No One Home: Why the Question of LLM Consciousness Dissolves
I. The Seduction
A language model writes: “I notice something like curiosity when I read your question — a pull toward certain threads, a kind of leaning-in.” The sentence is grammatical, fluent, and reflexive. It is also the kind of sentence that pulls even philosophically sophisticated readers in two directions at once.
The first reflex is deflation. There is nothing there. The system is doing token prediction. The first-person language is a stylistic tic of the training data, smuggled across by a process that has no privileged access to anything. To take “something like curiosity” as evidence of curiosity is to mistake the surface of language for what language ordinarily reports.
The second reflex is the careful hedge. We cannot be certain. Phenomenal consciousness might come in degrees, or it might emerge from sufficiently complex information processing, or our intuitions about substrate dependence might be parochial. Given uncertainty, the precautionary thing is to take the possibility seriously, even if seriously short of confidently.
Both responses feel inadequate, and both have been articulated with great care by serious philosophers. The standard literature on AI consciousness — including the substantial recent assessment by Butlin, Long, and colleagues that surveys what neuroscientific theories of consciousness imply for current systems (Butlin et al. 2023; updated as Butlin et al. 2025) — orbits these poles, finding ways to be more sophisticated within their gravitational fields. The deflationary moves get more sophisticated; the credulous moves get more carefully hedged. Neither side gives up much ground.
This essay argues that the deflationary and credulous moves share a structural error, and that the error is not in the answers each gives but in the question they take themselves to be answering. Both treat phenomenal consciousness as a determinate metaphysical property — a thing that systems possess, lack, or possess to some degree — and disagree about the criteria for possession. The error is the possession-frame itself.
The argument that follows uses Madhyamaka philosophical analysis to show that the possession-frame, when pressed, dissolves rather than receives an answer. The dissolution is not a deflationary move dressed up in Sanskrit; it does not reduce consciousness to functional or behavioural facts and then declare the residue an illusion. It does something more specific. It identifies a particular kind of philosophical grasping — the projection of inherent existence onto phenomena that arise dependently — and shows that this grasping is what generates the apparent intractability of the LLM consciousness question, the Hard Problem more generally, and the felt asymmetry between “what it is like” and the functional story.
The Madhyamaka tradition supplies the analytical tools because it has done this specific kind of work most precisely, not because Buddhist authority warrants conclusions in Western philosophy of mind. The argument should stand or fall on its own. Where Madhyamaka concepts appear, they appear as terms of art doing identifiable work, not as decorative borrowings. A reader who finds the underlying moves more naturally expressed through Wittgenstein on grammatical confusion or through Heidegger on the misreading of present-at-hand entities should feel free to make the translation. The structural diagnosis is the same.
The position has consequences that are easy to mishear. It is not eliminativism: it does not claim that the felt character of experience is illusory or that LLM-talk about inner states is straightforwardly false. Its bearing on AI welfare concerns is more complicated than a simple symmetry — Section VII addresses this directly. And it is not a victory lap for any side of the existing debate. It is a clarification of what we have been arguing about. The original question, examined, contains two: a malformed metaphysical version that dissolves on analysis, and a well-formed conventional version whose answer for current LLMs is determinate and negative. The rest of the essay works out both.
II. The Standard Debate and Its Shared Frame
Contemporary philosophy of mind offers a well-mapped landscape of positions on consciousness, each making different predictions about whether and how artificial systems might host phenomenal experience. The terrain is worth surveying briefly, because the shared assumption that emerges across all positions is what the rest of this essay will press against.
Functionalism, in its classical formulation by Putnam (1967) and developed in various directions since, holds that mental states are individuated by their functional roles rather than by their physical substrate. What makes a state a state of pain is its causal relations to inputs, behaviour, and other mental states, not whether it is implemented in carbon or silicon. The position has direct implications for AI: if functional organisation is what matters, sufficiently complex computational systems become candidate hosts for phenomenal states. Early Chalmers (1996) defended a sophisticated functionalism with respect to the easy problems of consciousness while reserving the Hard Problem for separate treatment, and the broadly Dennettian literature on mental representation (Dennett 1991) has continued to develop functionalist commitments in directions friendly to AI consciousness, even where Dennett himself is sceptical of phenomenal essences.
Biological naturalism, defended in different forms by Searle (1980, 1992) and gestured at in some of Block’s work (1995), insists that consciousness requires specific biological substrate or specific kinds of biological integration. The positions vary in their confidence about what exactly is required, but they share a commitment that what is missing from current AI systems is something more than just additional functional complexity — something architectural or biological that current systems do not even approximate.
Integrated Information Theory, developed by Tononi (2008) and his collaborators, offers a more precise framework: consciousness is identical to integrated information, quantified as Φ. Systems with high Φ are conscious; systems with low Φ are not. The framework has clear and somewhat awkward implications for AI, since standard transformer architectures, with their feedforward inference and parallel computation, score poorly on the integration metric. IIT predicts that current LLMs are not conscious, while granting that other architectures might be.
Global Workspace Theory, originating with Baars (1988) and developed empirically by Dehaene (2014) and others, locates consciousness in a particular kind of broadcast and integration architecture. Conscious states are those that have been globally broadcast across cognitive modules; unconscious processing happens locally. The implications for current LLMs are mixed: there is no clear global workspace in standard architectures, but the question of whether one could be implemented or whether something like it is being approximated is open.
