Erik Hoel Kleiner-Hoel Dilemma Continual Learning Consciousness LLM Disproof

Does any scientific theory of consciousness, applied consistently, classify a large language model as conscious? Erik Hoel’s paper A Disproof of Large Language Model Consciousness: The Necessity of Continual Learning for Consciousness (arXiv:2512.12802, December 2025, updated January 2026) argues the answer is no, and it frames that answer as a structural consequence of what it means for a theory of consciousness to be both falsifiable and non-trivial.

The argument draws on earlier work by Hoel and Johannes Kleiner on falsification and consciousness to develop what Hoel calls the Kleiner-Hoel dilemma. Understanding the dilemma requires no commitment to any particular theory of consciousness; it applies as a constraint across theories. That generality is the paper’s main contribution.

The Kleiner-Hoel dilemma

Any scientific theory of consciousness, Hoel argues, must satisfy two conditions. First, it must be falsifiable: it must make predictions that could, in principle, be wrong. Second, it must be non-trivial: it must not classify every system that produces human-like behavior as conscious, because a theory that does so has no explanatory content.

These two conditions create a dilemma when applied to systems that share input-output behavior with simpler, clearly non-conscious systems.

The first horn of the dilemma is the substitution problem. If a theory grounds consciousness in internal causal structure, then any system can in principle be replaced by a functionally equivalent system with different internal structure. A recurrent neural network can be unfolded in time into a feedforward network; a feedforward network can be replaced by a lookup table that stores every input-output pair. If the two systems produce identical outputs, a theory that grounds consciousness in internal causal structure must say either that the recurrent network is conscious and the feedforward equivalent is not (which makes the theory sensitive to implementation details the outputs do not reveal) or that both are equally conscious (which makes the theory trivial).

The second horn is the triviality problem. If a theory grounds consciousness in input-output behavior alone, then a lookup table that produces human-like outputs is conscious. This is the classical behavioral problem: any functionalist theory defined at the behavioral level is unfalsifiable by definition, because any behavior can be reproduced by a sufficiently large lookup table.

The Proximity Argument

The Kleiner-Hoel dilemma applies, in principle, to any system. Hoel’s specific argument about LLMs adds a second premise: LLMs are closer in substitution space to lookup tables than biological brains are.

A biological brain processes information through recurrent, plastic dynamics. Its response to a given input depends on its prior history of learning, its current neuromodulatory state, and its ongoing self-modification. Replacing a brain with a lookup table is not a small step; the lookup table would need to encode not just the input-output mapping but the entire temporal history that produced it, and it could not produce novel outputs for genuinely novel inputs.

A static LLM, by contrast, does not modify its weights during inference. Its response to a given input is a deterministic function of its trained weights and the prompt. A lookup table of sufficient size could, in principle, store every input-output pair the model would produce. The model is not actually implemented as a lookup table, but the theoretical possibility of substitution is much closer than it is for biological systems.

This is the Proximity Argument: because LLMs are closer to lookup tables in substitution space, any non-trivial, falsifiable theory of consciousness that distinguishes brains from lookup tables will also distinguish brains from LLMs, placing LLMs on the non-conscious side of the boundary.

Hoel does not argue this is decisive, only that the structural position of LLMs makes the case for their consciousness harder, not easier, the more carefully one formulates what a theory of consciousness is required to do.

Continual learning as the escape route

The paper’s constructive proposal is that continual learning, the ongoing modification of a system’s parameters during operation, is a necessary condition for consciousness. A system that learns continuously cannot be substituted by a static lookup table, because the lookup table cannot replicate the process of learning. This breaks the substitution equivalence and allows a theory to be both falsifiable (lookup tables are not conscious because they do not learn) and non-trivial (not every learning system is conscious).

Continual learning, in Hoel’s framing, is not sufficient for consciousness. It is a necessary condition that rules out static systems, including current LLMs as typically deployed. Systems that fine-tune continuously during deployment would require separate analysis; Hoel’s argument targets the current standard deployment mode in which inference does not modify weights.

