# Lipton Chapter 2 (Explanation) — Relevance to Generating Philosophy with AI ## 1. Chapter Summary Chapter 2 of Lipton's *Inference to the Best Explanation* asks what explanation is — or more precisely, what has to be added to knowledge to yield understanding. Lipton opens by noting that the question can be put in several ways, but all revolve around a single gap: > "Typically, someone who asks why something is the case already knows that it is the case. The person who asks why the sky is blue knows that it is blue, but does not yet understand why. The question about explanation can then be put this way: What has to be added to knowledge to yield understanding?" This gap between knowing-that and understanding-why is the chapter's organising problem. Lipton treats it as a test for models of explanation: any adequate account must allow for the difference between possessing a true belief and grasping why the thing believed is the case. Before canvassing models, Lipton addresses the why-regress — the child's discovery that every answer to a "why?" question can itself be met with "why?" He argues that the regress is not genuinely skeptical. Rather than requiring that explanations be self-explanatory or that every link in the chain itself be understood, Lipton observes that explanation does not work like a substance that has to be possessed in order to be transferred: > "Understanding is not like a substance that the explanation has to possess in order to pass it on to the phenomenon to be explained. Rather than show that explanation is impossible, the regress argument brings out the important facts that explanations can be chained and that what explains need not itself be understood." This matters for any model of explanation: a model that requires full understanding of the explanans is too strong. Lipton then makes a further observation about the epistemology of explanation. He notes that there is no skeptical argument against explanation analogous to Hume's argument against induction. The reason is structural: Hume exploits a gap between meeting our inductive standards and actually being correct, but for explanation we cannot clearly articulate a comparable gap: > "We do not appear to know how to make the contrast between understanding and merely seeming to understand in a way that would make sense of the possibility that most of the things that meet all our standards for explanation might nonetheless not really explain." This is a striking remark. It suggests that our grasp on explanation is, in some respects, worse than our grasp on inference: we can at least say what inferences are *trying* to do (track truth), but we lack a conception of understanding independent of whatever our explanations provide. The bulk of the chapter canvasses five models of explanation, finds each wanting, and sets up a sixth (the causal model) for the following chapter. Briefly: The **reason model** says that to explain a phenomenon is to give a reason to believe it occurs. Lipton rejects this on the grounds that explanation typically adds something beyond a reason for belief, since we already have a reason for believing the phenomenon when we know it obtains. The model also fails because of *self-evidencing explanations* — cases where the evidence for the explanans is the very explanandum. A galaxy's recession explains its red-shift, but the red-shift provides the evidence for the recession. Lipton argues that such explanations are "ubiquitous" and perfectly legitimate, but would be illicit if the reason model were correct, since accepting them as providing reasons for belief would be viciously circular. The **familiarity model** says that explanation makes the unfamiliar familiar. Lipton is more sympathetic to its "surprise" version — an explanation removes the tension between the phenomenon and our background beliefs — but argues that it fails because we often explain familiar phenomena that are not surprising. The rattle in a car is familiar and consistent with everything else one believes, yet still admits of explanation. The **deductive-nomological (D-N) model** says that explanation is deduction of the phenomenon from laws and initial conditions. Lipton notes that the D-N model handles self-evidencing explanations better than the reason model, and captures the aspiration of many scientific explanations. But it fails to account for what Lipton calls the "asymmetries of explanation": cases of deductive symmetry where the direction of explanation runs only one way. The Doppler law lets us deduce recession from red-shift just as easily as red-shift from recession, but the red-shift does not explain the recession. Lipton notes that the D-N model is isomorphic to the hypothetico-deductive model of confirmation, and suffers analogous weaknesses. The **unification model** says that understanding comes from seeing how a phenomenon fits into a unified whole. Lipton gives it qualified praise: > "This conception chimes with the ancient idea that to understand the world is to see unity that underlies the apparent diversity of the phenomena." He observes that it handles both the knowing-that/understanding-why gap and self-evidencing explanations: > "We can know that something is the case without yet being able to fit it together appropriately with other things we know, so there can be knowledge without understanding. Self-evidencing explanations are also accounted for, since a piece of a pattern may provide evidence for the pattern as a whole, while the description of the whole pattern places the piece in a unifying framework." But Lipton raises three objections. Unification is difficult to analyse without reducing it to features of vocabulary rather than features of the world. The model may not allow sufficiently for the why-regress, since a unifying explanation seems to need to be itself unified. And the model does not adequately capture singular causal explanations, which account for much of ordinary explanation without providing any strong form of unification. The **necessity model** says that explanation shows the phenomenon *had* to occur. Lipton finds it too strong — outside pure