Why Hoel’s Disproof of LLM Consciousness and Functionalism Fails

Hoel (2026) offers what he presents as a formal disproof of large language model (LLM) consciousness, building on the Kleiner-Hoel dilemma to argue that no non-trivial, falsifiable theory of consciousness could count contemporary LLMs as conscious. The result is striking because it claims theory-neutrality: it purports to settle a contested empirical question through formal reasoning alone. This paper argues that the proof fails, and that its failure is instructive about the limits of formal methods in consciousness science. I reconstruct Hoel's argument as six premises, accept the first, and reject the rest. The central error lies in a conflation of first-person and third-person consciousness. Hoel's framework requires that its prediction and inference functions both terminate in the space of phenomenal first-person experiences, yet neither function has direct empirical access to that space. Instead, both can only operate on third-person data. The dilemma between independence (falsification by substitution) and dependence (triviality) is generated by this impossible demand. Once we recognize that empirical science directly studies only third-person consciousness, and that its relation to first-person experience is a separate metaphysical question, the dilemma dissolves. Functionalist theories such as Global Workspace and Higher-Order Thought remain falsifiable; their convergence of prediction and observation is confirmation, not triviality. Hoel’s argument is also shown to depend on an unjustified asymmetry between brains and LLMs. Because lookup tables and universal approximation are substrate-independent results, the substitution chain that supposedly disproves LLM consciousness applies equally to biological brains, so the proof proves too much. Blocking this symmetry requires a positive commitment, whether substance dualism, Penrose-style non-computability, or biological naturalism. I argue this point without assuming computational functionalism and show that fine-grained functionalist considerations of the kind Cao (2022) raises, far from rescuing the argument, undermine Hoel's own substitution move. I further argue that the relevant lookup tables are physically impossible to construct for both systems, that intuitions about such impossible objects cannot ground a formal proof, and that the proposed continual-learning criterion is tautological, ad hoc, and inconsistently applied. Consciousness cannot be proven or disproven from armchair reasoning about computation; the path forward is empirical.