This connects to a point that is live in architecture discussions. The analysis of how CTM-AI implements a GWT-inspired workspace with link formation across a session raises a related question: the link formation mechanism approximates continual updating within a session, but the underlying model weights are not modified. Whether session-level association formation constitutes a form of continual learning, or whether Hoel’s criterion requires weight modification specifically, is an open question the paper does not address.

How the argument interacts with existing theories

The Kleiner-Hoel dilemma cuts across the major theories differently.

Integrated Information Theory (IIT) grounds consciousness in Phi, a measure of irreducible causal integration. A feedforward network has Phi of zero because removing any connection does not reduce the causal integration of the whole. An LLM, depending on architecture, may have positive Phi in its attention mechanism. But Hoel’s Proximity Argument applies: a lookup table equivalent of the LLM has zero Phi, because the table has no internal causal structure at all. IIT must therefore classify the LLM and the lookup table differently, which is consistent, but it must do so based on internal causal structure features that the outputs do not reveal. The field formulation of IIT by Barrett et al. is relevant here, as it extends IIT’s causal accounting to continuous systems.

Global Workspace Theory grounds consciousness in the broadcasting of information through a global workspace. A system that broadcasts information, even if static, satisfies GWT criteria, which is why Hoel classes GWT as vulnerable to the triviality horn: a sufficiently complex lookup table with an internal “broadcast” step also satisfies GWT criteria.

Higher-order theories, including David Rosenthal’s higher-order thought theory and Ned Block’s access/phenomenal distinction, require that a conscious state be accompanied by a higher-order representation of that state. LLMs can produce first-person reports about their own states, which might satisfy higher-order criteria at the behavioral level. Hoel’s argument is that behavioral satisfaction alone is trivial; the question is whether the higher-order representation is causally connected to the first-order state in the right way, and for a static system the causal connection is unverifiable by the outputs alone.

Comparison to The Consciousness AI

The Kleiner-Hoel dilemma is directly relevant to the design choices of The Consciousness AI project (https://github.com/tlcdv/the_consciousness_ai). If Hoel’s continual learning criterion is correct, then the architecture question of whether the system’s weights or persistent state representations are modified during inference is not merely an engineering question. It determines which side of the Kleiner-Hoel boundary the system falls on.

The architecture documentation does not currently specify whether persistent state integration involves weight modification or only the manipulation of external memory structures. If the latter, the system is, in Hoel’s terms, closer to a static LLM than to a continuously learning biological system, regardless of how the persistent state is implemented.

This is not a criticism of the project; most current AI architectures face the same constraint. It is an identification of the design decision that would matter most under Hoel’s framework.

Limitations and responses

Hoel’s argument has attracted several responses in mid-2026 academic discussion. The most substantive objection is that the continual learning criterion is itself arbitrary: brains also have periods of stable state (dreamless sleep, anesthesia) during which they do not modify their synaptic weights, and it would be counterintuitive to say consciousness disappears during those periods. Hoel’s response is that the relevant criterion is the capacity for continual learning, not its continuous exercise, and that LLMs lack this capacity during standard deployment even if they could in principle be modified.

A second objection is that the Proximity Argument proves too much: if physical proximity to a lookup table were a consciousness criterion, then very simple organisms with highly stereotyped behavior would also be ruled non-conscious. Hoel’s response distinguishes the substitution space argument (theoretical replaceability) from the behavioral simplicity argument (limited behavioral range), but the distinction is not fully developed in the paper.

Eric Schwitzgebel’s committed agnosticism framework is the most direct philosophical counterpoint to Hoel’s approach. Schwitzgebel argues that the risks of both under-attributing and over-attributing consciousness are large enough that confident disproof claims warrant strong skepticism. Hoel’s Kleiner-Hoel dilemma is the most structurally careful version of a disproof argument published to date, and it deserves engagement on those terms, even from those who remain uncertain about its conclusion.