mathematics, few explanations demonstrate logical necessity — and argues that weaker notions of necessity resist analysis. Lipton concludes that none of the five models is adequate as a general account, though each captures distinctive features of certain explanations. He prepares the ground for the causal model in Chapter 3. ## 2. Connections to the Generating Philosophy Project Several elements of this chapter bear on the project's active threads. I present these as parallel connections, not ranked by importance. ### The knowing-that / understanding-why gap and the evaluation of LLM output Lipton's opening question — what has to be added to knowledge to yield understanding? — maps onto a question the project's Section 0 raises via Dellsen and Bengson. Both Dellsen's dependency-modelling account and Bengson's six-property account are attempts to characterise understanding as something beyond mere true belief. The paper already draws on these accounts to argue that neither creates friction for AI-generated understanding, since the properties in question attach to the theory (or the model), not to the producer. Lipton's formulation of the gap sharpens this move. If explanation is what bridges the gap between knowing-that and understanding-why, and if explanation is a feature of the *text* — as the project argues for philosophy — then an LLM that produces an explanatory text has produced something that bridges the gap *for the reader*, regardless of whether the LLM itself "understands" anything. The gap is between the reader's knowing-that and their understanding-why; the explanatory text is the bridge. This is consistent with the project's artefact-level evaluation thesis: what matters is whether the text supplies whatever it is that turns knowledge into understanding, and that is assessable from the text. But Lipton also makes an observation that complicates things. He notes that "we do not have a clear conception of understanding apart from whatever it is our explanations provide." If our only grip on understanding comes through explanation, and if we cannot articulate a gap between meeting our standards for explanation and actually explaining, then the question of whether an LLM "really" explains (as opposed to producing text that meets all our standards for explanation) may not be well-formed. I interpret this as convergent with the project's "appearance/reality collapse" argument in Section 3: for competent readers, meeting the standards *is* explaining. Lipton's observation provides independent support for this claim from the philosophy of explanation itself, not just from metaphilosophy. I am speculating here, but this might be worth developing: Lipton's remark that we cannot clearly distinguish understanding from seeming-to-understand could serve as a direct response to Floridi et al.'s "abductive appearance" framing. If the appearance/reality distinction for explanation is itself unclear — if there is no Humean gap between meeting standards and actually explaining — then "mere abductive appearance" may not name a stable category. ### Self-evidencing explanation and philosophy as "textual all the way down" Lipton's discussion of self-evidencing explanations is, I think, directly relevant to the project's claim that philosophy's textual medium stands in a distinctive relationship to its contributions. A self-evidencing explanation is one where the explanans explains the explanandum, and the explanandum provides the evidence for the explanans: "the person passing on snowshoes explains the tracks and the tracks provide the evidence for the passing." The circularity is benign. The project's Section 2 argues that philosophy is textual all the way down — the text is the contribution, not a report of a contribution made elsewhere. This creates a structural analogy to self-evidencing explanation. A philosophical text presents an argument; the argument explains why its conclusion holds; and the only evidence that the argument is any good is the text itself — the quality of its distinctions, the precision of its premises, the cogency of its inferences. There is no laboratory result or physical observation that independently confirms the argument's explanatory force. The text is both the explanation and the evidence for the explanation's adequacy. If self-evidencing explanations are "ubiquitous" and benign, as Lipton claims, then the self-evidencing structure of philosophical texts is not a deficiency. It is the normal condition of explanation in a domain where the text constitutes (rather than reports) the contribution. And if this structure is benign, then an LLM that produces a self-evidencing philosophical text — one that presents an argument and simultaneously provides the textual evidence of its own cogency — is doing something explanatorily legitimate, not merely circular. This connection is, I think, underexploited in the current drafts. The project discusses the textual constitution of philosophy at length, but does not frame it in terms of self-evidencing explanation. Lipton's account would give the "textual all the way down" thesis a precise explanatory-theoretic articulation. ### The unification model, Dellsen, and what LLMs do when they "explain" The unification model has a specific connection to Dellsen's account of understanding, which the project already uses. Dellsen argues that understanding consists in grasping a sufficiently accurate and comprehensive dependency model — a representation of how phenomena stand in dependence relations to one another. This is structurally similar to what the unification model describes: understanding a phenomenon by seeing how it fits together with other phenomena. The difference is that Dellsen separates understanding from explanation (you can understand something by learning what it is *independent* of), while the unification model treats explanation and understanding as tightly linked. Lipton's observation that the unification model faces difficulty because "the notion of unification turns out to be surprisingly difficult to analyze" resonates with the project's use of Williamson on theoretical virtues. Williamson treats