{"id":"8c0f473c-3980-49e1-aa47-2820ba1a5f63","arxiv_id":"2507.09676","paper_version":1,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Because normative truths are largely asystematic, language models cannot rely on the systematicity of truth to self-complete and self-correct in ethics and politics, leaving final moral judgment to humans.","lead":"A philosophical essay argues that a key assumption behind AI optimism, that all true statements form a connected web, fails in ethics and politics. If correct, language models will be less able to complete missing moral knowledge, so people must stay responsible for final practical decisions.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's core inference—that absent systematicity, self-completion and self-correction must be hampered—is under-argued because it slides from truth-level asystematicity to corpus-level unlearnability; the bridging premise is asserted rather than defended.","rationale":"The reader identified the weakest assumption as the claim that LLMs can only move beyond training data by exploiting inferential redundancy of a domain. My concern is closely related but more specific and, I think, more damaging in a way the reader did not fully develop. The paper hinges on a slide from 'truth is not systematic' to 'the domain cannot be statistically modelled.' But the object of LLM modelling is not the truth directly; it is text generated under the influence of truth, belief, and argumentative practice. A domain can be asystematic at the level of its true propositions and still be highly regular at the level of its discourse: there are patterns in how conflicts are described, how judgments of importance are made, and how regrets are expressed. Indeed, the paper itself notes that LLMs could be trained to be sensitive to consistency and coherence indirectly through training objectives and human feedback (Section 3), and it discusses Kaleido, which successfully represents value conflicts using statistical methods (Section 4). The existence of Kaleido is evidence that the asystematicity of normative truth does not preclude learning that something about the domain is structured. The paper's response—that Kaleido cannot go beyond its training data—is fair, but it conflates the inability to go beyond training data in a specific way (via systematicity) with the general impossibility of going beyond it via any mechanism. The paper is carefully hedged and philosophically sophisticated, and I do not think it should be rejected. But the central conditional's antecedent and consequent are not tightly connected: the paper does not address the possibility that LLM competence in normative domains could be grounded in the statistical regularities of argumentation and human judgment rather than in the coherence of the underlying truths. My proposed concrete test would settle this empirically. Without such a test, the argument remains a plausible philosophical speculation rather than a demonstrated implication. I therefore recommend a CONDITIONAL verdict: accept the philosophical analysis of asystematicity and its implications for agency, but require the empirical bridging claim to be made explicit and tested before the stronger claim about LLM progress is drawn. This is a partial agreement with the reader because the reader saw the threat but did not identify the specific conflation I am flagging: the distinction between truth-level systematicity and corpus-level statistical regularity.","tokens_in":23235,"tokens_out":2281,"duration_ms":27629,"concrete_test":"Conduct a controlled empirical evaluation directly testing the paper's key inference. Construct two comparable corpora: one of normative texts (e.g., case-based ethics literature, policy documents, annotated value-conflict datasets like ValuePrism) and one of geographically systematic facts. Train or fine-tune identical small LLM architectures on held-out portions of each, then measure (a) accuracy of recovering held-out normative statements vs. geographical statements, and (b) success of self-correction when the training data is corrupted with plausible errors. If the LLM recovers and corrects held-out normative statements at comparable rates to geographical statements despite the presumed asystematicity of normative ground truth, the paper's central inference fails.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central argument is conditional: if normative truth is asystematic, then LLMs cannot leverage systematicity to progress toward comprehensiveness. The paper's contribution is precisely this bridge from a philosophical thesis about the structure of normative truth to an empirical claim about LLM capability. The bridge, however, is weaker than it appears. As the paper itself concedes (Section 3), LLMs are trained on corpora, not directly on the fabric of facts. The relevant question for LLM capability is therefore not whether the ground truth forms a systematic web, but whether the training distribution over text—including the vast literature of ethical deliberation, case-based reasoning, moral dilemmas, and situated judgments—is statistically structured enough for models to learn and generalize. A domain can be asystematic at the level of its true propositions and yet be highly regular at the level of its discourse: conflicts, trade-offs, and incommensurability themselves have recurring vocabularies, argumentative patterns, and contextual cues. The paper's own example of Kaleido (Section 4) is evidence that statistical methods can capture something real about value conflicts, and the paper's response—that Kaleido cannot go beyond its training data—conflates the inability to go beyond training data in one specific way (via systematicity) with the absence of any other mechanism. The Conclusion's admission that 'alternative ways for LLMs to move beyond their training data may yet emerge' is not a minor hedge; it names exactly the missing premise. The claim that human agency is ineliminable (Section 5) is better supported, because it rests on an argument about the first-personal nature of practical deliberation independent of LLM limitations.