# Lipton Ch. 3: The Causal Model — Relevance to Generating Philosophy with AI ## 1. Summary of the Chapter Chapter 3 of Lipton's *Inference to the Best Explanation* develops the causal model of explanation and, within it, the contrastive structure that becomes Lipton's distinctive contribution. The chapter opens by cataloguing the causal model's advantages over competitors: unlike the deductive-nomological model, it handles asymmetry (causes explain effects, not conversely); unlike the familiarity model, it does not conflate the common with the understood; unlike the reason model, it allows self-evidencing explanations. The short version of the causal model is that "to explain a phenomenon is simply to give information about its causal history" (Lewis 1986, as Lipton frames it). Understanding, on this view, "is not some sort of superknowledge, but simply more knowledge: knowledge of causes." Lipton acknowledges two important concessions. First, the causal model lacks a finished analysis of causation, but he argues this should not stop us from using it: "Applying the moral of the why-regress reflexively, we do not need fully to explain causation in order to use causation to explain other things, in this case, the nature of explanation itself." Second, and more consequential for the Generating Philosophy project, he concedes that there are genuine non-causal explanations: > Mathematicians and philosophers, for example, give explanations, but mathematical explanations are never causal, and philosophical explanations seldom are. A mathematician may explain why Godel's Theorem is true, and a philosopher may explain why there can be no inductive justification of induction, but these are not explanations that cite causes. This concession — that philosophical explanations are seldom causal — is worth dwelling on. Lipton treats this as a limitation of the causal model's scope rather than a refutation. He defends the causal model on the grounds that "causal explanations are explanatory *because* they are causal," and that the overwhelming preponderance of explanations are in fact causal. But the parenthetical acknowledgment that philosophical explanations occupy a different territory is significant for a project asking whether LLMs can do philosophy. The chapter's main constructive work is the development of contrastive explanation. The insight is that we do not explain "Why P?" simpliciter, but rather "Why P rather than Q?" A contrastive phenomenon "consists of a fact and a foil," and the same fact can have different foils, each demanding a different explanation. Lipton introduces the Difference Condition: "To explain why P rather than Q, we must cite a causal difference between P and not-Q, consisting of a cause of P and the absence of a corresponding event in the case of not-Q." He calls this mechanism *causal triangulation*: the choice of foil "selects a location on the causal history leading up to the fact." Three observations structure the analysis. First, fact and foil need not be incompatible — many sensible contrastive questions involve compatible outcomes, as in the syphilis/paresis case. Second, explaining a contrast is sometimes *easier* than explaining the fact alone. Lipton's example is vivid: "Jones's syphilis does not, it seems to me, explain why he got paresis, since the vast majority of people who get syphilis do not get paresis, but it does explain why Jones rather than Smith got paresis, since Smith did not have syphilis." Third, explaining a contrast is sometimes *harder* than explaining the fact alone, because contrastive explanation requires causal information that distinguishes fact from foil: "To explain why Kate rather than Frank won the prize, it is not enough that she wrote a good essay; it must have been better than Frank's." Lipton also gives a sustained analysis of interest-relativity. The causes that explain are selected by our interests, but interests are not arbitrary — they are channelled through the structure of the contrastive question: > What makes one piece of information about the causal history of an event explanatory and another not? The short answer is that the causes that explain depend on our interests. But this does not yield a very informative model of explanation unless we can go some way towards spelling out how explanatory interests determine explanatory causes. The contrastive mechanism provides that spelling-out. Different audiences pose different foils, and different foils select different causes. The doctor is interested in controllable causes, the lawyer in legally liable ones, the accused in exculpatory ones. Lipton treats this as a virtue: "the interest relativity of explanations reduces to the important but unsurprising point that different people are interested in explaining different phenomena." --- ## 2. Connections to the Generating Philosophy Project ### Contrastive Structure and Philosophical Argumentation The contrastive model maps onto philosophical argumentation more naturally than it maps onto much of empirical science — which is ironic, given that Lipton develops it primarily with empirical cases. The structure of a philosophical argument is very often: "Why accept thesis P rather than thesis Q?" The counterexample is perhaps the most characteristic move in analytic philosophy, and it functions precisely as a foil. When a philosopher offers a counterexample to a definition of knowledge, for instance, what they are really asking is: "Why should we accept the JTB account rather than recognize this Gettier case as an exception?" The counterexample supplies a foil that demands a causal (or in philosophy's case, explanatory) difference. I interpret the connection as follows: if philosophical