## Philosophy as Self-Grounding Domain: Exploration of the Position Space ### 1. What "Grounding" Means in This Context The symbol-grounding problem, as [[Luciano Floridi|Floridi]] et al. invoke it, is the claim that LLMs "lack grounded semantics connecting words to the physical world or perceptual experiences" (Floridi et al., p. 8, citing Harnad 1990). The worry is that an LLM outputting "fire" after "smoke" has learned a linguistic association, not a causal relation. Floridi puts it precisely: > "Humans infer real fire in the world, while LLMs predict 'fire' within sentences." (Floridi et al., p. 16) The self-grounding kernel turns on what counts as "the world" for a given domain of inquiry. For empirical sciences, the world is external: cells, particles, chemical reactions, perceptual stimuli. An LLM that outputs claims about cellular biology without ever having encountered a cell faces a genuine grounding deficit -- its words bear no causal-perceptual relation to the entities they denote. The question is whether philosophy's "world" has this same character, or whether it is constituted differently. The existing note states the claim as a slogan: "In philosophy, the Map IS the Land. The relations between concepts *are* the objects of study" ([[Philosophy as self-grounding domain]]). The paper draft in Section 3 puts it more carefully: "Unlike biology, grounded in cells, or physics, grounded in particles, philosophy is grounded in the space of reasons itself. The objects of philosophical study are logical and inferential relations, not external entities." I want to stress-test this by developing four distinct positions on what philosophy's "world" consists in, and what each implies for the LLM question. --- ### 2. The Range of Positions #### Position A: Strong Self-Grounding The strong version claims that philosophy's subject matter is entirely constituted by inferential and logical relations -- relations that are fully present in, and exhaustively represented by, philosophical texts. On this view, the "world" of philosophy just is the [[space of reasons]]: the network of entailment relations, consistency constraints, explanatory connections, and dialectical obligations that philosophical texts articulate and enact. There is no residue of philosophical subject matter that exists outside this network. The textual support comes from the observation that philosophy's verification procedures are largely internal. Validity is a property of argument form. Consistency is a property of sets of propositions. Dialectical robustness -- whether a view survives the objections that bear on it -- is assessable by examining the arguments and their relations. None of these require peering beyond the text at some external referent. This position gets additional support from the way [[John Bengson|Bengson]] et al. characterise philosophical method. Their Tri-Level Method involves accommodation of data, explanation, substantiation, and integration -- all of which are assessable by examining the theory's relation to other propositions. The Integration Criterion, for instance, requires that a theory's "claims and commitments integrate with each other and our best picture of the world" (Bengson et al., p. 119). But even "our best picture of the world" is specified propositionally -- it consists in "those uncontroversial claims furnished by common sense and our best physics, biology, chemistry, medicine, logic, and mathematics" (Bengson et al., p. 120). The background against which philosophical theories are assessed is itself a body of propositions, not a set of unmediated encounters with reality. If strong self-grounding is right, LLMs are well-positioned for philosophy specifically. An LLM trained on philosophical corpora has internalised the very network of inferential relations that constitutes philosophy's subject matter. The symbol-grounding problem evaporates, because the "symbols" (words in philosophical texts) do not need to be grounded in an external world -- they are already embedded in the inferential network that is philosophy's world. The map literally is the land. I am speculating here, but the strongest version of this position might claim that philosophy is more like mathematics than like any empirical discipline. [[Timothy Williamson|Williamson]] himself draws this comparison: "mathematics is a precedent for a successful discipline with an 'armchair' methodology that still has a key role for abduction" (Williamson, p. 362). In pure mathematics, the "objects" are abstract structures defined by axioms, and verification is entirely internal (proof from axioms). If philosophy's objects are similarly abstract -- logical and inferential relations rather than concrete entities -- then the grounding problem has no more purchase here than it does in mathematics. **What this implies for the LLM question:** If the strong position is right, philosophy is among the domains where LLMs have the *least* grounding deficit, not the most. The usual complaint -- that LLMs manipulate symbols without understanding what those symbols refer to -- loses its force when the symbols refer to other symbols, or more precisely, when the referents just are the inferential relations among propositions. #### Position B: Moderate Self-Grounding The moderate version holds that philosophy is *substantially more* self-grounding than empirical inquiry, but not entirely so. Some philosophical work genuinely requires only internal relations among propositions. But other philosophical work depends on inputs that cannot be fully captured in text -- moral intuitions, aesthetic responses, phenomenal experience, empirical findings that require understanding their evidential significance. The text in Section 3 gestures at this without fully developing it: "Philosophy is truth-directed, but it often lacks cheap external answer keys. Unlike empirical sciences with experimental verification, or mathematics with formal proof, philosophy relies heavily on public, text-assessable constraints." The word "heavily" is doing important work here -- it concedes that philosophy does not rely *exclusively* on text-assessable constraints. Bengson et al.'s enumeration of data-collection procedures supports the moderate position. They list: "Perception of items in one's environment; Introspection of one's own phenomenal states; Intuitions about thought experiments; Linguistic judgments about the grammaticality, meaning, or truth-conditions of sample sentences; Common sense verdicts about what exists, is known, etc.; Imagining future possibilities" (Bengson et al., p. 63). Several of these -- perception, introspection, imagination -- are not purely textual operations. They involve some kind of experiential contact with a domain that is not fully captured by its propositional representation. The moderate position acknowledges that the *theorising* stage of philosophy (where theories are constructed and evaluated against criteria) is largely self-grounding, while the *data collection* stage often requires world-contact. The Bengson model makes this architecturally clear: data collection (stage one) feeds into theorising (stage two), and the two stages are governed by different norms. An LLM might be well-equipped for stage two -- constructing theories that accommodate, explain, substantiate, and integrate -- while being systematically limited at stage one, where the inputs to theorising are generated. **What this implies for the LLM question:** The moderate position gives the self-grounding kernel real but bounded force. It says: LLMs can do the theory-construction and theory-evaluation work of philosophy (which is the bulk of what published papers contain) because that work is self-grounding. But they may lack access to some of the raw inputs -- the intuitive responses, the phenomenal data, the perceptual encounters -- that feed into philosophical theorising. This is a genuine limitation, but it may be less severe than it appears, because (a) philosophical data, once collected, are communicated propositionally and thus are present in texts, and (b) the vast majority of philosophical work involves responding to *already-collected* data, not generating new data from scratch. #### Position C: Domain-Variable Self-Grounding The domain-variable position holds that self-grounding is not a property of philosophy as such, but varies across philosophical subdisciplines. Some areas of philosophy are highly self-grounding; others are not. The clearest cases of self-grounding are formal logic and parts of metaphysics that deal with abstract structure. Williamson explicitly distinguishes these: "In philosophical logic, for example, an armchair methodology closer to that of mathematics is appropriate" (Williamson, p. 362). Evaluating whether a modal schema is valid involves checking instances and counterexamples -- a purely propositional exercise. Metaphysics of composition, identity, and modality similarly deals with relations among concepts that are fully articulable in text. At the other end, philosophy of perception is directly dependent on empirical findings. Williamson acknowledges this: "contemporary philosophy of perception is deeply informed by recent experimental results from the psychology of perception. Sometimes the experiments most relevant to a question in the philosophy of perception will not yet have been performed, or even conceived" (Williamson, p. 357). Here, the philosophical "world" includes empirical facts about how human perceptual systems work -- facts that are not internal to the space of reasons but come from laboratory results. Between these poles sit large swathes of philosophy. Ethics depends on moral intuitions that may require phenomenal experience or social embeddedness. Philosophy of mind engages with empirical cognitive science. Aesthetics involves responses to artworks that may not be fully communicable propositionally. Political philosophy, as Williamson notes, must "consider the historical outcomes of attempts to apply various political theories" (Williamson, p. 362) -- genuinely empirical input. **What this implies for the LLM question:** The domain-variable position gives the self-grounding kernel limited scope. It holds for logic, much of metaphysics, parts of epistemology, and meta-level philosophical methodology -- but not for philosophy of perception, philosophy of mind, applied ethics, or any area where empirical input or experiential contact plays a substantial role. The question then becomes: how much *interesting* philosophy falls in the self-grounding category? If the answer is "a lot but not all," the kernel is useful but not decisive. #### Position D: Against Self-Grounding The strongest case against the self-grounding thesis draws on Williamson's insistence that philosophy's evidence base is unrestricted. The passage is emphatic: > "From what evidence base should we start when applying abduction to the construction and selection of philosophical theories? As always, the answer is in principle: our total evidence. That is arguably no less than the total sum of human knowledge. It includes whatever knowledge the natural and social sciences, philosophy, and common sense have already gained. None of our knowledge is irrelevant in principle to philosophy, for any philosophical theory inconsistent with any of it is false (since what is known is true). In particular, there is no restriction to knowledge gained in some special 'conceptual' or 'a priori' or 'intuitional' or 'armchair' way." (Williamson, p. 356-357) This is a direct challenge to self-grounding. If philosophy's evidence base is "the total sum of human knowledge," then philosophy is not a closed system of inferential relations. It is open to the entirety of human experience, including empirical findings, common-sense knowledge, perceptual encounters, and practical wisdom. The "world" of philosophy, on Williamson's view, is not the space of reasons but the whole world -- everything we know. Williamson reinforces this: "There is no methodological firewall between philosophy and real life experiments" (Williamson, p. 357). Philosophy can and sometimes must generate new empirical data. Thought experiments are not always sufficient. The