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper argues that the optimism that LLMs can overcome gaps and inaccuracies in training data by exploiting the systematicity of truth fails in normative domains. Drawing on the value-pluralist tradition (Berlin, Williams, Nagel, Chappell), it claims that normative truths are at least partly asystematic: values are irreducibly plural, incompatible, and often incommensurable, so truths about values do not form a consistent, coherent, inferentially interlocking web. It then infers that because LLMs cannot leverage systematicity in these domains, their self-completion and self-correction will be correspondingly harder, and comprehensive modelling of normative domains will be hampered. The paper further argues that asystematicity amplifies the first-personal and personal dimensions of practical deliberation, so the final practical judgment cannot be outsourced to an AI. The argument is explicitly conditional and carefully qualified throughout.","tokens_in":23496,"tokens_out":4158,"duration_ms":52401,"significance":"If the argument succeeds, it identifies a principled structural limit on LLM performance in ethics and politics, and it grounds a positive role for human agency in AI-assisted practical deliberation. The paper is significant for both AI ethics and the philosophy of AI: it connects a long-standing value-pluralist literature to a concrete capability claim, engages with existing systems such as ValuePrism and Kaleido, and carefully distinguishes output-level consistency and coherence from generation that actually relies on the wider systematicity of truth. Its conclusions are presented conditionally, which is methodologically honest and makes the paper a useful starting point for further empirical work. The main limitation is that the inference from truth-level asystematicity to corpus-level unlearnability is asserted rather than demonstrated, and the paper's own concessions in the Conclusion leave this bridge under-protected.","major_comments":[{"comment":"The central inference from asystematicity of ground truth to reduced LLM capability is under-argued in the paragraph beginning \"The asystematicity of normative truths in turn has implications for the prospects of LLMs.\" LLMs are trained on corpora, not directly on the fabric of facts, and the paper itself notes in §3 that they internalise statistical patterns in text. A normative domain whose true propositions do not form a systematic web may nevertheless exhibit stable regularities in discourse about conflicts, trade-offs, and incommensurability; the paper does not explain why these corpus-level patterns cannot support learning and generalisation. Since the Abstract's claim that asystematicity renders progress \"correspondingly harder\" depends on this bridge, the argument needs either a defended premise that alternative mechanisms cannot compensate or a restriction of the conclusion to the claim that LLMs cannot rely on the specific route of systematicity.","section":"§4"},{"comment":"The Conclusion's concession that \"alternative ways for LLMs to move beyond their training data may yet emerge\" undermines the inference from \"LLMs cannot leverage the systematic harmony of these domains\" to \"it should to that extent be harder for LLMs to comprehensively model normative domains.\" If retrieval-augmented generation, symbolic reasoning, external memory, or human-in-the-loop scaffolding could supply the needed redundancy, the predicted difficulty would not follow. The paper should either address these mechanisms and explain why they cannot restore comprehensiveness, or state more modestly that the systematicity-based route is unavailable without thereby asserting an overall increase in difficulty.","section":"§6"},{"comment":"The claim that a pluralist model such as Kaleido \"cannot overcome the limitation imposed by the asystematicity of normative truth on its capacity to move beyond its training data\" is presented as a direct consequence of asystematicity, but the example only shows that one system trained on synthetic GPT-4 data does not extrapolate via systematicity. It does not test whether other statistical or non-statistical mechanisms could improve coverage. This is an empirical premise that needs support or explicit identification as an open empirical question.","section":"§4"}],"minor_comments":[{"comment":"The text refers to \"autoregressive LMMs\" where the intended term is presumably \"LLMs\"; this should be corrected.","section":"§3"},{"comment":"In the discussion of Sorensen et al., the citation appears as \"Sorensen et al,. 2024\" with a stray comma before the period, and the model name \"Value Kaleidoskope\" is spelled differently from the standard \"Kaleidoscope\" in the cited paper.","section":"§4"},{"comment":"The claim that Kaleido's relevance scores \"measure the statistical relevance of a type of consideration to a type of situation\" while importance \"goes significantly beyond such merely statistical relevance\" is asserted rather than argued; a sentence explaining why importance cannot be approximated by richer contextual features would help the reader assess this step.","section":"§5"}],"recommendation":"major_revision","confidential_remarks":"The paper is a good fit for Philosophy & Technology, and the conceptual material on asystematicity and first-personal practical deliberation is strong. The main revision should align the Abstract and Conclusion with the actually defended claim: rather than asserting that asystematicity makes comprehensive modelling harder, the paper should either defend the missing bridge premise about alternative learning mechanisms or explicitly restrict the conclusion to the unavailability of the systematicity-based route. This is a fixable load-bearing gap rather than a fatal flaw."