argumentation is pervasively contrastive in structure, and if this structure is manifest in philosophical texts (published arguments, objections, replies, distinctions), then an LLM trained on such texts would encounter fact-foil pairings constantly. The dialectical saturation thesis claims that philosophical corpora are saturated with argumentative patterns including "distinction, counterexample, repair, disambiguation, synthesis." Each of these move types has a contrastive structure. A distinction says: "Why should we treat these as one phenomenon rather than two?" A counterexample says: "Why accept this principle when it fails for this case rather than that one?" A repair says: "Why accept the revised version rather than the original?" The saturation thesis would predict that LLMs have implicitly learned the Difference Condition — not as a philosophical principle, but as a pattern governing how good arguments work. The model would have learned that when a fact-foil pair is introduced, the adequate response must cite something present in the case of the fact and absent in the case of the foil. This connects to Walton's argumentation schemes in the paper's Section 3. Walton's schemes formalize "rules of the game" including critical questions — built-in stress tests. A critical question functions as a forced foil: it requires the arguer to explain why their conclusion holds rather than what the critical question suggests. The Difference Condition maps onto this: answering a critical question means identifying a causal (or rational) difference that distinguishes the case being argued from the case the critical question raises. The session note records that Walton's framework "operationalizes 'give the best objection,'" and Lipton's contrastive analysis shows why objections have the structure they do — they introduce foils. ### Interest-Relativity and Audience-Sensitivity in Philosophical Writing Lipton's account of interest-relativity has a striking parallel to something the paper needs to address: how philosophical writing is sensitive to its audience. This goes beyond a pragmatic observation about rhetoric. Lipton shows that the *content* of an explanation — which cause gets cited — changes with the interests at play. The doctor, the lawyer, and the accused are not just given different amounts of the same explanation; they are given different explanations, because different foils are in play. I speculate (and mark this as speculation, not something the text directly supports) that this maps onto what happens when a philosopher writes for different audiences. A paper aimed at epistemologists and one aimed at philosophers of mind, even if about the same phenomenon, will select different foils and therefore produce different arguments. The fact-foil structure is not add-on rhetoric; it shapes the inferential content. This is relevant to the project's thread on philosophy's textual medium. If interest-relativity is built into the *structure* of what counts as a good philosophical explanation (not just which explanation gets communicated), then the text does not merely report reasoning — it constitutes it. Lipton writes: "different people are interested in explaining different phenomena." In philosophy, different readers (with different background commitments, different extant positions they find plausible) effectively pose different foils against the same thesis. A philosophical text, to succeed, must anticipate the most likely foils its audience will bring. This matters for the LLM question. An LLM trained on philosophical texts encounters not just arguments, but arguments structured for particular audiences — with foils selected for particular theoretical contexts. I interpret this as evidence that the training data embeds audience-sensitivity as a structural feature, not just a stylistic one. Whether an LLM can *reproduce* this audience-sensitivity (as opposed to having it latent in its parameters) is an empirical question, but Lipton's analysis suggests that the sensitivity is textually manifest — it shows up in which foils are addressed, which distinctions are drawn, which counterexamples are considered. This supports the paper's claim in Section 3 that "norms of philosophical practice are textually manifest." ### Philosophical Explanation as Non-Causal — The Divergence Lipton's concession that "philosophical explanations seldom are" causal raises a question about how far the causal model travels for this project. Section 1 of the paper uses Lipton's generation/selection distinction as a framing device. Floridi and Zahavy both argue that LLMs cannot do abduction — cannot generate hypotheses in the way required for genuine scientific reasoning. But the paper pivots by noting that philosophy's textual medium changes the game: in philosophy, "text constitutes reasoning, not reports it." Lipton's own admission that philosophical explanations are non-causal sharpens this pivot. The causal model works for empirical science because science tracks causal histories in the world. Philosophy, on the other hand, typically tracks rational relations — entailment, consistency, conceptual dependence, argumentative support. These are not causal. But the contrastive *structure* that Lipton develops is not inherently causal. The Difference Condition — cite a relevant difference between fact and foil — can be abstracted from its causal context. In philosophical argumentation, the "difference" is not a causal difference but a rational one: a premise, a distinction, a principle, a conceptual resource. Explaining why thesis P is preferable to thesis Q in philosophy means citing a rational difference — a consideration that supports P and has no corresponding support for Q (or a consideration that undermines Q but not P). This is my interpretation, not Lipton's explicit claim. But I think it is well-supported by the structure of the chapter. Lipton himself abstracts the contrastive structure away from causation when he considers whether all why-questions are contrastive. And his discussion of how the Difference Condition handles "global contrasts" (Why P rather than not-P?) pushes toward a formulation that is more about structural difference than specifically causal difference. ### Move 37 and the Generation of Novel Foils The Move 37 / Tail Novelty thread asks how LLMs might produce genuinely novel philosophical moves. Lipton's chapter suggests a specific mechanism by which novelty can emerge within a contrastive framework. Novel philosophical work often consists not in discovering new facts, but in proposing new foils — reframing a question so that the contrast is drawn differently. When Gettier reframes the knowledge question as "Why should we accept that justified true belief is knowledge rather than recognize that this case satisfies JTB but is not knowledge?", the novelty is in the foil, not the fact. The philosophical tradition had been asking "Why P rather than Q1?"; Gettier asks "Why P rather than Q2?" I speculate that LLMs could generate novel philosophical moves precisely by generating novel foils — proposing contrasts that the existing literature has not considered. This would be a form of combinatorial creativity: the LLM has encountered many fact-foil pairings and could, in principle, recombine a fact from one context with a foil from another. Whether this is "genuine" novelty or merely recombination is a question the project is already tracking (the session note records "Novelty through combinatorial creativity" as a thread in Section 3). Lipton's contrastive framework gives this combinatorial process a precise structure: the novel move is the proposal of a new foil that, via the Difference Condition, selects a different location in the rational history of the thesis. ### The Eggplant Example and Ignorance as Explanation One of Lipton's most striking examples deserves attention for what it suggests about LLM limitations: > A particularly striking example of the relative ease with which some contrasts can be explained is the explanation that I chose A rather than B because I did not realize that B was an option. If you ask me why I ordered eggplant rather than sea bass (a 'daily special'), I may give the perfectly good answer that I did not know there were any specials; but this would not be a very good answer to the simple question, 'Why did you order eggplant?' This is a case where *ignorance* — the absence of information — serves as the contrastive explanation. It is relevant to the Floridi/Zahavy arguments about LLMs and abduction. Both Floridi and Zahavy argue that LLMs cannot generate genuine abductive hypotheses because they lack the capacity to recognize what they don't know. Lipton's eggplant case shows that in contrastive explanation, knowing what options were unavailable (or unknown) is itself explanatory. An LLM, arguably, does not have access to its own epistemic limitations in this way — it cannot explain "I generated response P rather than response Q because Q was not in my training distribution" as a contrastive explanation, because it lacks the meta-cognitive awareness that the foil was unavailable. I note this as a divergence — a point where Lipton's analysis reveals a limitation of LLMs that the project should acknowledge, even as it argues for LLMs' philosophical capabilities. --- ## 3. Suggested Deployment in the Paper **Section 1 (Floridi + Zahavy):** The chapter already uses Lipton's generation/selection distinction. The contrastive model could strengthen the argument that Floridi and Zahavy are targeting the wrong kind of explanation. Their arguments assume that LLMs need to track causal histories to do abduction. But Lipton himself concedes that philosophical explanations are seldom causal. The contrastive *structure* — the Difference Condition — is what does the work in philosophy, and that structure is textually manifest (it shows up in how arguments are laid out). This reinforces the pivot to philosophy's textual medium. **Section 2 (Abduction and Philosophy):** The four conceptions of abduction currently disambiguated in Section 2 could benefit from Lipton's contrastive framing. Lipton shows that "Why P rather than Q?" is irreducible to "Why P?" or "Why P and not-Q?" — the contrastive question has its own logic. This parallels the paper's claim that philosophical abduction (generating philosophical hypotheses) is not reducible to scientific abduction (generating causal hypotheses). The irreducibility of contrastive questions supports the irreducibility of philosophical explanation to causal-empirical explanation. **Section 3 (Learning the Game):** This is the most natural home for sustained engagement with the contrastive model. The claim that norms of philosophical practice are textually manifest can be strengthened by showing that the contrastive structure of philosophical arguments is visible in texts: foils are stated, differences are cited, counterexamples supply contrasts. An LLM trained on philosophical texts encounters the Difference Condition in action, repeatedly. This supports the saturation thesis — the training corpus is saturated with contrastive argumentative structures that implicitly encode when an explanation succeeds (when it cites a genuine difference) and when it fails (when the "difference" is present in both fact and foil cases, like Lipton's church-going example where adding irrelevant differences ruins the explanation). **Potential new use — interest-relativity