discipline is porous to external input in a way that undermines the "map is the land" metaphor. Bengson et al. make a compatible point by insisting that philosophical data are "what inquirers have good epistemic reason to take to be genuine features of reality itself" (Bengson et al., p. 45, paraphrasing their epistemic reasons theory). The data are *about reality*, not about the space of reasons as such. In aesthetics, the datum is that "certain works of art are masterpieces" (Bengson et al., p. 36) -- a claim about actual artworks, not about inferential relations. In philosophy of action, the datum is that "some attempts are unsuccessful" -- a claim about events in the world. The data that philosophy must accommodate are not purely internal to a propositional network. The anti-self-grounding position can also point to the role of thought experiments. Thought experiments are often treated as philosophy's analogue of empirical data. They invite the reader to imagine a scenario and consult their intuitive response. The response -- "that seems like knowledge," "that doesn't seem fair," "that looks like personal identity" -- is a kind of experiential datum. It is grounded in something more than inferential relations: it draws on the reader's sense of how the world works, on their moral sensibility, on their perceptual imagination. Williamson discusses thought experiments at length and treats them as genuine evidence-generators for philosophy. If thought experiments require experiential engagement that goes beyond propositional manipulation, then philosophy is not fully self-grounding. **What this implies for the LLM question:** If the anti-self-grounding position is right, the self-grounding kernel does not help. Philosophy is as dependent on world-contact as any other discipline, and the symbol-grounding problem applies to it with (nearly) full force. An LLM manipulating philosophical texts without experiential grounding is in the same position as an LLM manipulating biological texts without ever having encountered a cell -- it can reproduce patterns, but it lacks the anchoring in reality that underwrites genuine understanding. --- ### 3. What Each Position Implies for Floridi's Specific Complaints Floridi identifies three deficits in LLMs: (1) lack of truth-directedness, (2) lack of verification capacity, and (3) lack of grounded semantics. How much does self-grounding neutralise each? **Grounded semantics** is the complaint most directly addressed by self-grounding. If philosophy's "world" is constituted by inferential relations, then an LLM that has internalised those relations from training on philosophical texts has something analogous to grounded semantics for this specific domain. The symbols in the LLM's weights are "grounded" in the inferential network that constitutes philosophy's subject matter. Positions A and B both support this conclusion. Position D denies it. **Truth-directedness** is less clearly neutralised. Even in a self-grounding domain, there is a difference between producing propositions that cohere with the inferential network and producing propositions that are *true*. Floridi's point is that LLMs "lack an epistemic compass" (Floridi et al., p. 20) -- they cannot distinguish true from merely plausible. In a self-grounding domain, truth might be more closely approximated by coherence than in empirical domains (where coherence can diverge wildly from reality). But coherence and truth are not the same thing, even in logic. This is where the self-grounding claim faces pressure from within: even if philosophy's world is the space of reasons, there are truths about that space (which inferences are valid, which theories cohere, which explanations succeed) that an LLM might get wrong while producing plausible-sounding text. I interpret this as a genuine residual worry that self-grounding does not fully resolve. **Verification** is partially neutralised. Floridi argues that LLMs "perform prior predictive sampling but lack an external feedback loop for posterior evaluation" (Floridi et al., p. 7). In philosophy, as the paper draft argues, verification is largely internal: "validity, consistency, dialectical robustness." An LLM can, in principle, check whether an argument is valid by tracing its logical structure -- this is an internal operation on propositions, not an external check against reality. But whether this amounts to genuine verification or merely pattern-reproduction-that-looks-like-verification is precisely what is at issue. The self-grounding claim does not settle this debate; it relocates it. Instead of asking "can the LLM verify claims against the world?", we ask "can the LLM verify claims against the inferential network?" -- and the answer depends on how much of the inferential network the model has genuinely internalised versus merely memorised surface patterns of. --- ### 4. Whether Self-Grounding Would Be a Good or Bad Thing for Philosophy There is a non-trivial question about what it *means* for philosophy if the self-grounding thesis is correct. On the positive reading, self-grounding means philosophy has an unusually rigorous internal verification structure. Unlike disciplines that depend on messy empirical input, philosophy can check its own work by examining the logical and inferential relations among its claims. This is a strength: it means philosophical evaluation can be thorough, systematic, and fully articulable. Bengson et al.'s Tri-Level Method, with its ordered criteria from accommodation through integration, embodies this -- it is a set of publicly specifiable norms for internal assessment. The fact that these norms are (in Bengson et al.'s phrase) "familiar from the way many philosophers go about their business" (Bengson et al., p. 107-108) suggests that philosophy already operates as a self-checking system. On the negative reading, self-grounding means philosophy is a closed system -- a "coherentist bubble," as the session note puts it ([[Sessions/Generating Philosophy]], current question 4: "Self-grounding vs coherentist bubble?"). If philosophy's "world" is just the space of reasons, and the