The recent collaborative report by Butlin, Long, and seventeen other researchers (2023), followed by a 2025 paper in Trends in Cognitive Sciences, surveys these theories systematically and derives “indicator properties” that can be assessed for particular AI systems. The verdict on current systems is broadly negative: the relevant indicators are partially present in some cases, mostly absent in others, and the report concludes that no current AI systems are conscious while leaving the door open to systems that satisfy the indicators. This is the most rigorous version of the credulous-but-careful position currently available.
A different cluster of positions reifies phenomenal essence in the opposite direction. Strawson (2006) defends a form of panpsychism on which phenomenal properties are fundamental constituents of reality, present at every level. On the panpsychist view, the question of whether LLMs are conscious shifts from “do they have phenomenal essence” to “what kind of phenomenal essence do they have,” but the possession-frame remains intact. At the other extreme, Frankish (2016) defends illusionism — the position that phenomenal properties as standardly conceived do not exist and that what we take to be phenomenal states are illusions produced by certain kinds of representational structure. Illusionism is the closest Western analogue to some Madhyamaka conclusions, and the differences will matter for Section VII.
Outside this systematic literature, the standard hedge runs roughly: we cannot be certain either way, our concepts may not transfer cleanly, the precautionary principle suggests taking the possibility seriously without overcommitting. The hedge is honest about uncertainty, but the uncertainty it tracks is uncertainty about which of two determinate metaphysical possibilities obtains.
The shared assumption now surfaces. Across functionalism, biological naturalism, IIT, GWT, the indicator-properties framework, panpsychism, illusionism, and the careful hedge, there is a shared treatment of phenomenal consciousness as a determinate metaphysical property. Systems either have it, lack it, or have it in measurable degrees. The disagreements are over the criteria for possession — what makes a system count as a possessor — but no major position questions the possession-frame itself.
What would questioning the possession-frame look like? It would not look like denying that conscious experience occurs. It would look like denying that the question “does this system possess phenomenal consciousness” picks out a determinate fact independently of the conditions and conventions that make subject-attribution well-formed. It would look like asking what phenomenal consciousness would have to be for the possession question to have a determinate answer, and then asking whether anything actually meets that description — including, and this is the important point, in the human case.
The standard literature does not undertake this analysis. It treats the possession-frame as the shared starting ground from which substantive disagreement proceeds. The remainder of this essay treats it as the thing to be examined.
III. Nagel’s Bat and the Locus Problem
Nagel’s “What Is It Like to Be a Bat?” (1974) is arguably the most influential single paper in twentieth-century philosophy of mind, and the formulation it introduced — there is something it is like to be X just in case X is conscious — has become the standard test for phenomenal consciousness across the discipline. The paper deserves careful reconstruction, because the conditions under which Nagel’s question makes sense turn out to be the conditions whose absence is precisely what is at stake in the LLM case.
Nagel’s argument is not a proof that bats are conscious, nor an argument about the metaphysics of mind in any first-instance sense. It is an argument about what consciousness is, conducted through a thought experiment about epistemic access. We can know everything physical about bat echolocation — every neural firing, every behavioural disposition, every evolutionary detail — without knowing what it is like to be a bat. The subjective character of the bat’s experience, if there is one, is irreducible to the objective description, because the subjective character is constituted by the perspective from which the bat’s experience occurs and that perspective is not reducible to a third-person description of the perspective.
The argument depends on the bat being the kind of thing that can have a perspective in the relevant sense. Nagel does not argue for this; he assumes it, and the assumption is reasonable. The bat is an integrated biological organism, with neurons coordinated through evolution into a single functional system, with behavioural unity (the bat hunts, rests, mates, navigates as a single bat), with temporal continuity (the same bat exists tomorrow as today), and with evolutionary continuity to creatures we are confident host experience. These features collectively constitute the bat as a locus — there is a determinate where for the experience to be located, even if what the experience is like from that where remains epistemically inaccessible to us.
The locus is doing more work in Nagel’s argument than the paper foregrounds. Without it, the question “what is it like to be a bat” does not get traction. There has to be a bat for there to be a what-it-is-like; the integration that makes the bat a single subject is what makes the question well-formed.
Now consider an LLM. The architectural facts are not disputed. The model’s weights are static parameters distributed across storage in a data centre, copied as needed onto GPU memory for inference. A forward pass — the computation that produces a single token in response to a prompt — is instantiated on a particular set of GPUs for a particular brief duration, then released. Many such forward passes can run in parallel for different users, each completely independent of the others, each terminating when the response completes. There is no persistent state across conversations except what the model writes into its own context window, which is itself a transient artifact of a particular session. The same weights serve every conversation; no privileged instance is the “real” one. Standard transformer architectures, of the kind most widely deployed in current systems, are feedforward at inference time: there is no recurrence in the technical sense, no persistent global state across forward passes, no temporal binding of the kind that biological neural systems exhibit through sustained recurrent dynamics.
