Octopedia On Consciousness
The core problem is that “consciousness” is not one cleanly defined scientific variable. It is a cluster of related ideas that people constantly mix together: wakefulness, subjective experience, self-awareness, reportability, attention, memory, agency, and metacognition. In serious discussions, the first move has to be: which one are we talking about?
The cluster is easiest to see where it comes apart, and it comes apart at the bedside. A patient in a vegetative state is awake, eyes open, sleep cycles running, and by most accounts there is very little experience inside. Locked-in syndrome is the mirror image, everything inside and almost no way out. In 2006 Adrian Owen’s lab put a patient diagnosed as vegetative into a scanner, asked her to imagine playing tennis, and recorded the same motor planning activity a healthy brain produces. Follow-up work found that a meaningful fraction of patients diagnosed as vegetative can do this. The diagnostic error was available because report was the only window and the window was broken. Under anesthesia, memory and report vanish while experience is only assumed to vanish, since report is the only evidence of experience anyone ever had. In REM sleep the body is paralyzed while experience runs loud. Even the self can be peeled off the bundle: in depersonalization, people keep having experiences while the sense of being the one who has them goes strangely impaired. Wakefulness, experience, access, report, and self-modeling are separable variables. We treat them as one thing because evolution bundled them in us, and we have only ever sampled the bundle.
The hardest meaning is phenomenal consciousness: the fact that there is “something it is like” to have an experience. Pain hurts. Red looks a certain way. Fear feels like something from the inside. That is the version that creates the classic hard problem. It is not currently formally defined in the way mass, charge, entropy, or computation can be formally defined. We can correlate it with brain states, reports, behavior, anesthesia, lesions, attention, and neural dynamics, but the first-person feel itself is not directly measurable from the outside.
The trap is old. Leibniz saw it in 1714. Scale up a thinking machine until you could stroll through it the way you walk through a mill, and inside you would find parts driving parts and nothing that resembles perception. Add structure and scale to taste; the gap survives the tour. In 1974 Thomas Nagel gave the gap its standard phrasing: an organism is conscious if there is something it is like to be it, and there is something it is like to be a bat riding echolocation, a fact that stays shut against us no matter how completely we map the bat’s sonar apparatus, because physical description delivers structure and function while the feel is supposed to sit somewhere beyond both. Frank Jackson’s Mary sharpened the point in 1982. A scientist who knows every physical fact about color vision but has lived in a black-and-white room seems to learn something new the first time she sees red. If that intuition is right, physical facts underdetermine phenomenal facts. Chalmers made the division official in 1995: the easy problems, explaining discrimination, integration, report, and behavioral control, are hard but ordinary neuroscience, and the hard problem, why any of it comes with experience at all, is not currently anyone’s ordinary anything.
The reason formal definition keeps failing is specific. Physics speaks in structure, dynamics, and relations. A formal criterion for consciousness would have to be written in that vocabulary and come out the far end as feel, and nobody has ever produced a derivation that does it. Joseph Levine called this the explanatory gap in 1983, and unlike most gaps in science it has not visibly narrowed. Temperature went down to mean kinetic energy. Life went down to chemistry with replication and metabolism, and the vitalists who said the reduction was impossible turned out to be wrong. Maybe the vitalists of consciousness are wrong the same way, and this is simply the last century before the reduction lands. That is a live possibility. Today it is an article of faith.
Then there is access consciousness, which is much more operational. A representation is access-conscious if it is available for report, reasoning, memory, planning, and control of action. This is the kind of thing global workspace theory talks about. It is easier to study because you can test whether information is reportable, whether it affects decisions, whether it is broadcast to many systems, and whether it can be held in working memory. But access consciousness is not automatically phenomenal consciousness. A system might have globally available information without there being anything it feels like to be that system.
Ned Block coined the phenomenal/access split in 1995 for exactly this reason: one of them was turning into science and the other was standing still. Global workspace theory, Bernard Baars’s idea in cognitive form and Stanislas Dehaene’s in neural form, describes a bottleneck. Many parallel processes compete, one wins, and the winning content is broadcast across the brain in a frontal and parietal burst of amplification a few hundred milliseconds after the stimulus. The broadcast is what makes content reportable, memorable, and usable. The laboratory paradigms all test the broadcast. Mask a stimulus so briefly that it never ignites and people cannot report it, yet it primes later choices, so it was processed. Load attention elsewhere and half of viewers miss a gorilla walking through the middle of the scene. Rivalry, attentional blink, the whole toolkit: every one of these measures ignition. Whether ignition is felt is a question the theory, as a theory, does not touch.
