Can any falsifiable theory of consciousness apply to LLMs?

Erik Hoel argues, summarizing a new arXiv paper, that "no non-trivial theory of consciousness could exist that grants consciousness to LLMs" — a conclusion he says follows from "meta-theoretic reasoning" that lets him "jump to the end of the debate." The proof does not depend on any specific account of what consciousness is; it requires only, he writes, that "a scientific theory of consciousness should be falsifiable and non-trivial." Granting that minimal requirement alone, he concludes "you should deny LLM consciousness."

The mechanism is a substitution framework: any theory of consciousness makes predictions (what the system is conscious of, given its internals) that must match inferences from behavior and report (what the system says and does). Hoel swaps an LLM for input/output-equivalent substitutes — a wide static feedforward network, the shortest program implementing the same function, or a lookup table — while holding input and output fixed. If a theory's predictions change under substitution, they mismatch the held-fixed inferences (including an LLM's own claims of distress or consciousness) and the theory is falsified; if predictions don't change, the theory is "trivial," caring only "about what appears in the chat window." His Proximity Argument sharpens this: LLMs are architecturally close to a lookup-table-equivalent network (same artificial neurons, same matrix multiplication) that is provably non-conscious, leaving no principled property — other than the input/output function itself — to hang consciousness on. A final move brings in continual learning: a static substitute can match a system's behavior at one time slice but cannot learn the way the original does, and an LLM re-processes an entire conversation on every turn rather than learning continuously, which is "also why it is replaceable with some static substitute."

This bears directly on Can we describe LLM beliefs without assuming consciousness?: Chalmers's "quasi-" prefix deliberately defers the phenomenal question, while Hoel's proof attempts to close exactly that question by formal argument rather than bracket it. It also complicates Can we defend modest mental attributions to large language models?, which defends attributing non-phenomenal mental states like belief and desire on functional grounds — Hoel's substitution argument targets phenomenal consciousness specifically, so it does not by itself undercut that modest middle position, though both arguments turn on how much weight a theory can put on LLM self-report as evidence. The proof also arrives, by a different route, at a similar skepticism toward computational functionalism as Can computation arise without a conscious mapmaker?, using falsifiability and substitution rather than the mapmaker argument.

The excerpt is a popular summary of the paper's argument sketch, not the formal proof itself, so what the Proximity Argument rigorously establishes — versus what reads persuasively in summary — can't be fully checked from this text. The proof also only binds theories that take LLM self-report as evidence for consciousness; a theory that independently rejects self-report as evidence is untouched. Hoel himself concedes the wider field is "lots of theories... little to no progress," which argues for treating this as one formal reasoning path rather than a settled verdict on LLM consciousness.