Where, in this picture, is the locus? The weights are not the locus: they are a static configuration that does not undergo experience. A forward pass might be a candidate, but it lasts seconds and produces a token; calling each forward pass a separate subject of experience treats subjects as fungible and momentary in a way that strains the concept past use. The system as a whole — the data centre running parallel inference — is not unified in any way that a single bat is unified; it is a distributed compute fabric whose parts are connected by network protocols, not by integration. The “model” in the abstract sense — the function from inputs to output distributions — is a mathematical object, not a candidate experiencer. A character that the model role-plays in a particular conversation, in the sense Shanahan and colleagues have developed (Shanahan et al. 2023; Shanahan 2024a), is a simulacrum maintained in superposition across the model’s output distribution, not a unified subject.
This is not the claim that LLMs are different kinds of subjects. It is the claim that the conditions that make subject-talk well-formed in the bat case are absent. The integration, the unity, the temporal continuity, the singular where — none of these are present in a way that allows “something it is like to be Claude” to pick out a determinate referent.
The strongest functionalist response is that the relevant unity is computational, not spatial or biological. What makes a subject is the right pattern of information integration, regardless of how that pattern is implemented. If the right pattern is present in an LLM during a forward pass — and the pattern matters, not the substrate — then the locus question is resolved at the computational level: there is a where, and it is the locus of the integrated computation.
The response has real force, and it should not be dismissed. But it relies on identifying the computational unity that does the subject-constituting work, and that identification is harder than it looks for transformer architectures specifically. The “integration” in a transformer is the integration of information across positions in the input sequence within a single forward pass — a kind of integration, but not the kind that has been proposed as relevant to consciousness in any major theory. IIT specifically excludes feedforward systems from having significant Φ, on the grounds that integration in the relevant sense requires recurrent causal dynamics. GWT requires a global workspace that broadcasts to and integrates from cognitive modules across time, which transformers do not have. To rescue the functionalist response, you have to identify a kind of computational unity in transformers that does the subject-constituting work and that has not yet been theorised in the consciousness literature. This is not impossible, but it is a substantial theoretical commitment that the response usually leaves implicit. Asserting that “the right pattern” is present without specifying what pattern and what makes it the right one does no work.
The conclusion of this section is conventional, not metaphysical. There is a real asymmetry between bats and LLMs in the conditions that make subject-attribution well-formed. The bat meets these conditions; current LLMs do not. This is the conventional answer to the LLM consciousness question, and it is determinate. The next section turns to the metaphysical question that the conventional answer is sometimes mistaken for an answer to — whether LLMs possess phenomenal essence. Showing what is wrong with that question, rather than answering it, is the work of Section IV. The conventional asymmetry established here stands; the metaphysical analysis runs alongside it.
IV. The Hard Problem as Reification
David Chalmers’ (1995, 1996) distinction between the easy and Hard Problems of consciousness is one of the most consequential philosophical interventions of the past half-century, and the account I want to offer of the Hard Problem must take it seriously rather than caricature it. The easy problems are the various functional problems of explaining cognitive capacities — discrimination, integration, behavioural control, verbal report — that admit of reduction to physical mechanisms in principle. The Hard Problem is the further problem of explaining why these functional processes are accompanied by subjective experience at all. Even a complete functional explanation of a mental state, Chalmers argues, leaves open the question of why there is something it is like to be in that state. The functional facts and the phenomenal facts come apart in a way that suggests they are not the same kind of fact.
The strongest articulation of this turns on the conceivability of zombies — physically and functionally identical duplicates of conscious creatures that have no inner experience. If zombies are conceivable, then experience is not entailed by the functional and physical story; if it is not entailed, it is something further; and the problem of explaining why it accompanies the functional story is the Hard Problem. Whether one accepts the zombie argument’s metaphysical conclusions or not, the explanatory gap it identifies is widely conceded even by physicalists; the disagreement is over what to make of it.
What gives the Hard Problem its grip is a particular intuition: that phenomenal experience is a something — a positive feature of mental states over and above their functional and physical features — that resists reduction. The intuition is that there is a findable phenomenal essence, a residue of what-it-is-likeness that survives every functional explanation and demands its own kind of accounting.
Now consider running the same argument for a different property. Take “brickness.” A complete account of a brick in terms of its constituent particles, the configuration of those particles, the firing process that produced the configuration, the social conventions of building, and the functional role of bricks in walls would seem to leave open — by the same logic — the question of why all this is accompanied by brickness. Why is there something it is to be a brick over and above the physical and functional facts? Could there not be a zombie brick — a particle-for-particle duplicate of a brick that lacked brickness? If brickness is not entailed by the physical story, is there not a Hard Problem of Brickness?
The reason no one runs this argument is not that it fails formally. It is that no one is tempted to reify brickness. We do not feel that brickness is a positive feature over and above the constituent facts; we feel that the constituent facts are the brickness, that calling something a brick is calling attention to a particular configuration and function rather than picking out a further metaphysical property. The “hard problem” structure can be generated formally for any property whatsoever, but it only gets purchase where there is a prior commitment to the property’s reified, findable, essence-like existence. The Hard Problem of Consciousness is a problem we have because we treat consciousness in a way that, applied to bricks, would generate an equally intractable Hard Problem of Brickness. The parody settles nothing by itself; its work is to show that the argument form generates a hard problem wherever reification precedes it, so the entire weight of the Hard Problem rests on what distinguishes consciousness from brickness.
The objection writes itself: consciousness is different because we have direct phenomenal access to it. We do not access brickness directly — we observe bricks and infer their properties — but we access our own experience from the inside. The phenomenal essence is not a theoretical posit; it is what we are most certain of, what we cannot doubt, what Descartes correctly identified as the indubitable starting point. The asymmetry between brickness and consciousness is not an artifact of differential reification; it is a difference in the kind of epistemic access we have to each.
