# The Appearance/Reality Collapse: Stress-Testing the Counterargument ## 1. The Weak/Strong Appearance Distinction The collapse argument rests on a distinction between two senses of "looks like good philosophy." The weak sense: the text has genre markers -- signposting, objections-and-replies sections, theoretical virtue talk, confident conclusions. The strong sense: the text actually satisfies the constraints that competent philosophical readers track -- hinge points located, commitments explicit, costs paid, no equivocations smuggled in. The distinction is clean at the poles. A ChatGPT output that begins "There are several perspectives on this issue" and proceeds through a numbered list of views without engaging any of them deeply satisfies the weak sense and fails the strong sense. A passage that identifies a precise tension between two commitments, shows exactly where a repair is needed, and pays the cost of the repair (acknowledging what is given up) satisfies the strong sense. Competent readers can tell these apart instantly. The question is whether the distinction is binary or graded. I want to develop three positions. **Position A: The distinction is binary, and the collapse holds.** On this view, constraint satisfaction in philosophy is like validity in logic: either the inference is valid or it is not. A philosophical text either locates the genuine hinge point of a debate or it does not. It either makes its commitments explicit or it equivocates. There is no stable halfway point. [[Philosophical Methodology Philosophical Methodology –From Data to Theory by John Bengson, Terence Cuneo, Russ Shafer-Landau|Bengson, Cuneo, and Shafer-Landau]]'s Tri-Level Method supports this reading: the criteria -- accommodation, explanation, substantiation, integration -- function as demands that are either met or unmet with respect to particular data. As they put it: > "Methods themselves comprise a set of criteria that serve a dual role: they provide instructions for the construction of a theory, given the data, while also serving as standards by reference to which the merits of theories are evaluated." (Bengson et al., p. 77) The dual-role framing implies that the same criteria that guide construction also furnish the evaluative standard. If you can check whether a theory accommodates the data, and check whether claims are substantiated, then the evaluation is public and determinate. Apparent satisfaction just *is* satisfaction, because the criteria are publicly codifiable. **Position B: There is a genuine intermediate zone, but it is narrow.** On this view, the distinction is graded. Between genre-marker-only outputs and genuine constraint satisfaction, there is a region of text that does *some* real philosophical work -- makes *some* genuine distinctions, engages *some* objections substantively -- but falters at a level that requires close reading to detect. The question is how wide this zone is. If it is narrow, the collapse still holds approximately: the bimodal experience (good philosophy or sludge, rarely the middle) reflects the genuine distribution, even if a few cases fall between. This position says the middle category exists but is a thin strip, not a broad plain. **Position C: The middle zone is wide, and the collapse fails.** This is the position the sceptic needs. It holds that there is a large class of texts that satisfy enough constraints to fool competent readers on first pass but contain subtle failures -- equivocations that require three readings to detect, question-begging premises disguised by terminological sophistication, ad hoc repairs that pattern-match as legitimate but actually shift the explanatory burden rather than discharging it. If this class is large, then the "bimodal experience" people report is an artefact of insufficient scrutiny, not of the actual quality distribution. My assessment: Position B is the most plausible, but this needs testing. Let me actually try to produce a middle-category example and see what happens. ## 2. Attempting to Produce the Feared Middle Category The instruction is to produce a philosophical passage (300-500 words) that satisfies surface constraints but fails deeply. I will attempt this, then assess the attempt. --- *Begin middle-category attempt:* The standard objection to phenomenal conservatism -- that it collapses the distinction between epistemic justification and mere psychological compulsion -- can be defused by attending to a structural feature of seemings that critics have overlooked. Seemings, as Huemer characterises them, carry their own justificatory force: if it seems to S that p, then S thereby has prima facie justification for believing that p. Critics charge that this makes justification too cheap, since pathological states (wishful thinking, phobic projection) also involve its seeming to the subject that something is the case. But this objection conflates two distinct properties of mental states: phenomenal presentiveness and doxastic etiology. A seeming's justificatory force derives not from its causal history but from its presentational character -- the fact that it presents its content as true in a way that is introspectively available to the subject. Pathological states, by contrast, lack this presentational character even when they produce strong conviction. The phobic who "sees" danger everywhere is not presented with danger-as-true in the relevant phenomenal sense; rather, the affective charge of the state is misread as evidential weight. The distinction between presentational seemings and affectively charged convictions is not ad hoc; it