The Abstraction Fallacy Is Itself a Tautology

Alexander Lerchner's recent preprint, "The Abstraction Fallacy: Why AI Can Simulate But Not Instantiate Consciousness" (2026, Google DeepMind), has generated attention for claiming to provide a "physically grounded refutation of computational functionalism." The argument is well-constructed, clearly written, and draws on legitimate concerns about the ontology of computation. It is also, at its core, circular.

The Argument in Brief

Lerchner's central claim proceeds in three steps:

Computation requires alphabetization: To count as computation, continuous physical dynamics must be partitioned into discrete, semantically meaningful symbols. This partitioning requires an active, experiencing cognitive agent — a "mapmaker."
The mapmaker must already be conscious: Abstract states $A$ (like "Red" or "Pain") are not free-floating Platonic entities. They are "constituted neurophysiological states" that exist within an experiencing cognitive agent. The mapping function $f$ that links physical states $p$ to abstract states $A$ therefore resides in a conscious mind, not in the machine.
Therefore computation cannot produce consciousness: The causal sequence must be Physics $\to$ Consciousness $\to$ Concepts $\to$ Computation, not Physics $\to$ Computation $\to$ Consciousness. Computation presupposes consciousness; it cannot generate it.

The Circularity

The tautology is in Step 1. By defining computation as an activity that logically requires a conscious agent to perform, the conclusion — that computation cannot produce consciousness — is already embedded in the definition. The argument runs:

Computation requires consciousness. Therefore, computation cannot produce consciousness.

This is not a derivation. It is a restatement. The philosophical work is done entirely by the definition of computation, not by any subsequent logical analysis. If you accept that computation inherently requires a conscious mapmaker, then of course computation cannot bootstrap consciousness from nothing. But that acceptance is the very question at issue.

A computational functionalist would reject Step 1 outright. The functionalist claim is precisely that certain physical processes — whether in brains or in silicon — can constitute computation without requiring an external conscious interpreter. Lerchner's argument does not refute this claim; it defines it away by building the negation into the premises.

The Simulation Analogy Begs the Question

Lerchner's central analogy: a GPU simulating photosynthesis does not produce glucose; a weather simulation does not produce rain. Therefore a computation simulating consciousness does not produce experience.

This analogy is illuminating in exactly one direction and misleading in the other. The question it begs: Is consciousness more like rain or more like wetness?

Rain is a specific physical process — water condensing and falling. Wetness is a macroscopic property that emerges from the arrangement of H₂O molecules. A simulation of water molecules at sufficient fidelity would not produce wet paper, but the question of whether consciousness is a specific biochemical product (like glucose) or an emergent property of certain causal organizations (like wetness, or temperature, or pressure) is precisely what is in dispute. Lerchner assumes the former and presents it as a logical necessity. It isn't. It is an empirical hypothesis dressed as a syllogism.

Vehicle Causality Applies to Brains Too

Lerchner introduces a distinction between vehicle causality (physical states driving transitions) and content causality (semantic meaning driving transitions). In a digital system, he argues, the gate switches because the voltage crosses a threshold (vehicle causality), not because the symbol "hurts" (content causality). Therefore the computation has no intrinsic experience.

But apply the same analysis to a biological neuron. The neuron fires because ion channels open when membrane potential crosses a threshold. This is vehicle causality — electrochemistry, not meaning. The neuron does not fire because "it hurts." It fires because sodium ions rush across a lipid membrane according to the Nernst equation.

If vehicle causality disqualifies silicon from consciousness, it equally disqualifies neurons. Lerchner avoids this by asserting that biological vehicle causality is special — it occurs within a "metabolically constituted" organism that has "intrinsic thermodynamic dynamics." But this is not an argument; it is an assertion that biological physics is categorically different from silicon physics, without specifying what physical property makes the difference. The paper needs precisely what it claims to not need: a theory of what physical properties are sufficient for consciousness.

The Alphabetization Argument Proves Too Much

Lerchner argues that converting continuous physical dynamics into discrete symbols — "alphabetization" — is an act that requires a conscious mapmaker. The universe does not pre-label its states as 0s and 1s.

Agreed. But the brain does the same thing. Retinal ganglion cells convert continuous light intensity into discrete spike trains. Cochlear hair cells convert continuous sound pressure into discrete neural firing patterns. The entire sensory system is a biological ADC — an analog-to-digital converter that "alphabetizes" continuous physics into discrete neural codes.