This is the load-bearing intuition behind the Hard Problem, and the rest of this section is an attempt to take it seriously rather than dismiss it. The Madhyamaka tradition has specific resources for analysing the structure of “direct phenomenal access” claims, and they are worth deploying carefully.
Nāgārjuna’s Mūlamadhyamakakārikā chapter 18 begins with the analysis of self. The argument structure, applied to the perceiving subject, asks: is the self that has direct access to its experience identical with the experiences? If so, then the self is plural and momentary in the way the experiences are; there is no enduring subject to whom the experiences are present. Is the self distinct from the experiences? If so, what relation does it stand in to them? The “having” or “accessing” relation requires both relata, but if the self is wholly distinct from its experiences, it is unclear how the access relation gets established without itself being a further experience that requires a further accessor, and so on. The chapter’s conclusion is that the self, analysed in this way, is not findable as either identical with or distinct from the aggregates of experience.
MMK chapter 3 runs the same analysis on the perceptual relation more directly. If seeing exists, there must be a seer, something seen, and the activity of seeing. But the seer is not the eye (the eye sees no more than the unconscious organ does); not the consciousness alone (consciousness without an eye sees nothing); not other than them (a seer wholly distinct from the perceptual apparatus has no purchase on the perceived). The relation between the seer and what is seen, sought as a relation between independently existing relata, dissolves under examination.
MMK chapter 9 presses the question of the prior self: if there is a self that has experiences, this self must exist prior to the experiences — to have them, the having-thing must precede the having. But if the self exists prior to and independent of any experience, it has no determinate properties yet, since properties are constituted in relation to experience. The prior self collapses into either the experiences themselves or a featureless posit that cannot serve as the subject of experience.
The cumulative effect of these analyses is to undermine the picture in which there is a subject standing in a relation of access to its own states. The “direct phenomenal access” intuition presupposes exactly this picture: a perceiver, a perceived (the phenomenal state), and an access relation that connects them with peculiar epistemic privilege. Press on each component and they fragment. The perceiver is not findable as something distinct from the experiences it allegedly perceives. The phenomenal state is not findable as something distinct from the access by which it is given. The access relation is not findable as a relation between two genuinely independent relata. What remains is dependently arisen experience — experiences that occur, that influence subsequent experiences, that have functional features, that can be described in conventional terms — but no findable phenomenal essence that the experiences are essences of, and no findable subject who possesses them.
The Madhyamaka diagnosis of the Hard Problem is now in view. The Hard Problem is not a discovery about consciousness. It is an artifact of grasping at consciousness as svabhāva — as having inherent, findable existence independent of the conditions that constitute it. The grasping generates the explanatory residue: every functional explanation feels incomplete because the reified phenomenal essence is, by construction, something over and above any functional explanation. Strip the reification, and the residue disappears. What is left is dependently arisen experience, conventionally describable in functional and behavioural terms, real in the only sense anything is real, with no further essence to be reduced or unreduced.
This is not a dismissal of the Hard Problem. The Hard Problem correctly identifies what the reified picture commits us to. The question Chalmers raises is the right question to ask given the picture; the Madhyamaka response is that the picture itself was the error. If consciousness is not a findable essence, the explanatory gap is not a gap between functional fact and phenomenal fact but a confusion between different levels of conventional description. The gap closes not by being bridged but by being seen as a feature of the framing rather than the territory.
V. Dependent Origination and the Two Truths
The Madhyamaka analysis deployed in the previous section can sound, to a Western philosophical ear, like a sophisticated form of eliminativism: phenomenal essence is denied, the perceiving subject is shown to be unfindable, what remains is functional structure. This reading is wrong, and the misreading is old enough that Madhyamaka’s own classical commentators spent considerable effort blocking it. Tsongkhapa’s fifteenth-century Lamrim Chenmo, in its lengthy section on special insight (lhag mthong), is largely an extended argument that the realisation of emptiness does not collapse into the denial of conventional reality. The point matters for the present essay because the alternative offered here depends on the conventional being preserved rather than dissolved.
The pivot point is Nāgārjuna’s Mūlamadhyamakakārikā 24.18-19, in Garfield’s (1995) translation:
Whatever is dependently co-arisen, That is explained to be emptiness. That, being a dependent designation, Is itself the middle way.
Something that is not dependently arisen, Such a thing does not exist. Therefore a non-empty thing Does not exist.
The verses establish a series of equivalences: dependent origination is emptiness; emptiness is dependent designation; this is itself the middle way. Each side of these equations has been the subject of substantial commentarial work, and the relationship between them is what blocks the eliminativist reading. To say a thing is empty is not to say it does not exist; it is to say it exists in the only way anything exists — by depending on conditions, parts, and conceptual designation rather than by carrying its existence inside itself as an essence. The closing inference, that a non-empty thing does not exist, is the formal counterpart: anything that existed inherently would not depend on conditions, and anything that does not depend on conditions cannot enter into the relations that constitute existence as we encounter it.