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, the punchline. This is a genuinely useful paper, and the part that matters most isn't the part that's easiest to attack. The novel move is tying Amodei's 'web of truth' optimism to value pluralism: if normative truths really are asystematic, then the inferential redundancy that supposedly lets LLMs self-complete and self-correct just isn't available in ethics and politics. Section 5, on the irreducibly first-personal character of practical deliberation, is the strongest section and stands independent of the LLM discussion.\n\nWhat the paper does well: it's explicit about its conditionality—'insofar as' and 'to that extent' do real work throughout. It engages seriously with the value pluralist tradition (Berlin, Williams, Nagel, Chang) and with the existing AI ethics literature (Sorensen et al.'s ValuePrism/Kaleido, Goodman on hard choices). The discussion of training as indirect pattern-learning is fair and doesn't overclaim what current models do. The author also flags his own prior work only as elaboration, not as sole basis, so no circularity problem.\n\nThe soft spot is the bridge from truth-level asystematicity to corpus-level unlearnability. The paper admits in Section 3 that LLMs are trained on text, not on facts, and in the Conclusion that other mechanisms for moving beyond training data may emerge. But the main argument never really confronts the possibility that normative discourse—the way we talk about conflicts, trade-offs, incommensurability—is itself statistically regular, even when the underlying truths are asystematic. That's a real gap. The Kaleido example actually undercuts the author a bit: it shows an LLM can learn to generate and filter considerations about value conflicts. The response is that Kaleido can't go beyond its training data in the systematicity-based way, but that's exactly the premise that needed defending. This is the weakest link, but it's not fatal because the conclusion is hedged as 'harder' rather than 'impossible.' Still, a referee should push on it.\n\nThe human agency argument, by contrast, is solid. The distinction between the impersonal 'What is to be done?' and the first-personal 'What should I do?' is a classic Williams point, applied cleanly to AI advisors. Even a perfect normative model can't make the agent's judgment for them.\n\nWho's this for? Philosophers of AI and AI ethics researchers. It's not an empirical paper; judging it as one would be a category error. It deserves serious peer review—the kind where the referee pushes on the corpus-level bridge and the author either concedes more or engages head-on with the discourse-regularity objection. I'd accept and engage.","headline":"A serious value-pluralist challenge to LLM progress optimism, with a soft empirical bridge but a robust agency thesis.","tokens_in":24074,"tokens_out":3396,"would_cite":true,"duration_ms":39296,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper argues that LLMs can only fill gaps in their data by exploiting the systematicity of truth, and that normative domains are too asystematic for that to work.","keywords":["large language models","systematicity of truth","value pluralism","normative domains","self-completion","self-correction","practical deliberation","AI ethics"],"falsifier":"A controlled benchmark would settle it: take matched sets of facts from a systematic domain (e.g., geography) and a normative domain (e.g., what values bear on novel moral situations), withhold or corrupt the same fraction of each, and test whether an LLM recovers the withheld truths at comparable rates after training on the rest. If the model extrapolates normative truths as successfully as geographical ones despite the lack of inferential redundancy, the central claim would be false.","tokens_in":23025,"feed_emoji":"⚖️","tokens_out":6709,"duration_ms":67122,"temperature":0.7,"pith_summary":"This paper argues that the optimistic assumption behind large language models' ability to comprehensively model the world—that true statements hang together in a systematically interlinked web—does not hold in normative domains. Drawing on value pluralism, it holds that truths about ethics and politics are often irreducibly conflicting, incommensurable, and disconnected, so LLMs cannot exploit inferential redundancy to fill gaps or correct errors there. As a result, the paper concludes, comprehensive modelling of normative domains should be harder for LLMs, and the final practical decision must remain the agent's own. A sympathetic reader should care because this gives a principled reason to expect AI moral advice to remain advisory no matter how much training data accumulates.","feed_headline":"LLMs rely on truth's web, and ethics has no web to lean on","feed_subtitle":"If normative truths are asystematic, models cannot infer missing values, and final choices stay with us.","key_machinery":"The central machinery is the distinction between systematic and asystematic truth. Systematicity names the property of a body of truths being not merely consistent (free of contradiction) but coherent: truths stand in relations of rational support, so each truth can be recovered from others—an inferential redundancy that LLMs can in principle exploit to fill gaps and correct errors. The paper then uses value pluralism—the thesis that values are irreducibly multiple, often incompatible, and incommensurable—to establish that normative truths lack this redundancy, and uses the distinction between an impersonal 'What is to be done?' and a first-personal 'What should I do?' to show why the resulting judgements of importance remain the agent's own.","core_discovery":"The central claim is that LLMs' capacity to progress beyond incomplete and inaccurate