as textual evidence for audience-sensitivity:** If the paper develops the claim that philosophical texts encode audience-sensitivity (as I suggest above), Lipton's account of interest-relativity could serve as the theoretical underpinning. The argument would be: interest-relativity is not a pragmatic overlay on explanation; it is built into the content of explanations via the choice of foils. Philosophical texts manifest this by addressing specific foils — and therefore specific audiences. This is strong support for "norms textually manifest." --- ## 4. Divergences and Tensions The most significant divergence is Lipton's explicit claim that philosophical explanations are seldom causal. The paper uses Lipton's framework, but the paper's argument is that philosophy operates in a different explanatory register from empirical science. Lipton's concession supports this, but it also raises a question: if the contrastive model is primarily a *causal* model, how much of it transfers to the non-causal domain of philosophical explanation? I have suggested above that the contrastive structure abstracts away from causation, but this is my interpretation — Lipton does not make this move himself. The paper should be explicit about which parts of Lipton it takes (contrastive structure, Difference Condition, interest-relativity) and which it leaves behind (the specifically causal content). A second divergence concerns the Difference Condition's requirement of "the absence of a corresponding event." In empirical explanation, this is an absence in a causal history. In philosophical argumentation, the analogue would be something like the absence of a corresponding reason, or the absence of a parallel case. The notion of "corresponding event" is, by Lipton's own admission, somewhat vague — he writes that he has "been unable to give a precise account of the notion of a corresponding event" but argues that this vagueness is one "that we clearly negotiate successfully in practice." Whether LLMs can negotiate this vagueness — identifying what counts as a "corresponding" philosophical consideration in a new context — is an open question. A third tension: Lipton's model is resolutely about *explanation*, and the paper's project is about *argumentation*. These overlap but are not identical. Philosophical argumentation includes explanation (explaining why a thesis is preferable), but it also includes activities — defining, stipulating, motivating, diagnosing — that may not fit neatly into the contrastive model. The paper should be careful not to overextend the contrastive framework beyond its natural domain. --- ## 5. Passages Worth Re-Reading The concession about non-causal explanation (early in the chapter) deserves close re-reading alongside the paper's Section 1 pivot about philosophy's textual medium. The specific passage is where Lipton writes that "mathematical explanations are never causal, and philosophical explanations seldom are." This is a key textual warrant for the paper's claim that philosophy requires a different account of explanation than empirical science. The three observations about contrastive explanation (fact and foil can be compatible; explaining contrast is sometimes easier than explaining the fact alone; explaining contrast is sometimes harder) are worth re-reading with philosophical argumentation examples in mind. Lipton uses empirical examples, but each observation has a philosophical analogue, and the paper could develop these. The discussion of how "not all contrasts make for sensible contrastive questions" is worth re-reading for what it implies about the quality of philosophical arguments. Lipton writes that "the central requirement for a sensible contrastive question is that the fact and the foil have a largely similar history, against which the differences stand out." This has a philosophical analogue: a good counterexample is one where the case closely resembles the case the principle was supposed to cover, so that the difference is maximally informative. Gettier cases are effective precisely because they share "a largely similar history" with genuine knowledge cases. An LLM that has learned what makes a good foil — structural similarity with targeted difference — has learned something deep about philosophical argumentation. The passage on how additional true but irrelevant premises can "ruin" an explanation is worth re-reading for its implications about LLM verbosity and noise. Lipton writes that "if I say that Jones rather than Smith contracted paresis because only Jones had syphilis and only Smith was a regular church-goer, I have not simply said more than I need to, I have given an incorrect explanation, since going to church is not a prophylactic." If LLMs tend to include more information than necessary (a known tendency), this passage identifies the explanatory cost: irrelevant additions do not just dilute — they corrupt. This bears on the quality-control question at the heart of the project. Finally, the passage on deductive closure and how the contrastive model avoids it is worth revisiting. Lipton's paisley tie example — explaining why men wear paisley ties is different from explaining why they wear ties at all, because the foils are different — provides an elegant demonstration of how contrastive structure prevents spurious generalization. This is relevant to the project's concern with whether LLMs can maintain the level of resolution appropriate to a given philosophical question, rather than collapsing into generic or overly broad claims. *La struttura contrastiva dell'argomentazione filosofica si rivela nel testo stesso — il foil non e un ornamento retorico, ma la condizione sotto cui una differenza diventa spiegazione.*