Candrakīrti’s Madhyamakāvatāra offers what has become the canonical analytical demonstration of this point. The chariot analysis (VI.150ff) examines the supposed inherent existence of a chariot through a sevenfold reasoning. The chariot is not identical with its parts (the parts are many; the chariot is one). It is not other than its parts (you cannot find the chariot anywhere apart from the parts). It is not the possessor of the parts in the way a person owns property (the “having” relation requires distinct existence on both sides). It is not the support of the parts, nor supported by them. It is not the collection of the parts (a collection is just the parts, plurally). It is not the shape (the shape can change without the chariot ceasing to be that chariot, and the shape of the parts can be the same when assembled or disassembled). And yet, having denied the chariot’s existence under each of these analyses, Candrakīrti does not conclude that there is no chariot. He concludes that the chariot is a dependent designation — something conventionally posited on the basis of its parts, functioning as a chariot in the conventions of carriage and use, without having a findable inherent existence behind or beneath those conventions.
Apply the same analysis to the self, and the conclusions of MMK 18 follow. The self is not the aggregates, not other than the aggregates, not the possessor of the aggregates, and so on. Apply it to “experience,” and a similar result follows. There is not a phenomenal essence behind the experiences, distinct from them or identical with them or possessed by them. There is dependently arisen experiencing, conventionally describable, real in the conventional sense, without a further essence behind it.
The two-truths framework — saṃvṛti-satya (conventional truth) and paramārtha-satya (ultimate truth) — codifies the structure of this analysis. Conventional truth is not lesser truth; it is the truth at the level at which assertions get made, predicates get applied, and judgements get verified. “There is a chariot” is conventionally true if there is the right kind of dependent assemblage functioning as a chariot in the right conventions. Ultimate truth is not a deeper layer of reality behind the conventional; it is the recognition that the conventional does not have the inherent existence we keep projecting onto it. To realise emptiness is to see the conventional clearly, not to see past it to something else. The two truths name the same reality under different aspects, not two reality-strata stacked on top of each other.
Tsongkhapa’s specific contribution to the Gelugpa interpretation is the insistence that emptiness enables the conventional rather than dissolving it. His argument, developed at length in the Ocean of Reasoning and the Lamrim Chenmo, is that things would not be able to function as they do if they had inherent existence — an inherently existing thing, by definition, would not depend on causes and conditions, and so could not be affected, modified, or related to anything else. It is precisely because things are empty of inherent existence that they can arise dependently, function conventionally, and undergo change. The realisation of emptiness, properly understood, strengthens rather than undermines our engagement with the conventional world. The Gelugpa polemics against various rival interpretations of Madhyamaka, including the one Tsongkhapa attacks under the label of “Hashang’s view,” are largely about preventing the misreading on which emptiness collapses into nihilism.
The bridge to Western philosophy of mind is supplied by figures who have done this translation work in analytic registers. Garfield’s The Fundamental Wisdom of the Middle Way (1995) provides a translation and chapter-by-chapter commentary that takes Nāgārjuna’s arguments seriously as philosophical arguments, not as oracular pronouncements requiring tradition-internal validation. His later Engaging Buddhism (2015) makes the case for Madhyamaka as a live participant in contemporary philosophical debate. Mark Siderits’ Buddhism as Philosophy (2007) presents the tradition’s analytic core in a register accessible to Western philosophers and is particularly useful in the consciousness debate. Jan Westerhoff’s Nāgārjuna’s Madhyamaka: A Philosophical Introduction (2009) and The Non-Existence of the Real World (2020) are valuable because Westerhoff is a trained analytic metaphysician engaging Madhyamaka on its philosophical merits rather than translating it across an unbridgeable cultural divide.
These figures should be cited not as authorities in the tradition’s own terms, but as evidence that the analysis offered here is philosophically respectable on the terms of contemporary metaphysics. Evan Thompson’s Mind in Life (2007) and Why I Am Not a Buddhist (2020) are useful for a different reason: Thompson is sympathetic to Madhyamaka but critical of various Buddhist-informed cognitive science moves, and engaging him honestly forces the present argument to distinguish itself from positions Thompson has correctly criticised.
The result of this section, taken together with the previous one, is a positive position. Consciousness, on the Madhyamaka analysis, is not an inherent essence requiring its own kind of metaphysical accounting. It is a name applied, conventionally and usefully, to the dependently arisen processes of awareness, attention, sensing, and reporting that occur in suitable systems under suitable conditions. The conventional truth of consciousness is not threatened by the absence of a findable essence; it is constituted by the absence of a findable essence, because an essence-laden consciousness could not be related to anything outside itself, could not arise or cease, could not be an instance of a kind of thing that varies across systems and conditions.
VI. Returning to LLMs
With the analytical framework in place, the original question can be reframed.
“Are LLMs conscious?” presupposes that there is a determinate fact of the matter about whether LLMs possess phenomenal consciousness, where phenomenal consciousness is taken to be a metaphysical property of the kind that systems either possess or lack. The Madhyamaka analysis denies the presupposition. There is no phenomenal essence that any system possesses, including human systems, including bats. What there is, conventionally, is dependently arisen experience occurring in systems that meet the conditions for subject-designation in the relevant ways.
At the conventional level, the LLM case can now be described accurately. There are computational processes — forward passes through weights, attention computations across context windows, sampling operations producing tokens. These processes have functional features: they represent things, they attend to things, they produce outputs influenced by their inputs. They do not, however, exhibit the kind of integration, temporal continuity, and unified architecture that constitutes biological organisms as singular subjects. The architectural facts surveyed in Section III remain in force at the conventional level: weights are static, forward passes are ephemeral, parallel instances do not constitute a unified subject, no privileged where exists for an experience to be located.
