Source
- Birch (2025), AI Consciousness: A Centrist Manifesto, LIT-111, §§10–12, as read in NOTE-183.
- Schwitzgebel (2025), AI and Consciousness, LIT-191, chs. 6–10, as read in NOTE-104.
- Seth (2024), Conscious artificial intelligence and biological naturalism, LIT-135, §§3, 5, as read in NOTE-169.
- Butlin et al. (2023), LIT-056, §§1.2, 2, 3, as read in NOTE-052.
- Block (2002), The Harder Problem of Consciousness, LIT-639, §§IV–X and XIII, as read in NOTE-501.
- Block (2009), Comparing the Major Theories of Consciousness, LIT-632, the sections on the workspace, the biological theory and machine consciousness (pp. 1111–1113, 1119), as read in NOTE-505.
- Chalmers (2023), Could a Large Language Model be Conscious?, LIT-411, §§2–4 and fn 30, as read in NOTE-345.
What was actually shown
Behaviour. Four readings independently deny that humanlike behaviour is evidence of consciousness in a system trained on human output:
- Birch’s gaming problem (§10): a system incentivised to reproduce our markers empties them of evidential value.
- Schwitzgebel’s Mimicry Argument (ch. 7): a known mimic’s resemblance to the model no longer licenses the inference to the underlying feature. It undercuts and does not rebut (p. 55).
- Seth’s note on the Turing test (fn. 41).
- Butlin et al. §1.2.3.
Architecture. Here the readings diverge on the answer while agreeing on the structure:
- Butlin et al. derive 14 indicators on the explicit assumption of computational functionalism. They claim only that more indicators raise the probability.
- Seth (§§3.9, 5.2) makes conscious AI require functionalism and silicon substrate flexibility, and doubts both. For him the indicator route is conditional on what he rejects.
- Birch names this the Janus problem (§11). A global workspace found in an LLM would be read by a functionalist as evidence for consciousness, and by a biological naturalist as evidence that workspace theory is insufficient. “Both responses are reasonable” (p. 22). So the deadlock is at two levels: over the answer, and over what counts as evidence.
- Schwitzgebel adds that global workspace, higher-order and recurrence theories face minimal instantiation: trivial machines meet their minimal criteria (ch. 8). All of them rest on a narrow vertebrate evidence base (ch. 9).
The claim is the record’s synthesis across the four. Birch and Seth state its two halves most directly, and NOTE-169 records that Birch’s Janus sceptic “is Seth’s position in effect”.
What Block adds
Why the evidence is missing, for the idealised case. Block (2002), LIT-639, imagines Commander Data, a “merely superficial” functional isomorph of us whose physical realizers share nothing with our brain mechanisms beyond what toasters share. Functional likeness is a reason to ascribe consciousness, but a defeasible one (C2). A fundamentally different physical basis is a reason to doubt it, but one that “falls below the epistemic level of a ground for rational belief” (C3). And (C4): “We have no conception of a ground of rational belief to the effect that a realization of our superficial functional organization that is physically fundamentally different … is or is not conscious.” His consciousness is meta-inaccessible: “Not only do we lack a ground of belief, but we lack a conception of any ground of belief.” That is this account’s evidential half in its strongest form. The two sides lack the evidence, and also any agreed idea of what would count as evidence. Block also shows why shared laws would not break the tie. Whether Stevens’s power law is “a law of consciousness or a law of the human realization of consciousness” is “of a piece with” whether Data is conscious: “We cannot settle one without the other” (§X). That is the Janus problem of Birch (LIT-111) for laws rather than architectures.
Three limits keep this from being more than a source for the evidential half. Block calls meta-inaccessibility “a premise rather than a lemma or a conclusion”. The argument is conditional on naturalism and phenomenal realism, and a deflationist may run it backwards. Data is stipulated to be a functional duplicate, not a mimic trained on our output, so the paper gives no support to the mimicry clause. And Block holds that it is “the subjective default” that Data is not conscious while the question stays open. That puts him among the readings below that contest which way the prior leans.