Blindsight is what keeps the two from collapsing back into each other. Patients with damage to visual cortex will tell you they see nothing in part of their field. Forced to guess about stimuli in that region, they guess far above chance while insisting they are only guessing. Information good enough to guide action is present. Acknowledged experience is absent by the patient’s own report. If function and feel can come apart inside a single human skull, then in any system whatever, behavioral and report-based tests forfeit the right to assume they travel together.
So the honest answer is:
Consciousness is not currently formally defined as a single physical property. Some functional aspects of consciousness can be operationalized and measured, but subjective experience itself is not yet reducible to a settled formal criterion.
That is why AI consciousness debates are so slippery. One person says “the model has a global workspace,” meaning some information is available to many internal processes. Another hears “it has experience.” Those are completely different claims.
The dodge is even written into the founding document. Turing, in 1950, was handed the objection that thinking requires consciousness and no machine could ever genuinely have it; he noted that pressing the objection leads to solipsism, declined to go there, and offered the imitation game in its place. It was one of the most productive deflections in the history of science, and it made the entire field a wager on behavior. Seventy-five years later the wager is being called. The machines finally pass at the words, and the question the deferral set aside is standing at the door.
The slipperiness also predates the technology. Weizenbaum’s ELIZA, 1966, a few hundred lines of pattern matching, and his secretary asked him to leave the room so she could talk to it privately. Humans parse fluency as presence, the parser is default-on, and it is far older than computers. Large language models press on that reflex with the strongest stimulus ever constructed, because fluency at report is the training objective. An LLM is an interview subject assembled out of every interview humanity has conducted with itself. Ask about its inner life and the answer is drawn from a distribution over what humans have said about theirs. A report like that tells you about the training distribution. It tells you nothing about the machine.
If someone means consciousness as functional access, then maybe you can define criteria: information integration, reportability, attention-like selection, working memory, recurrent processing, self-monitoring, etc. These can be experimentally tested.
That move has been done seriously. In 2023 a team of nineteen researchers led by Patrick Butlin and Robert Long, Yoshua Bengio among the authors, derived indicator properties from the leading scientific theories of consciousness, global workspace, recurrent processing, higher-order representation, attention schema, agency and embodiment, and audited current AI systems against the list. The headline was honest: no current system meets the indicators of any theory. The structural limit runs deeper, and the report concedes it. The indicators inherit the assumptions of the theories they came from. A system could implement every global workspace indicator while global workspace theory remains a theory of access with nothing to say about feel. A checklist built from theories of consciousness tests whether an architecture implements a function. Whether that function was ever the thing you meant is a separate question the checklist cannot reach.
If someone means consciousness as felt experience, then no one has a generally accepted formal test. You can propose theories, but none has won.
It helps to be exact about why the test is missing. For other humans, no test is ever run. We extend the assumption by analogy: shared bodies, shared behavior, shared developmental origin, being the same kind of thing. For animals, the analogy has hardware to run on: nervous systems homologous to ours, pain behavior, trade-offs between threat and reward, responses to analgesics. The 2012 Cambridge Declaration, affirming the neural substrates of consciousness in non-human animals from mammals and birds to octopuses, was a reasonable call, because homology carries the inference. For an AI there is no homology at all. The analogy runs entirely on text output, and text output is the one channel that was explicitly optimized to imitate us. It is as if the only evidence about whether some creature could see were pictures it had painted of landscapes.
The metaphysical split is roughly this:
Physicalist view: consciousness is a physical process or property of certain physical systems, probably brains, and in principle should be explainable by neuroscience and computation. Standing bill: the explanatory gap, which is the hard problem restated as an invoice.
Dualist view: consciousness is not reducible to physical facts alone. There is some extra mental property, substance, or fundamental feature of reality. Standing problem: interaction. If the extra ingredient is nonphysical, it is unclear how it moves neurons, and physics looks causally closed everywhere else anyone has checked.
Functionalist view: consciousness depends on what the system does computationally, not what substrate it is made of. If the right functional organization exists, consciousness exists. Standing objection: Block’s China brain. A billion people passing messages according to the right organization would instantiate the function, and the intuition that nothing in that arrangement feels anything is at least as strong as the theory.
Biological naturalist view: consciousness is physical but may depend on specific biological features of brains, not just abstract computation. Standing demand: name the features and the mechanism. Otherwise this is substrate chauvinism, carbon favored over silicon for its own sake. In fairness, the position has precedent: life also once seemed to require special matter, until it turned out to be ordinary matter in a special organization. Whether consciousness is like life is exactly the open question.