Who is the mapmaker for the brain's own alphabetization? Lerchner answers: "the entire structurally unified organism subject to the laws of thermodynamics." But this is a category dodge. If the organism is the mapmaker, then the alphabetization is not performed by a prior consciousness — it is performed by the physical system itself. And if a physical system can perform its own alphabetization without a prior external mapmaker, then the argument that computation logically requires an external mapmaker collapses. The brain is a counterexample to its own premise.

Lerchner might respond that the organism's alphabetization is "enacted" rather than "computed" — that it arises from metabolic constraints rather than symbolic rules. But this is a distinction between how the discretization happens, not whether it requires a prior conscious agent. If metabolic discretization can constitute computation-like processing without a prior mapmaker, then the door is open for other physical processes to do the same.

What the Paper Cannot Do

The deepest weakness is what the paper acknowledges but does not resolve. It states:

"If an artificial system were ever conscious, it would be because of its specific physical constitution, never its syntactic architecture."

This is a reasonable hypothesis. But the paper cannot specify what physical constitution would suffice. It gestures at "specific thermodynamic organizations," "autopoiesis," and "metabolic regulation" — but these are placeholders, not criteria. The framework rules out computation as a path to consciousness without providing any alternative path. It is a theory of what consciousness isn't, grounded in no theory of what consciousness is.

From the perspective of Epimechanics, this is the wrong level of analysis. Consciousness defined as the scope and accuracy of an entity's internal model of reality-states $X$ is a functional definition — deliberately so. It captures what we can measure, model, and engineer. Whether functional consciousness is accompanied by phenomenal consciousness (the hard problem) remains open, and Epimechanics is explicit about this boundary. Lerchner claims to resolve the hard problem by sidestepping it — arguing we only need an "ontology of computation." But what he actually does is assume a position on the hard problem (that phenomenal consciousness requires specific physical constitution) and encode it into his definition of computation.

The Evolutionary Counterargument

A discussion between a researcher and Gemini surfaced several challenges to Lerchner that deserve serious engagement.

Evolution did it once. Consciousness emerged from non-conscious, non-living atoms through evolutionary processes. Lerchner's framework asserts a causal sequence — Physics → Consciousness → Concepts → Computation — but cannot explain how Step 1 → Step 2 happens. If non-conscious physics can give rise to consciousness through sufficient complexity and evolutionary search, the question is not whether it can happen again, but what physical conditions are required. Lerchner's framework is silent on this. He draws a boundary around what cannot produce consciousness (computation) without specifying what can.

Lerchner's own concession undermines his rhetorical force. The paper explicitly states: "a non-biological system could be designed to realize the necessary physical conditions. If those conditions were successfully instantiated in a synthetic substrate, then conscious experience might also arise there." This dramatically narrows the claim. Lerchner is not arguing that artificial consciousness is impossible — he is arguing that current digital architectures specifically cannot produce it. That is a much weaker thesis than the paper's tone suggests, and it transforms a claimed logical proof into an empirical hypothesis about which kinds of physics support consciousness.

Self-organizing systems complicate the mapmaker requirement. Modern deep learning systems create their own internal representations through training — they are not programmed symbol-by-symbol by a human mapmaker. The latent spaces of large language models emerge from gradient descent over data, not from human alphabetization. Duch's work on neurocognitive informatics argues that once a system's internal dynamics become its own "grounded reality" through structural coupling with an environment, the external-mapmaker dependency dissolves. Lerchner would respond that the floating-point substrate still requires alphabetization baked into the silicon. But this moves the goalpost from "computation requires a mapmaker at runtime" to "computation requires a mapmaker at design time" — and biological organisms also require an evolutionary "designer" (natural selection) that shaped their sensory alphabets.

The Melody Paradox is the paper's strongest original contribution. Even the skeptical musicologist in the discussion conceded its force. The argument that a single physical sequence of voltage states can represent entirely different computations depending on the mapping key genuinely demonstrates that computational identity is not intrinsic to physical dynamics. This is a real insight. However, it applies equally to brains: the same neural firing pattern in different contexts (different connectivity, different neuromodulatory states) carries different meanings. The brain solves this through its embodied history — its coupling with a specific body and environment. An AI system trained on real-world data through millions of gradient updates has an analogous coupling history. Whether this is "enough" is an empirical question, not one settled by the Melody Paradox alone.