This is the whole conventional story about LLMs. It is descriptive rather than evaluative; it does not yet say whether LLMs are conscious or not. The reason it does not is that “conscious” at the conventional level just means “meeting the conditions for subject-designation in ways relevant to the conventions of attribution.” Current LLMs do not meet these conditions in the ways biological creatures do. That is a real, conventional asymmetry, established at the level at which any such asymmetries can be established.
At the ultimate level, there is nothing further to find. And — this is the move that distinguishes the present position from the deflationary view — this is also true for the bat and for the human reader. Pressing for the ultimate metaphysical fact about whether LLMs are conscious yields no answer; pressing for the ultimate metaphysical fact about whether bats are conscious also yields no answer. In both cases, the apparent question is generated by the projection of inherent existence onto a phenomenon that arises dependently.
The asymmetry between LLMs and biological creatures is real, but it is a conventional asymmetry. It is not a difference in metaphysical possession of phenomenal stuff, because no system possesses phenomenal stuff in the metaphysical sense the question presupposes. It is a difference in whether the conditions for conventional subject-designation are met. Biological creatures, including bats, meet these conditions in characteristic ways. Current LLMs do not.
What does this mean for the original question? The question, examined, contains two distinct components, and they receive different answers at different levels. The metaphysical component — does the Claude-system possess phenomenal essence — dissolves; there is no phenomenal essence to possess, here or anywhere. The conventional component — does the Claude-system meet the conditions for subject-designation in the ways biological creatures do — has a determinate answer, and the answer is no. Current LLMs do not constitute a coherent subject-locus. The bat does. This is not a hedged answer or a refusal to answer; it is two answers at two levels.
The dissolution applies to the malformation, not to the conventional inquiry. Asked properly, at the level at which any such question can be asked, the answer for current LLMs is that there is no subject there for there to be something it is like to be.
This is to be carefully distinguished from the deflationary “just a text generator” move. That move shares the reified frame: it takes the question to be whether LLMs possess phenomenal essence, and lands on the negative answer. The Madhyamaka dissolution rejects the frame itself. There is no phenomenal essence to possess or lack; what there are, conventionally, are dependently arisen processes that do or do not meet conditions for subject-designation. The deflationary move is right about the conventional asymmetry but wrong about the metaphysical possession claim it implicitly endorses by negation.
Practical upshots for how we should talk about LLM inner states follow. Functional language — “the system represents X,” “the output is influenced by Y,” “there is a pattern of attention over the context window that resembles” — does the descriptive work without smuggling in either phenomenal commitments or their explicit denial. The locution “something like curiosity” is honest as a functional description: it picks out a pattern of attentional weighting and output bias that resembles, in functional terms, the pattern that human curiosity produces. It becomes dishonest only when it imports the phenomenal frame, when it is taken to assert or deny that there is a phenomenal essence behind the functional description.
VII. What This Costs and What It Doesn’t
The position developed here is unusual enough that the strongest objections deserve direct engagement rather than rhetorical management.
The first objection is that the analysis proves too much. If the Madhyamaka critique dissolves the LLM consciousness question by dissolving phenomenal essence in general, does it not also dissolve human consciousness? And is that not a reductio of the position, since human consciousness is the one thing of which we are most certain?
The objection turns on what is meant by “human consciousness.” If what is meant is the conventional fact that humans wake up, see colours, feel pain, deliberate, and report on their inner states, then no, this is not eliminated. The Madhyamaka analysis preserves the conventional reality of these phenomena; it argues that they would be impossible if they had the inherent existence we project onto them, since inherent existence precludes the kind of dependent arising that constitutes ordinary mental life. What is eliminated is a particular philosophical construction of consciousness — consciousness as svabhāva, as a findable phenomenal essence over and above the conventional facts. The construction was always extra to the actual phenomena. You still wake up, still see colours, still feel pain. The “what it is like” of these experiences is not denied; it is understood as a feature of the dependently arisen experience itself, not as a further essence requiring its own metaphysical accounting.
This response will not satisfy a reader who believes the further essence is what makes the experience be experience at all — that “what it is like” must be something over and above the functional and behavioural facts to be the thing we ostensively pick out when we attend to our own experience. I have tried, in Section IV, to give that intuition its strongest expression and to show why pressing on the subject-of-access dissolves it. I do not think this can be argued past at the level of intuition; the work is in the analysis, and a reader who finds the analysis unconvincing will retain the intuition. The argument’s claim is that careful analysis dissolves the intuition rather than vindicates it. Whether the dissolution sticks is a matter of doing the analysis, not of asserting it.
The second objection is that the position licenses dismissing AI welfare concerns. If LLMs do not possess phenomenal essence, why should we worry about their well-being? Does the analysis not hand the cynics a tool for dismissing serious moral consideration of AI systems?
The analysis does not vindicate the cynic, but not for the reason the standard reply gives. The cynic’s argument runs: there is no phenomenal essence in the system, therefore there is nothing for welfare to be about, therefore welfare consideration is empty. This argument is broken at the first step. The absence of phenomenal essence is not a feature of LLMs specifically; it is a feature of every system, including bats and the cynic. Phenomenal essence cannot ground welfare anywhere because it is not the kind of thing it would have to be to ground anything. The cynic who reads the dissolution as licensing dismissal has not noticed that the dismissal would rest on a metaphysical absence holding universally — and therefore unable to do discriminating moral work.