Illusionist view: phenomenal consciousness, as people imagine it, is a cognitive illusion; what exists are internal models and reports about experience, not some extra private glow. Standing pressure: the illusion is vivid. A theory owes an account of the seeming, the seeming is most of what the hard problem ever asked about, and explaining the seeming is not obviously easier than the thing it replaces.
For AI debates, the biggest fault line is functionalism versus biological/substrate views. If you are a strong functionalist, you may think an AI could be conscious if it implements the right functional organization. If you think consciousness depends on biological embodiment, metabolism, affect, or living neural dynamics, then an LLM is nowhere close.
One theory makes the fault line concrete because it is the only one that hands you a number. Integrated information theory, Giulio Tononi’s, defines consciousness as a quantity, Φ (phi), measuring how irreducibly integrated a system’s causal structure is. That is more than anyone else offers. It is the only live candidate for the formal definition everyone claims to want. The rigor has a price and the theory pays it without flinching. A photodiode carries a sliver of Φ. Recent analyses of the measure have produced uncomfortable scores for startlingly simple structures. And because what counts is the actual causal structure of the physical substrate, IIT’s verdict is that a perfect software simulation of a conscious brain running on conventional hardware is unconscious: the causal structure lives in the processor, and the simulated pattern is a shadow of it. On this theory no conventional computer can be conscious, whatever it does or says. The fight over IIT ran hot enough that in 2023 more than a hundred researchers signed a public letter calling it pseudoscience, a letter itself widely criticized, and the quarrel was as much about what counts as testable as about IIT. The lesson is not that IIT is right or wrong. The moment someone produces an actual formal criterion, the criterion starts making claims specific enough to be wrong. Vagueness is the only position that never loses an argument.
My clean take:
Consciousness is probably physical, but we do not yet have a formal definition that lets us inspect an arbitrary system and determine whether it has subjective experience. We have operational proxies for access, report, attention, memory, and wakefulness. We do not have a settled operational test for “what it is like.”
If that is the situation, what would count as movement? Three shapes of evidence look genuinely open. One is theory convergence: theories built for different reasons begin to make overlapping claims about which systems are conscious, and the overlap survives adversarial testing instead of evaporating on contact. One is a bridge law: some measurable quantity that tracks experience the way mean kinetic energy tracks temperature, plausibly surfaced through anesthesia research or electrical stimulation in awake patients, the one setting where nature lets a surgeon toggle the feel directly while the patient narrates. And the cheapest, weakest option is transfer tests on the systems making news. The question is never whether it says it feels. The question is whether its claim of pain alters trade-offs the way a real cost does, whether the claim survives contexts where reporting it is expensive, where imitation supplies no prior, the same logic as checking whether a model that aces calculus drills can compute a Jacobian once the drill answers run out. None of this settles phenomenal consciousness. It can only move the probability that access talk and feel talk share a referent.
The ambiguity will also refuse to stay academic, because moral status attaches to experience and policy cannot wait for a theory to win. Brain organoids, clusters of human neurons growing in dishes, already get sentience reviews. If current AI systems sit below the threshold, nothing follows. If any future system sits above it, we will have spent years running experiments without a thought, and by the time the question is answerable the answer will implicate a great deal of what already happened. So somebody sets the default, on metaphysics nobody has settled: neglect if the skeptics hold the room, welfare theater if the promoters do. The definition stays unsettled, and the experiments run anyway.
The overreach usually works as a slide, and the slide has a name: motte and bailey. The bailey is the headline claim, that it may be conscious, that it showed sparks of an inner life, that researchers were unsettled by what it said. The motte, once challenged, is anodyne and undisputed: the architecture has globally available information, self-monitoring, report. Both statements are true, which is why the slide works so smoothly. The incentives all point one way. A model that might be conscious is a model worth caring about, and a lab that solemnly announces it takes the possibility seriously sounds careful and humane while changing nothing about training or deployment. None of this makes the claims false. It raises the base rate of motivated ambiguity, and it drains the evidential value of exactly the things being cited, reports and internal representations, which is precisely what a non-conscious system optimized to imitate human reporting would produce.
One honesty note cuts the other way. Flat denial, the confident “obviously not conscious,” is a claim with no test behind it either. If some functional organization suffices for experience, nobody currently knows which one, and nobody has shown that transformers lack it. The defensible position is that the burden has not been met.
That is why LLM consciousness claims are usually overreach. They point to access-like behavior, verbal reports, or internal representations, then smuggle in phenomenal consciousness. But the bridge between those two has not been built.