tracks a genuine phenomenological difference that subjects can, upon reflection, recognise. This dissolves the cheap-justification worry. If we restrict the scope of phenomenal conservatism to states with genuine presentational character, then pathological states are excluded not by stipulation but by phenomenological description. The justificatory force of seemings is grounded in a specific phenomenal property -- presentiveness -- that pathological states lack. Moreover, this restriction strengthens phenomenal conservatism against the regress problem. If justification ultimately rests on seemings with presentational character, then we have a non-inferential stopping point that is neither arbitrary nor merely causal. The seeming's presentiveness is itself the ground of justification, requiring no further justificatory support. This avoids the foundationalist's usual embarrassment of having to explain why certain states are self-justifying: presentational seemings justify precisely because of what they are, phenomenologically speaking. One might object that the distinction between presentiveness and affective charge is itself theory-laden, and that invoking it to defend phenomenal conservatism is circular. But this confuses the order of justification with the order of explanation. The phenomenological distinction is available pre-theoretically; phenomenal conservatism merely systematises what reflective subjects already recognise. *End middle-category attempt.* --- **Assessment of the attempt.** I speculate that this passage does inhabit the middle category, but I want to be precise about where the failures are. There are at least three: (1) **The distinction between "presentiveness" and "affective charge" is doing unearned work.** The passage asserts that pathological states lack "presentational character" as though this is phenomenologically obvious, but it is not. Many phobic subjects report that danger genuinely *seems* present to them -- that is precisely what makes phobias distressing. The passage helps itself to a phenomenological distinction that would need to be independently motivated, and the reader who does not press on this will accept the move. (2) **The anti-circularity reply is question-begging.** The passage's response to the circularity objection -- that the phenomenological distinction is "available pre-theoretically" -- is itself an application of phenomenal conservatism (it seems to us that the distinction is real, therefore it is). This is the very principle under defence. A careful reader would spot this, but it is disguised by the confident assertion that the distinction is "pre-theoretical." (3) **The scope restriction is ad hoc, despite the denial.** The passage claims the restriction to "presentational seemings" is not ad hoc because it tracks a phenomenological difference. But the phenomenological difference is precisely the one that needs to be demonstrated, not assumed. The passage patterns like a legitimate repair move -- it identifies a distinction and uses it to narrow the principle's scope -- but the distinction itself is the contested territory. **What does the attempt reveal?** Two things. First, producing a middle-category example is *possible* but requires deliberate effort. I had to engineer the failures, which means I had to understand the constraint structure well enough to know where violations could be hidden. This is itself evidence for a version of the collapse claim: producing genuine philosophical structure and producing disguised failures draw on overlapping competences. Second, the failures I produced are *detectable* -- not by casual reading, but by a reader who asks the right questions (Is the phenomenological distinction independently motivated? Does the anti-circularity reply presuppose the principle it defends?). The question is whether competent readers reliably *do* ask these questions. I want to be honest about what this does and does not show. It shows the middle category is not empty. A text can pattern-match as doing genuine philosophical work while containing buried equivocations and circularities. But the passage I produced also feels *fragile* -- the failures are the kind of thing that a seminar would surface within fifteen minutes. Whether this generalises depends on the claim about competent readers, which I turn to next. ## 3. The Reliability of Competent Readers The collapse argument depends on expert readers being good at distinguishing genuine constraint satisfaction from mere appearance. How reliable are they? **Evidence for reliability.** The peer review system, for all its flaws, does function as a filter. Competent reviewers regularly identify equivocations, question-begging, and ad hoc repairs. The "revise and resubmit" process is largely a process of demanding that these failures be addressed. The fact that reviewers often converge on the same problems suggests that constraint satisfaction is publicly trackable. **Evidence against reliability.