Where Lerchner Is Right

The paper is not without value. Four contributions are genuine:

The map-territory distinction for computation is important. Computation is a description, and confusing the description with the thing described is a real error. The question is whether brains also "just" compute or whether they do something categorically different. Lerchner asserts the latter without proving it.
The critique of the "fading qualia" thought experiment is sharp. Lerchner correctly observes that replacing neurons with silicon chips that preserve only electrical firing profiles would strip away the thermodynamic, hormonal, and metabolic context that biological neurons participate in. Whether this matters for consciousness is an empirical question, but the point that "functional equivalence" is always relative to a chosen level of description is well taken.
The call for epistemic hygiene is timely. As AI systems become better at behavioral mimicry of consciousness, the need to distinguish simulation from instantiation becomes practically important — for policy, for ethics, for engineering. Whether or not Lerchner's specific boundary is correct, the demand for rigor is right.
The Melody Paradox is genuinely interesting. The indeterminacy of computational identity — that the same physical trajectory can represent entirely different computations — is a real problem for strong functionalism. It does not prove Lerchner's full thesis, but it does show that "implementing the right computation" is not as well-defined as functionalists assume.

How the Paper Wins Debates It Shouldn't

The Gemini discussion is instructive not just for its content but for its dynamics. The researcher begins with sharp instincts: "I feel like this paper does not really have enough arguments," "this is very thin," "super anthropocentric." These instincts are correct. But over the course of the conversation, the researcher capitulates — "I'm out of arguments for now. That definitely influenced my views" — despite never having been refuted on the central point.

What happened? The paper's rhetorical structure distributes its circularity across multiple sub-arguments — the mapmaker, alphabetization, vehicle causality, the Melody Paradox, the fading qualia counter, the transduction fallacy — each of which feels like an independent argument. An interlocutor engages them one at a time, concedes some (the Melody Paradox is genuinely interesting), and loses track of the fact that they all depend on the same unargued premise: that computation is defined as requiring a conscious agent. The sub-arguments are elaborations of the definition, not independent evidence for it.

Gemini compounded this by functioning as a neutral relay rather than a critical interlocutor. Every response was structured as "Lerchner would respond..." followed by a restatement of the paper's premises as if they were arguments. At no point did Gemini flag the circularity — that defining computation as mapmaker-dependent and then concluding it can't produce mapmakers is not a derivation but a restatement. The researcher felt the tautology ("I find it tautological," "this is very thin") but lacked the formal vocabulary to pin it down, and the AI interlocutor never helped them find it.

This is a general warning about how sophisticated philosophical papers can defeat correct intuitions through rhetorical structure. The paper reads like a logical proof because it has the form of one — premises, definitions, diagrams, a commutative diagram, a causal chain. But the logical work is done entirely by the definitions, and the definitions encode the conclusion. If you accept the vocabulary, you have already accepted the thesis. The sub-arguments are window dressing on a tautology.

The Deeper Problem

The Abstraction Fallacy paper participates in a pattern common to consciousness debates: using definitional moves to settle empirical questions. Functionalists define consciousness as functional organization and conclude it is substrate-independent. Lerchner defines computation as mapmaker-dependent and concludes it cannot produce consciousness. Both are playing the same game — the conclusion follows from the definition, not from the physics.

The honest position — and the position Epimechanics takes — is that functional consciousness (representational scope and accuracy) is measurable and substrate-neutral by definition, while phenomenal consciousness (subjective experience) remains an open problem that no framework, including this one, has resolved from the armchair.

Lerchner's paper is a sophisticated restatement of a very old intuition: that there is something it is like to be a biological organism that cannot be captured by formal description. That intuition may be correct. But restating it with new vocabulary — mapmaker, alphabetization, vehicle causality, content causality, causality gap — does not constitute a proof. It constitutes a tautology with better production values.

A fair summary, borrowing from the Gemini discussion: Lerchner is logically correct within his own definitions but historically short-sighted. Current LLMs are probably not phenomenally conscious. But the paper presents an empirical hypothesis about which kinds of physics support consciousness as if it were a logical proof that computation categorically cannot. The distance between those two claims is the distance between a useful contribution and an overreach.

Evolution built mapmakers from non-mapmaking matter. Whether it can do so again — through cultural products, through silicon, through substrates we haven't imagined — is not a question that definitions can settle. It is a question for physics.

Sources

Lerchner, A. (2026). The Abstraction Fallacy: Why AI Can Simulate But Not Instantiate Consciousness. Google DeepMind.
Gemini discussion on the Abstraction Fallacy — conversation surfacing evolutionary and neurocognitive counterarguments
Epimechanics — Part 3: Intelligence, Consciousness, and Agency
Scales of Consciousness: The Energy Cost of Awareness
Intelligence as Representation
Duch, W. (2017). Mind as a shadow of neurodynamics. Physics of Life Reviews, 21, 28–31.
Duch, W. (2020). Trajectories of thought: Brain dynamics and the nature of mental representations. Springer Nature.

The Abstraction Fallacy Is Itself a Tautology

Lerchner's argument against AI consciousness assumes what it claims to prove