason-based" property as a property of the theory, not the theorist: a theory is reason-based when it is "positively supported by considerations, beyond mere coherence, that speak in favor of its accuracy." Whether those considerations exist is a fact about the theory's content, not about who produced it. The deepest issue, then, is whether philosophical quality is fully textually manifest. If it is, source knowledge is irrelevant to evaluation and relevant only to triage. If it is not, there is philosophical work left for source knowledge to do. I find myself genuinely uncertain which is correct, and I think that uncertainty is itself philosophically significant -- it marks an unresolved question about the nature of philosophical justification that the LLM debate has brought into sharp focus but did not create. --- *La domanda se il sapere da dove viene un testo ne modifichi il valore filosofico ci costringe a chiarire che cosa intendiamo per 'valore filosofico' -- e questo chiarimento potrebbe essere la parte piu difficile dell'intera questione.*