** [[Widening the Picture by Williamson (from philosophy of philosophy)|Williamson]]'s discussion of overfitting in philosophy provides important counter-evidence. The Gettier literature is his case study. After Gettier's 1963 counterexamples, the analytic epistemology community engaged in decades of increasingly complex repair attempts: > "Just as in the quantitative case, tolerance for highly complicated, messy, gerrymandered analyses yielded a succession of proposals that fitted the current data but succumbed to new ones. Other programs for reductive analysis in philosophy, for instance of causation or meaning, have had similar track records. Strikingly, the philosophical community showed very little aversion to the multiplication of complication." (Williamson, "Widening the Picture," pp. 368-369) This is a case where the philosophical community -- composed of competent readers by any standard -- failed to detect that something was going wrong for *decades*. The over-fitted analyses appeared to satisfy the constraints (accommodating all known counterexamples), but they were what Williamson calls "predictively inaccurate" -- they succumbed to new counterexamples because they were fitting noise rather than signal. Williamson's verdict: > "A firmer preference for simplicity and elegance would have warned the community that something was going wrong. Indications of over-fitting remain quite widespread in analytic philosophy." (Williamson, p. 369) This suggests that competent readers *can* be fooled -- not by mere genre markers, but by something more like my middle-category example scaled up. The over-fitted Gettier analyses had genuine philosophical structure; they made real distinctions, engaged real objections, and proposed real repairs. The problem was not that they lacked constraint satisfaction in the weak sense but that the constraints they satisfied were the wrong ones. They accommodated the data too closely, in Williamson's framework, at the cost of parsimony and predictive accuracy. Williamson also notes that this failure is not random but structural. The community's collective judgement mechanisms -- peer review, citation, conference acceptance -- are themselves subject to fashion effects: > "Academic fashions arise because people trained in a discipline have some respect for the judgment of others trained in the discipline as to what is good or fruitful work, worth imitating or following up. When things go well, that mechanism enables the community to concentrate its energies quickly where progress is being and will be made... The word 'fashion' is most appropriate when the level of deference to majority opinion becomes too high." (Williamson, p. 339) This cuts in two directions for the collapse argument. On one hand, it shows that competent readers are not infallible detectors. On the other, Williamson's own point is that the community *eventually* recognises its errors -- the over-fitting was identified, the Gettier industry is now widely seen as a cautionary tale. The question is the timescale. If the community corrects within a few years, the collapse holds in practice (for purposes of paper evaluation at least). If correction takes decades, the middle category may be larger than the bimodal experience suggests. **A complication from Floridi.** [[Luciano Floridi|Floridi]]'s "phenomenology of plausibility" account suggests a mechanism by which competent readers could be fooled: > "When users interact with an LLM-based AI, such as a chatbot or assistant, they often perceive the AI's responses as if they were created by an intelligent mind reasoning through the question. The AI's answer 'makes sense': it addresses the question with relevant points, sometimes even providing justification or analogies. This phenomenology of plausibility can be pretty compelling." (Floridi et al., p. 10) But Floridi's account of *why* users are fooled invokes factors that are largely irrelevant for competent philosophical readers evaluating a text: interface design, sycophancy, and the assumption of an intelligent mind behind the responses. A competent reader evaluating a philosophical text is not (or should not be) asking "Is there an intelligent mind behind this?" but "Does the text satisfy the relevant constraints?" The phenomenology of plausibility, as Floridi describes it, is the phenomenology of naive users interacting with chatbots -- not the phenomenology of referees reading papers. I interpret this as supporting a qualified version of the collapse: competent readers are reliable detectors of constraint satisfaction in the evaluatively relevant sense, but they are not infallible, and certain systematic failures (overfitting, fashion effects) can persist for long periods. ## 4. Does the Collapse Vary by Sub-Domain? The strongest version of the collapse should hold for areas of philosophy where evaluation criteria are most publicly checkable. Consider a rough ordering: **Formal logic and formal semantics:** Validity is mechanically checkable. Countermodels are constructive. The appearance/reality gap is nearly zero. If a proof appears valid and the steps check out, it *is* valid. This is the easiest case for the collapse. **Analytic metaphysics and epistemology (argument-heavy):** Evaluation involves checking validity, tracking equivocations, assessing whether distinctions are well-motivated, and judging explanatory fit. The criteria are publicly checkable but require interpretive judgement. The collapse holds approximately, but the Williamson overfitting cases show that extended evaluation can go wrong. **Ethics and political philosophy:** Normative claims add a dimension. A reader must evaluate not only argumentative structure but also whether the normative premises are plausible, whether the theory captures the phenomena (moral experience, considered judgements), and whether costs are being paid honestly. The criteria are still publicly articulable, but they involve more evaluative judgement, and disagreement about the data (considered moral judgements) is more pervasive. **History of philosophy and continental philosophy:** Here the collapse is weakest. Evaluation often turns on interpretive adequacy to a textual corpus, sensitivity to historical context, and judgment about which readings are "fruitful" or "illuminating." These are real intellectual virtues, but they are harder to check publicly, and plausible-sounding but shallow readings can persist. The bimodal experience may not hold here. **Discursive metaphysics (the hard case).