space of reasons is constituted by what philosophers have written, then philosophy risks becoming circular: theories are assessed against data that are themselves theoretical commitments, and the whole system floats free of reality. This is the insular reading. It suggests that self-grounding is not a mark of rigour but of disconnection from the world. I speculate that the truth lies in recognising that self-grounding is a matter of degree, not kind. Even the most abstract metaphysics must, on Williamson's account, cohere with "the total sum of human knowledge." The inferential network is not hermetically sealed; it has input nodes where empirical knowledge, common-sense verdicts, and experiential data enter. What self-grounding means is that the *processing* of those inputs -- the construction and evaluation of theories -- is internal to the network. The inputs come from outside; the theorising is internal. This is the moderate position (B), and I interpret it as the most defensible. --- ### 5. Edge Cases and Hard Cases **Moral intuitions.** When a philosopher reports that it seems wrong to push the fat man off the bridge to save five, is this an internal-to-the-space-of-reasons datum or an experiential datum? Bengson et al. list "intuitions about thought experiments" as a data-collection procedure (p. 63), and it seems clear that the intuitive response draws on something beyond propositional manipulation -- a sense of the moral weight of the scenario. An LLM can reproduce the standard responses to trolley problems (because those responses are well-documented in texts), but whether it can generate a genuinely novel moral intuition -- one that has not appeared in the training data -- is unclear. This is a limitation that self-grounding does not resolve. **Aesthetic judgments in philosophy of art.** Bengson et al. note that "certain works of art are masterpieces" is a datum in aesthetics (p. 36). Assessing whether a work is a masterpiece plausibly requires perceptual engagement with the work. An LLM cannot look at a painting or listen to a symphony. For philosophy of art that depends on fine-grained aesthetic discriminations, self-grounding is weakest. **Philosophy of mathematics.** I speculate that this is a case where self-grounding is very strong. Mathematical objects are abstract, mathematical verification is internal (proof), and mathematical intuition -- while real -- is largely communicable through formal notation. An LLM trained on mathematical texts is plausibly in contact with mathematics' "world" in a way that an LLM trained on biological texts is not in contact with biology's. **Thought experiments requiring imagination.** Williamson treats thought experiments as evidence-generators. But some thought experiments require imagining scenarios that are novel -- that have no close analogue in existing texts. The "Mary's Room" thought experiment in philosophy of mind, for instance, asks us to imagine what it would be like for a colour-deprived scientist to see red for the first time. The philosophical force of the example depends on our imaginative engagement with the scenario. An LLM can reproduce the example and its standard interpretations, but whether it can generate a *new* thought experiment with comparable force is an open question. This is where the boundary between "self-grounding" and "requiring experiential input" becomes genuinely unclear. --- ### 6. Assessment The self-grounding kernel is real but its force is domain-variable and degree-sensitive. The text sources support the following graduated assessment: The kernel is **strongest** for: formal logic, abstract metaphysics, meta-level philosophical methodology, parts of epistemology that deal with the structure of justification, and any philosophical work that is primarily responsive to the existing dialectical landscape (which is textual and therefore available to LLMs). The kernel is **weakest** for: philosophy of perception, philosophy of mind engaging with cognitive science, applied ethics depending on novel moral intuitions, aesthetics requiring perceptual engagement with artworks, and any philosophical work that depends on genuinely new empirical input. The kernel **does not fully neutralise** Floridi's complaints. It substantially weakens the grounded-semantics objection for the self-grounding subdomain. It partially weakens the verification objection (because philosophy's verification is largely internal). It leaves the truth-directedness objection largely intact (because coherence within a propositional network is not the same as truth, even in philosophy). The most defensible version of the thesis, I interpret, is the moderate one (Position B): philosophy is substantially more self-grounding than empirical disciplines, the theorising stage is highly self-grounding, and this gives LLMs an underappreciated affordance in this domain specifically -- but philosophy is not entirely self-grounding, and the data-collection stage involves forms of world-contact that LLMs lack. The paper's existing formulation ("the map IS the land") is rhetorically effective but needs qualification. A more precise formulation might be: "In philosophy, the terrain that theories must navigate is substantially constituted by the map -- the existing network of arguments, positions, and inferential relations -- in a way that has no parallel in empirical disciplines." The worry about a "coherentist bubble" is real but addressable. Williamson's insistence on an unrestricted evidence base functions as the corrective: philosophy is self-grounding in its *method* (internal verification, propositional assessment) but not in its *evidence base* (which includes the total sum of human knowledge). The self-grounding thesis applies to the processing, not to the input. This is a subtle but important distinction that the paper should make explicit. *La questione se la filosofia possa fondarsi interamente su se stessa -- se la mappa dei ragionamenti coincida davvero col territorio -- resta aperta, e forse la sua apertura e proprio cio che distingue il sapere filosofico da quello puramente formale.*