training data depends on the systematicity of truth—the consistency and inferential coherence of true statements—and that in normative domains this systematicity is largely absent. On the paper's account, value pluralism shows that values are irreducibly diverse, incompatible, and incommensurable, so normative truths form a fragmented and tension-ridden landscape rather than a web. Consequently, the very mechanism that promises self-completion and self-correction in systematic domains—deriving missing truths from surrounding truths—cannot be relied on in ethics and politics. The paper further claims that this asystematicity intensifies the first-personal and personal character of practical deliberation: deciding what matters in a hard choice cannot be outsourced to an algorithm because it requires the agent's own judgement of importance and authenticity.","pith_inferences":["Editorial inference: this argument predicts an empirically testable asymmetry—a model trained on sparse or corrupted data should recover withheld facts more reliably in systematic domains than in normative domains, once difficulty is matched.","Editorial inference: if the paper is right, current alignment techniques that aggregate human preferences may not merely risk 'washing out' value conflicts; they may also create an illusion of comprehensiveness precisely where the model is depending on training data rather than inference.","Editorial inference: the same asystematicity constraint would apply to future models with retrieval or memory scaffolding only if those mechanisms import normatively resolved answers from outside the model; otherwise they just shift the gap rather than closing it."],"forward_implications":["LLM self-completion and self-correction from training-data gaps should work in systematic empirical domains such as geography, but should falter precisely on normative questions where values conflict.","Value-pluralist training data that explicitly represents conflicts can help surface considerations, but cannot supply the inferential redundancy needed to extrapolate beyond the data.","The less systematic a domain is, the more first-personal judgement is required, so even a highly capable moral advisor leaves the final 'should I really do this?' question to the agent.","Fine-tuning methods aimed at internal consistency and coherence cannot substitute for the systematicity missing from the normative landscape itself."],"supporting_citations":[{"why":"Supplies the target thesis: the claim that true things form a web LLMs can use to overcome gaps and errors in training data.","marker":"Amodei, 2024"},{"why":"Provides the account of systematicity as inferential redundancy and the example showing how an excised truth can be recovered from the rest.","marker":"Rescher (2005, 5)"},{"why":"Supplies the value-pluralist claim that the realisation of some values can only be obtained at the expense of others.","marker":"Berlin, 2002, 213–214"},{"why":"Defines incommensurability as the absence of a common currency or lexical priority rule, grounding the asystematicity claim.","marker":"Berlin & Williams, 1994, 306"},{"why":"Provides the account of conflicts of ought used to show that conflicting normative truths can be genuine and irreducible.","marker":"Williams, 1973, 171"},{"why":"States the key contrast that evaluative beliefs do not form a consistent system because they are not attempts to describe a single world.","marker":"Nagel, 2001, 108–9"},{"why":"Argues that moral philosophy seeks no hidden underlying order, supporting the absence of systematicity in normative domains.","marker":"Wiggins & Williams, 1978, xxxviii–xxxix"},{"why":"Documents the value-pluralist dataset and model approach and its limits, serving as the current state of the art the argument must accommodate.","marker":"Sorensen et al. (2024)"},{"why":"Supplies the notion of hard choices that remain on a par once all information is in, anchoring the argument that agents must supply judgements of importance.","marker":"Chang (2017)"}],"fun_headline_variants":["Ethics lacks truth's web, so AI can't self-correct","When truth isn't systematic, AI can't do ethics for you","Normative truths are fragmentary, so LLMs can't infer gaps","AI's self-correction fails in ethics, where truths clash"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument rests on the premise that LLMs can only move beyond their training data by exploiting a domain's inferential redundancy, so that where that redundancy is absent, no other mechanism—retrieval, symbolic reasoning, memory, or human-in-the-loop scaffolding—can make coverage comprehensive.","fun_headline_variants_meta":{"raw":{"variants":["Ethics lacks truth's web, so AI can't self-correct","When truth isn't systematic, AI can't do ethics for you","Normative truths are fragmentary, so LLMs can't infer gaps","AI's self-correction fails in ethics, where truths clash"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00082,"raw_usage":{"total_tokens":3583,"prompt_tokens":936,"completion_tokens":2647,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":552,"completion_tokens_details":{"reasoning_tokens":2571}},"tokens_in":552,"tokens_out":2647,"duration_ms":18504,"temperature":1.0,"reasoning_tokens":2571,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T17:50:44.799121+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A controlled benchmark would settle it: take matched sets of facts from a systematic domain (e.g., geography) and a normative domain (e.g., what values bear on novel moral situations), withhold or corrupt the same fraction of each, and test whether an LLM recovers the withheld truths at comparable rates after training on the rest. If the model extrapolates normative truths as successfully as geographical ones despite the lack of inferential redundancy, the central claim would be false.","supporting_citations":[],"review_version":1}