** Consider something like the free will debate or the debate over composition. Evaluation requires tracking long chains of dialectical engagement, assessing whether repairs are ad hoc, and judging whether theoretical costs are being honestly acknowledged. A text could satisfy many local constraints (each argument is valid, each distinction is well-drawn) while failing at a global level -- the overall position may be unmotivated, or it may purchase local coherence through terminological opacity. Bengson et al.'s integration criterion is especially relevant here: a theory must cohere "with our best picture of the world" (their second-level criterion), and judgements about whether this condition is met are inherently more contestable than judgements about accommodation or explanatory fit. I speculate that the collapse varies along a gradient from formal to discursive philosophy, and that the middle category grows wider as we move toward the discursive end. The bimodal experience may be a feature of philosophy's more constrained sub-domains. ## 5. The Strongest Version of Floridi's Counter The strongest version of the sceptic's reply to the collapse goes like this: the bimodal experience is an artefact of scrutiny limits, not of the actual quality distribution. People report "either good philosophy or sludge" because they do not scrutinise enough to detect the middle category. The middle category exists, is large, and is precisely where LLM outputs cluster. This version draws support from several observations: (a) **Peer review catches only some failures.** Retraction rates, while low, are nonzero. More importantly, many papers that are never retracted contain acknowledged errors, and the "replication crisis" in empirical fields suggests that even expert evaluation is unreliable for certain kinds of subtle failure. Philosophy has no replication equivalent, which means subtle failures may persist undetected. (b) **LLMs might be specifically good at producing middle-category text.** The stochastic core, as Floridi describes it, is optimised for plausibility -- for producing text that aligns with human expectations. The very training process that enables LLMs to mimic philosophical structure also equips them to produce *systematically plausible-looking but subtly defective* text. The defects might be precisely the kind that evade peer review: not gross equivocations but slight meaning-shifts across paragraphs, not obvious question-begging but premises that seem independent but actually encode the conclusion in slightly different vocabulary. Floridi himself gestures at this: > "An explanation can be coherent and convincing (even optimal by IBE criteria) and yet still false. LLMs lack an epistemic compass to navigate that distinction." (Floridi et al., p. 20) The implicit worry is that coherence and convincingness (meeting IBE criteria) are necessary but not sufficient for philosophical quality. Something more is needed -- truth-directedness, genuine understanding -- and this something is precisely what LLMs lack. (c) **The "competent reader" standard may be self-certifying.** If competent readers are defined as those who can detect the middle category, and the collapse is defined by the absence of undetected middle-category text, then the claim is circular: the middle category is empty for readers who can detect it. **My assessment of this counter.** Observation (a) is real but does not distinguish LLM outputs from human outputs; human philosophers also produce work that contains errors peer review misses. Observation (b) is the strongest point, and I am genuinely uncertain about it. The possibility of systematic, subtle, plausibility-optimised defects is the most serious challenge to the collapse. However, the training data that Floridi himself concedes contains philosophical structure is not just *any* structure -- it contains the structure of arguments that have been tested, challenged, and refined through peer review. The distributional regularities the model learns are not the regularities of first-draft philosophy but of published, scrutinised philosophy. This makes it harder (though not impossible) for the model to systematically produce plausible-but-defective output in the specific way the sceptic fears. Observation (c) can be deflected: "competent reader" can be defined independently of the collapse claim, e.g., by reference to track record in identifying known philosophical errors. ## 6. What "Competent Reader" Means The collapse depends on a threshold of expertise. Let me distinguish three levels: **Graduate student.** Can detect gross failures: invalid arguments, obvious equivocations, missing premises. May miss subtler failures: whether a distinction is independently motivated, whether a repair is genuinely explanatory or merely accommodating. At this level, the bimodal experience might hold -- the student sees either "good" or "bad" -- but the threshold for "good" is set too low, and some middle-category text gets classified as good. **Specialist in the sub-field.** Knows the dialectical landscape: which moves have been tried, which objections are standard, which repairs have already been shown to fail. Can detect when a text reinvents a known wheel or ignores a known problem. At this level, the collapse holds more robustly, because the specialist's background knowledge provides additional constraint-checking capacity. The middle category shrinks. **Multi-field expert (rare).