The disputed premise, stated by a partisan. Block (2009), LIT-632, says that the global workspace’s “architectural aspects … can just as easily be realized in silicon-based computers as in protoplasm”, which makes it “a form of what philosophers call functionalism”. On his biological theory, broadcasting is what consciousness does, not what it is. Only “machines that have the right biology can have consciousness” on that theory. The electrical–chemical–electrical transfer of neural coding “may well be” necessary, a possibility “it would be foolish to discount … without evidence”. This is the architecture half of the claim from the other side. Butlin et al.’s workspace indicators (LIT-056) are evidence only on the premise Block rejects, and Block names no result that would show the chemical step matters or does not. He also notes that “there is a real empirical difference among the views that each side seems to think favors its own view”, which is the two-level deadlock in his words. The chapter favours the biological theory (“it is hard to avoid the impression that the biology of the brain is what matters”), so it is a source for the shape of the dispute and not for its answer.
What Chalmers adds
Behaviour, independently and earlier. Chalmers’s 2022 talk, LIT-411, holds that a model’s self-reports are fragile and that it “has learned to imitate those claims” from people talking about consciousness, so “the evidence is much weaker” (§2). That is the undercutting move of Birch’s gaming problem and Schwitzgebel’s Mimicry Argument, made before either. His tenth challenge, a model that describes features of consciousness it was not trained on, is a proposal for the training-immune marker this account names as a refuter, not an instance of one.
Architecture, as an instance. The obstacles that do the work in §3
(recurrence, a global workspace, unified agency) are theory-derived, and
building them into extended models is treated as progress. That is the
functionalist face of the Janus problem. Biology is set aside with a
reference to earlier work, though it enters the credences at 1/3. The
resulting figures, under 10% for current models and 25% or more for
conscious extended models within a decade, are credences under assumed
theories, which promote_when says cannot settle the question. They add
a fifth position on the prior. The classification of objections as
temporary is stated as THEORY-460.
What this does not say
- It does not say AI is or is not conscious, nor which way the prior should lean. The readings contest exactly this:
- Seth: unlikely on current trajectories, and false positives are the likelier error. He supports this weakly and concedes it is “not watertight”.
- Birch: flicker and shoggoth hypotheses cannot be ruled out for current LLMs.
- Butlin et al.: no current system is a strong candidate, though their abstract’s “no current AI systems are conscious” outruns the body.
- Schwitzgebel: we “will not know” before mass deployment.
- It does not say the deadlock is permanent. Birch argues it can be broken over the long run by comparative animal studies (§12, Fig. 4). Schwitzgebel’s iterative natural-kind strategy may converge, but on a timescale longer than AI development. Both are proposals, argued informally.
- Mimicry is a defeater, not evidence of absence. Schwitzgebel allows the output to remain weak evidence if one credits “the likeliest way to be an effective mimic is to actually be conscious” (fn. 95). How much post-training erodes mimic status is unanalysed (NOTE-104, C6).
- Biological naturalism is not shown to be testable or untestable here. Birch holds that a brute metabolism claim would be untestable, and that an explanatory one is testable. Seth frames his as explanatory, but names no property that separates conscious from non-conscious life.
Connections
-
LIT-159 and LIT-216 (Schwitzgebel): the same liberality pressure from the materialist side. Any criterion generous enough for rabbits and aliens credits the United States.
-
LIT-185: a defence of IIT against the expander-grid objection, which Schwitzgebel cites against IIT’s application to AI.
-
LIT-096 (Nagel): the definition of consciousness that Birch and Seth share.
-
LIT-982 (Seth & Bayne, Box 3; THEORY-298): the general form of this account’s evidential point. For infants, patients, animals and organoids as for AI, what counts as evidence is itself disputed, the measurement problem for creatures.
-
THEORY-606 (Shanahan, LIT-465) denies this account’s presupposition rather than taking a side: if consciousness words mean only what is public, there is no fact beyond an informed consensus, and a disembodied chat agent is not even a candidate.
-
THEORY-577 (the meta-problem challenge, LIT-425) extends the mimicry point to puzzled phenomenal reports, which a simple theorem-prover agent mirrors in part (fn 26), and asks of any theory whose indicators are used that its basis explain the reports.
THEORY-045 locates this deadlock for affect. On Barrett’s computational version, a system that predicts and regulates a body would have affect. Damasio & Carvalho and Seth tie affect to living tissue.