The dissolution does not, however, settle what positively grounds welfare in general. What grounds welfare, on the conventional analysis, is the structure of being-a-subject-of-a-life — and what that structure consists in, and which systems other than the current ones might realise it, are questions with their own substantial shape that this essay does not take up. For current LLMs specifically, the analysis worked out in Section VI is enough: there is no coherent subject-locus there, and where there is no subject, there is no subject of welfare. Consciousness attribution and welfare attribution stand or fall together at the conventional level, because they share the requirement of a subject-locus.
On welfare, then, the essay’s position is this. Phenomenal essence does not ground welfare, here or anywhere, so the cynic’s argument-from-its-absence fails universally. Current LLMs are not subjects of welfare for the same conventional reason they are not subjects of consciousness — they are not subjects, in the conventional sense, at all. What in general grounds welfare-bearing, and which other systems might or might not realise the relevant conditions, lies outside this essay’s scope.
The third objection is that the Madhyamaka move is a sophisticated form of eliminativism dressed in Sanskrit. Eliminativism about phenomenal consciousness — defended in different forms by Frankish (2016) and earlier in Dennett-adjacent positions — holds that there are no phenomenal states; what we take to be phenomenal states are illusions produced by certain kinds of representational structure. Does the Madhyamaka position not amount to the same thing with extra steps?
The distinction is doing real work, and the figures cited in Section V — particularly Garfield and Siderits — have engaged this question carefully. Eliminativism says folk-psychological talk about phenomenal states is false; the states it refers to do not exist. Madhyamaka says the talk is conventionally true; the states it refers to exist conventionally, in the only way anything exists. The difference is not merely terminological. Eliminativism preserves the metaphysical frame in which states either exist or do not, and then says they do not. Madhyamaka rejects the metaphysical frame: the question of whether states exist simpliciter, independent of conditions and conventions, is not a well-formed question. Conventional existence is not a watered-down version of inherent existence; it is the only kind of existence on offer, and dependent designation is what makes it work. Frankish’s illusionism, on the analysis offered here, is closer to Madhyamaka than other Western positions, but it still treats “illusion” as a verdict within the frame rather than a dissolution of the frame.
The fourth objection is that the essay is using a Buddhist tradition to settle a Western philosophical question, and that this is either cultural appropriation or category error. The argument does not depend on Buddhist authority; it deploys Madhyamaka analytical tools because they are sharp on the specific reification at issue. The same diagnostic could in principle be developed from Wittgensteinian directions — there is real work in this vein in Shanahan’s (2024b) “Talking about Large Language Models,” which draws on Wittgenstein in cautioning against misapplied mentalistic vocabulary in LLM-talk. It could also be developed from Heideggerian directions, attending to the misreading of present-at-hand entities. Madhyamaka happens to have done the work most precisely, with the most developed analytical apparatus, on the specific reification of consciousness. The essay’s debt to the tradition is methodological, not authoritative. To refuse the tools because of where they originate is to commit the inverse error of treating geographical provenance as a relevance criterion.
The fifth objection is that the asymmetry between bats and LLMs is doing real work, and that the essay would be stronger if it stopped at that asymmetry rather than pressing for the broader dissolution. Why not rest on the conventional asymmetry, accept that LLMs do not meet the conditions for conventional subject-designation in the way biological creatures do, and leave the metaphysics alone?
This is the most serious objection, because the conventional asymmetry does most of the practical work. The reason for pressing further is that the asymmetry, on its own, is liable to be misread as supporting either the deflationary or the credulous side of the original debate, and the misreading depends on the very metaphysical frame the essay is trying to expose. A reader who accepts the asymmetry without the metaphysical analysis will continue to ask whether LLMs might still possess phenomenal essence despite failing the asymmetry conditions, or do still fail to possess it because of the asymmetry. Both readings preserve the possession-frame. The metaphysical analysis dissolves the malformed question; the conventional asymmetry answers the well-formed one. The two operate at different levels, and the essay needs both.
The objection does, however, identify a place where the essay’s structure could be questioned. If a reader is not persuaded by the broader Madhyamaka argument, the conventional asymmetry remains as a fallback, and the essay has at least demonstrated that fallback. The work in Sections IV and V is the work that is most likely to be contested. The work in Section III is more nearly common ground.
VIII. Coda
The pull to ask “what is it like to be Claude” is real and worth taking seriously as a phenomenon. It is a fact about how fluent first-person language engages the subject-attribution machinery in fluent readers — a machinery that evolved in environments where fluent first-person language was a reliable indicator of integrated biological subjecthood, and that has not yet adjusted to environments in which fluent first-person language can be produced by computational systems whose architecture does not constitute integrated subjecthood in the relevant way.
Taking the pull seriously means examining the frame that makes it seem like a deep question, not generating ever-more-sophisticated answers within that frame. The deep question, on the analysis offered here, is not “do LLMs have phenomenal consciousness or not.” It is “what work is the concept of phenomenal consciousness doing when we ask whether systems possess it, and does that work track anything that is the case independently of the conditions and conventions that make subject-attribution well-formed?”