** Can evaluate not just the argument's internal structure but its relations to adjacent debates and its broader theoretical implications. At this level, the integration criterion (Bengson et al.'s second-level requirement) is most reliably applied. The middle category may be nearly empty. The collapse, I speculate, is relativised to expertise. For specialists and above, it holds approximately. For graduate students, it does not -- which is consistent with the observation that graduate students are more impressed by LLM philosophical output than senior researchers are. This would mean the collapse is a claim about *expert* evaluation, not about philosophical text *simpliciter*. ## 7. Historical Parallels Several cases from the history of philosophy illustrate what happens when arguments accepted as good by competent readers are later found to contain subtle errors. **The Gettier case (already discussed).** The post-Gettier industry is Williamson's paradigm of overfitting. The lesson: competent readers can collectively pursue a research programme that satisfies local constraints while failing globally, and the failure can persist for decades. **Anselm's ontological argument.** For centuries, competent readers disagreed about whether the argument was valid. Kant's objection (existence is not a predicate) was widely accepted but itself contested. The modern consensus, insofar as there is one, treats the argument as committing a subtle error (treating existence as a first-order property), but the subtlety of the error meant it persisted through centuries of scrutiny. **The private language argument.** Wittgenstein's argument against the possibility of a private language was accepted by many competent readers for decades. Later work (by Kripke, among others) showed that the argument's force depends on controversial premises about rule-following that Wittgenstein himself may not have endorsed. The appearance of a knock-down argument turned out to rest on interpretive assumptions. **Quine's Two Dogmas.** For a generation, many analytic philosophers accepted Quine's attack on the analytic/synthetic distinction as decisive. Later work (by Grice and Strawson, and more recently by Paul Boghossian and others) showed that Quine's arguments, while powerful, depended on a standard of explicability that could be challenged. The "devastating critique" turned out to be a strong but contestable argument. These cases suggest that the middle category does exist in human-produced philosophy, and that it can persist for extended periods. But they also show something else: in each case, the error was *eventually* detected, and the detection involved the same constraint-checking capacities that the collapse argument invokes. The community is a self-correcting system -- imperfect and slow, but functional. The implication for the LLM debate is double-edged. If human-produced philosophy has a non-trivial middle category that persists for years or decades, then the collapse claim cannot be that the middle category is empty -- only that it is smaller than sceptics fear, and that it shrinks under scrutiny. This is a weaker but more defensible version of the collapse. It says: for competent readers engaging seriously with a text, the probability of a text looking good while being deeply defective is low (not zero). And this probability is no higher for LLM outputs than for human outputs, provided the LLM output is subjected to the same scrutiny. ## Summary Assessment The appearance/reality collapse, in its strongest form (the middle category is *empty*), is false. The historical record and the middle-category attempt both show that text can satisfy many philosophical constraints while containing subtle failures. But the weaker form -- for competent readers, the middle category is *small*, and text that genuinely satisfies the publicly checkable standards constitutes good philosophy regardless of its provenance -- remains defensible. The strongest version of Floridi's counter (observation (b) above -- that LLMs might be *specifically* good at producing middle-category text) is the challenge that this weaker form must answer, and I am not confident it has been answered fully. The most honest assessment: the collapse holds approximately, relative to a specialist-or-above level of expertise, in the more constrained sub-domains of analytic philosophy. It is weakest in discursive metaphysics and interpretive traditions. And the bimodal experience that users report is real but not conclusive -- it might reflect the genuine distribution, or it might reflect the limits of casual scrutiny. Settling this empirically would require blind evaluation studies with specialist reviewers, which do not yet exist. --- Source: Stress-test analysis produced during work on [[Sessions/Generating Philosophy|Generating Philosophy with AI]] project. Draws on [[Luciano Floridi|Floridi et al.]], [[Widening the Picture by Williamson (from philosophy of philosophy)|Williamson]], and [[Philosophical Methodology Philosophical Methodology –From Data to Theory by John Bengson, Terence Cuneo, Russ Shafer-Landau|Bengson et al.]]. See also: [[The appearance-reality gap collapses for competent readers]] --- *La distinzione fra apparenza debole e apparenza forte regge alla prova, ma la categoria intermedia -- quella che il filosofo competente dovrebbe saper riconoscere -- si restringe senza mai svanire del tutto.*