unification, elegance, and simplicity as virtues that make one theory preferable to another — but he does not offer an analysis of what unification *is*. The difficulty of analysing unification may actually help the project. If unification is best grasped through practice — through seeing what counts as a unifying explanation in case after case — then it is exactly the kind of norm that could be learned from a large corpus of examples. A competence in recognising and producing unifying explanations might be acquirable through pattern-learning, which is what LLMs do by architecture. The difficulty of *analysing* unification does not entail the difficulty of *recognising* it; indeed, the gap between doing and describing that Lipton emphasises throughout the chapter suggests the opposite. We are good at recognising unifying explanations and bad at saying what makes them unifying — which is the signature of a tacit competence, and tacit competences are precisely what statistical learning over large corpora is suited to acquire. This is speculative on my part. But it connects to the dialectical saturation thesis: if the norms of good explanation (including unification) are tacit but learnable from text, then a model trained on philosophical texts has absorbed not just the *content* of philosophical explanations but the *evaluative standards* by which explanations are judged good or bad. ### Asymmetries of explanation and what it takes to "get the direction right" Lipton's discussion of explanatory asymmetry — the fact that explanation runs in one direction even when deduction runs in both — has implications for evaluating LLM philosophical output. The Doppler law lets us deduce recession from red-shift and red-shift from recession, but only the recession explains the red-shift. Getting the direction of explanation right is part of what makes an explanation good. In philosophy, asymmetry takes a different form. Philosophical explanations are not typically causal (as Lipton will argue the best explanations are in later chapters), but they do have directionality. One explains a phenomenon by appeal to something more basic — a principle, a distinction, a structural feature — not by appeal to something that is itself explained by the phenomenon. Getting this direction right is a mark of philosophical competence: knowing which things are explanatorily prior to which, and not reversing the order. The question for the project is whether LLMs can get this direction right. If explanatory direction in philosophy is encoded in the corpus — if the training data contains thousands of instances of correct explanatory ordering, with implicit signals marking which way the explanation runs — then the model may have absorbed these directional patterns. The asymmetry would be tacitly present in the learned distributions. But if explanatory direction depends on something beyond what is statically present in text — on a grasp of metaphysical priority, say, or on understanding why the explanation runs this way rather than that — then this is a potential locus of failure. Lipton's observation that the D-N model cannot account for asymmetry is relevant here: a model that generates deductively correct text (premises entailing conclusions) might still get the direction of explanation wrong, producing valid arguments that explain the wrong thing. I am not sure how significant this worry is for the project. The current drafts do not discuss explanatory direction explicitly. But it might be worth flagging as a potential text-internal failure criterion: one thing competent readers check for is whether the explanatory direction is right, and this is assessable from the text. ### The why-regress and the scope of LLM explanatory competence Lipton's resolution of the why-regress — that explanations need not themselves be understood — has a bearing on what we should expect of LLM philosophical output. The project argues that LLMs can produce philosophy that satisfies the discipline's evaluative standards. Lipton's point suggests a more modest but potentially useful claim: an LLM-produced explanation can provide understanding to a reader even if the LLM does not itself understand the explanation, just as a drought explains a poor crop even if we do not understand the drought. Understanding is not a substance the explanation has to possess in order to transfer it. This is, I think, a cleaner way of making the point that the project already makes about artefact-level evaluation. The project currently argues that we evaluate texts, not producers, and that provenance is irrelevant. Lipton's formulation adds something: understanding is not a property the producer needs to have in order for the product to provide it. The analogy is precise. Just as the drought need not "understand" itself in order to explain the crop failure, the LLM need not "understand" the philosophical argument in order for its text to explain the phenomenon to a competent reader. ### The gap between doing and describing Lipton repeatedly emphasises that our explanatory practices outrun our ability to describe them. We discriminate between good and bad explanations with great reliability, but we cannot give a principled account of how we do this. None of the five models captures the full range of our explanatory practice. This doing/describing gap connects to the dialectical saturation thesis. If the norms governing explanation are better grasped in practice than in theory — if competence outstrips articulability — then those norms are the kind of thing that can be learned from examples rather than from rules. LLMs learn from examples. The fact that we cannot fully articulate the norms of good explanation does not mean they are unlearnable; it means they are best learned the way LLMs learn: by exposure to a very large number of instances, with implicit evaluative signals (which explanations the philosophical community treats as successful, which it criticises and rejects) serving as training data. This is structurally the same point as the dialectical saturation thesis makes about argumentative norms, extended to explanatory norms specifically. ## 