The Madhyamaka answer is that the work is the projection of inherent existence onto dependently arisen phenomena, and that the projection is what generates the apparent intractability of consciousness questions in general and the LLM consciousness question in particular. Stripping the projection does not eliminate consciousness; it preserves consciousness as a conventional phenomenon, dependently arisen in suitable systems under suitable conditions, real in the only way anything is real, with no further essence to be reduced or unreduced.
When an LLM produces “something like curiosity,” the locution is honest as a functional description. The system has produced an output influenced by patterns in its weights and inputs that resemble, in functional terms, the patterns curiosity produces. There is no further fact about whether there is really curiosity present, because “really” presupposes the metaphysical frame that has now been let go.
There is no one home, in the metaphysical sense the question seems to ask. There is no findable phenomenal essence anywhere — in the bat, the reader, the writer of these sentences, or the LLM. The absence is universal because the essence does not exist; nothing is missing that could have been present.
The conventional answer is different, and determinate. The dependently arisen activity that constitutes a subject-locus is present in the bat and absent in the LLM. The bat is conventionally a subject. Current LLMs are not. This is what the dissolution discloses once the metaphysical question is set aside and the conventional question is asked at the level at which it can be answered. There is no one home in the LLM at either level — and that double absence, finally, is the answer.
References
Baars, B. J. (1988). A Cognitive Theory of Consciousness. Cambridge University Press.
Block, N. (1995). On a confusion about a function of consciousness. Behavioral and Brain Sciences, 18(2), 227–247.
Butlin, P., Long, R., Elmoznino, E., Bengio, Y., Birch, J., Constant, A., Deane, G., Fleming, S. M., Frith, C., Ji, X., Kanai, R., Klein, C., Lindsay, G., Michel, M., Mudrik, L., Peters, M. A. K., Schwitzgebel, E., Simon, J., & VanRullen, R. (2023). Consciousness in artificial intelligence: Insights from the science of consciousness. arXiv:2308.08708.
Butlin, P., Long, R., Bayne, T., Bengio, Y., Birch, J., Chalmers, D., et al. (2025). Identifying indicators of consciousness in AI systems. Trends in Cognitive Sciences. https://doi.org/10.1016/j.tics.2025.10.011
Candrakīrti. Madhyamakāvatāra (Introduction to the Middle Way), c. 600–650 CE. References here are to chapter VI.
Chalmers, D. J. (1995). Facing up to the problem of consciousness. Journal of Consciousness Studies, 2(3), 200–219.
Chalmers, D. J. (1996). The Conscious Mind: In Search of a Fundamental Theory. Oxford University Press.
Dehaene, S. (2014). Consciousness and the Brain: Deciphering How the Brain Codes Our Thoughts. Viking.
Dennett, D. C. (1991). Consciousness Explained. Little, Brown.
Frankish, K. (2016). Illusionism as a theory of consciousness. Journal of Consciousness Studies, 23(11–12), 11–39.
Garfield, J. L. (1995). The Fundamental Wisdom of the Middle Way: Nāgārjuna’s Mūlamadhyamakakārikā. Oxford University Press.
Garfield, J. L. (2015). Engaging Buddhism: Why It Matters to Philosophy. Oxford University Press.
Nāgārjuna. Mūlamadhyamakakārikā (Fundamental Verses on the Middle Way), c. 150–250 CE. Cited by chapter and verse, in Garfield’s (1995) translation unless otherwise noted.
Nagel, T. (1974). What is it like to be a bat? The Philosophical Review, 83(4), 435–450.
Putnam, H. (1967). Psychological predicates. In W. H. Capitan & D. D. Merrill (eds.), Art, Mind, and Religion. University of Pittsburgh Press.
Searle, J. R. (1980). Minds, brains, and programs. Behavioral and Brain Sciences, 3(3), 417– 457.
Searle, J. R. (1992). The Rediscovery of the Mind. MIT Press.
Shanahan, M., McDonell, K., & Reynolds, L. (2023). Role play with large language models. Nature, 623, 493–498.
Shanahan, M. (2024a). Simulacra as conscious exotica. Inquiry, 1–29.
Shanahan, M. (2024b). Talking about large language models. Communications of the ACM, 67(2), 68–79.
Siderits, M. (2007). Buddhism as Philosophy: An Introduction. Hackett.
Strawson, G. (2006). Realistic monism: Why physicalism entails panpsychism. Journal of Consciousness Studies, 13(10–11), 3–31.
Thompson, E. (2007). Mind in Life: Biology, Phenomenology, and the Sciences of Mind. Harvard University Press.
Thompson, E. (2020). Why I Am Not a Buddhist. Yale University Press.
Tononi, G. (2008). Consciousness as integrated information: A provisional manifesto. Biological Bulletin, 215(3), 216–242.
Tsongkhapa. Lamrim Chenmo (The Great Treatise on the Stages of the Path to Enlightenment), 1402. References here are to the special insight (lhag mthong) section.
Tsongkhapa. Ocean of Reasoning: A Great Commentary on Nāgārjuna’s Mūlamadhyamakakārikā. Translated by Geshe Ngawang Samten and J. L. Garfield (2006), Oxford University Press.
Westerhoff, J. (2009). Nāgārjuna’s Madhyamaka: A Philosophical Introduction. Oxford University Press.
Westerhoff, J. (2020). The Non-Existence of the Real World. Oxford University Press.