3. Suggested Deployment in the Paper **Section 0 (Introduction):** Lipton's formulation of the knowing-that/understanding-why gap could reinforce the Dellsen/Bengson material. The observation that we lack a conception of understanding independent of our explanatory practices supports the claim that understanding is assessable through the products of explanation (texts, theories) rather than through access to the producer's inner states. **Section 1 (What LLMs Aren't Doing):** The chapter is already drawn on here for the generation/selection distinction and the formulation of IBE. No further deployment seems necessary in Section 1. **Section 2 (Abduction and Philosophy):** Two connections could be developed. First, the self-evidencing explanation concept could articulate what "textual all the way down" means in explanatory-theoretic terms: philosophy produces self-evidencing explanations, and this is benign. Second, the remark about the absence of a Humean gap between meeting standards and actually explaining could directly address the "abductive appearance" worry. If there is no stable distinction between seeming-to-explain and actually explaining, then "abductive appearance without substance" is not a well-defined category. **Section 3 (Learning the Game):** The doing/describing gap supports the claim that philosophical norms are tacit but learnable. The difficulty of analysing unification specifically supports the claim that evaluative standards are practice-patterns rather than articulable rules, and therefore acquirable through statistical learning. These could reinforce the existing Walton/Bengson material. **Section 4 (Demonstration):** The asymmetry discussion could inform the evaluation of worked examples — one thing to check is whether the LLM gets explanatory direction right. This would also serve as an honest concession: getting direction right is non-trivial, and checking for it is part of what competent readers do. ## 4. Divergences Lipton's chapter is focused on explanation in general, not on philosophy specifically, and his examples are drawn predominantly from the natural sciences (red-shift, drought, bridge collapse, snowshoe tracks). His concern is with causal explanation of events and regularities in the physical world. The project, by contrast, is concerned with philosophy, where explanation is typically not causal in Lipton's sense. Philosophical explanations appeal to logical relations, conceptual dependencies, and structural features of theories, not to physical causes. Lipton's preferred model (the causal model, developed in Chapter 3) may therefore be less relevant to the project than the models he rejects — particularly the unification model, which is closer to what philosophical explanation involves. The project should be attentive to this: borrowing Lipton's framework for understanding IBE does not commit the project to his preferred model of explanation. There is also a divergence regarding the scope of the why-regress resolution. Lipton argues that what explains need not be understood. But in philosophy, there is a stronger expectation that explanations be transparent — that the reader can follow every step and see why each move is made. A philosophical explanation that appealed to an opaque principle might be viewed with suspicion in a way that the drought example is not. This does not undermine the basic point (the LLM need not understand its own output for the output to explain), but it suggests that the *text* needs to exhibit a higher degree of self-transparency than explanations in the natural sciences typically do. Finally, Lipton does not discuss the possibility that explanation is a literary or textual practice in the way the project requires. His account is primarily epistemological: explanation is about the relationship between the world and our understanding of it. The project's distinctive move — that philosophical explanation is constituted by text, not merely reported in it — goes beyond anything Lipton considers. This is not a conflict, but it means the project is extending Lipton's framework rather than straightforwardly applying it. ## 5. Passages Worth Re-Reading The passage on self-evidencing explanations (beginning with "Suppose you ask me why there are certain peculiar tracks in the snow...") is worth close re-reading. It establishes that the circularity of self-evidencing explanations is benign and that such explanations are "ubiquitous." This is the passage that most directly supports the "textual all the way down" thesis. The remark on the absence of a Humean gap for explanation deserves attention: "We do not appear to know how to make the contrast between understanding and merely seeming to understand in a way that would make sense of the possibility that most of the things that meet all our standards for explanation might nonetheless not really explain." This passage, if deployed carefully, could do significant work against the "abductive appearance" framing. The discussion of the unification model's strengths and weaknesses (beginning "According to the unification model, we come to understand a phenomenon when we see how it fits together with other phenomena into a unified whole") is worth re-reading for its connection to Dellsen, and for the specific observation that unification is difficult to analyse without reducing it to vocabulary. The implication — that unification may be a practice-level competence rather than a theoretically articulable property — has not been drawn out in the current drafts. The opening paragraph, with its formulation of the knowing-that / understanding-why gap, is quotable and could appear in Section 0 alongside the Dellsen/Bengson material. The passage on the doing/describing gap — "we discriminate between things we understand and things we do not, and between good explanations and bad explanations, but we are strikingly poor at giving any sort of principled account of how we do this" — deserves attention for its direct relevance to the dialectical saturation thesis: the norms are real, effective, and learnable from practice despite being resistant to explicit articulation.