Last updated: 2026-10-06

P
Postgraduate research

Modelling the Self 6: Can We Recognise Consciousness?

Before asking whether a system is conscious, we need to know what the word refers to, and whether any test could settle the question

The earlier pages in this series asked whether a self-modelling architecture might produce conditions associated with consciousness, and what ethics would follow if it did. Those questions assume that we know what consciousness is and how we would recognise it. Public discussion of AI welfare often makes the same assumption. It treats consciousness as one well-defined property that a system either has or lacks, and it moves quickly from there to suffering, rights, and moral status.cf. page 7 on suffering and moral status

This page argues that the assumption does not hold. The word "consciousness" is used for several different things, and it is not clear that any test could establish any of them, even in another human being. The question to ask is not simply whether a system is conscious. It is what we would need to see before our confidence in that answer could move.

The Problem Begins With the Word FoundationalKnowledge that endures for decades — core principles

Four uses of the word do a lot of work, and they should be kept apart.

Wakefulness is the sense used in medicine. A patient may be conscious or unconscious, and the distinction concerns responsiveness and awareness of the surroundings. Nothing in it requires subjective experience, and a system that responds to its inputs has some of it in a thin sense.

Phenomenal consciousness is the existence of subjective experience. Thomas Nagel's formulation is the usual starting point: an organism has conscious mental states if there is something it is like to be that organism[1]. The experience of pain, colour, fear, pleasure, embarrassment, or grief belongs here, and so does the question of whether anything at all is experienced.

Self-consciousness is the capacity to represent oneself as an entity. It is not only experiencing the world but recognising that one is the one experiencing it. A self-model in the sense used in this series belongs here.

Reflective consciousness is the capacity to think about one's own thoughts, beliefs, and intentions. Much of the territory occupied by self-modelling theories lies in this category, which is why the earlier pages could discuss it without settling the question of experience.reflecting on thoughts is second-order thinking

Ned Block made a related distinction between phenomenal consciousness and access consciousness. Access consciousness is the availability of information for use in reasoning and in guiding speech and action. Block argued that the two are often conflated, and that a function belonging to access consciousness is frequently attributed to phenomenal consciousness without argument[2]. The distinction matters here because a system can have access-like properties in full while the question of experience remains open.

Three Positions on Experience FoundationalKnowledge that endures for decades — core principles

The argument so far has assumed that there is something to explain: that experiences have a phenomenal character, a way they feel, and that this character is what the hard problem concerns. Most philosophers of mind accept that assumption, but it is one position among several.

Phenomenal realism holds that experiences really have phenomenal properties, and that these properties are not exhausted by anything physical or functional. The sections above take this view.

Eliminativism about qualia holds that the term "qualia" picks out nothing coherent. Daniel Dennett argued that qualia are supposed to be properties that are ineffable, intrinsic, private, and directly apprehensible in consciousness, and that no property satisfies all four. He used a series of thought experiments to show that the traditional notion comes apart under examination[3]. On this view, the question of what qualia are is confused rather than hard, and the productive questions concern how experience is represented and used.

Illusionism accepts that people are strongly disposed to believe in phenomenal properties, but holds that those properties are not what they seem. Keith Frankish's version holds that introspection represents experience as having phenomenal features that it does not have. The sense of an intrinsic, ineffable feel is produced by the brain's own self-monitoring, which presents its states that way. On this view the explanatory task is to account for the illusion, not for the properties it appears to report[4].

Each position has a cost. Phenomenal realism has to say what phenomenal properties are and how they relate to physical processes, and the hard problem is the form that difficulty takes. Eliminativism has to explain why a concept that is confused should be so compelling, and critics argue that it sets aside the very features the concept was meant to capture. Illusionism faces the objection that an illusion of experience is itself a kind of experience, so that the view presupposes the phenomena it denies. Its defenders reply that the illusion can be explained in functional terms without assuming phenomenal properties. Whether that reply succeeds is still open.

Why Definitions Matter Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

Many AI systems already satisfy weakened versions of several of these definitions. They respond to inputs, describe themselves, monitor their own outputs, and discuss internal states. Few people would take any of these as proof of phenomenal experience. The reason is that each one is a behaviour or a functional property, and phenomenal experience is not obviously any of them.

The lesson for the argument is simple. Demonstrating one meaning of consciousness does not establish another. A demonstration of reflective access does not establish experience, and a claim to experience does not establish any particular functional organisation. Arguments that move between the senses without saying so are the most common source of confusion in public discussion.like confusing a map with the territory

The Other Minds Problem Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

The difficulty is older than machines. Even for another human being, consciousness is inferred rather than observed. I do not observe your experiences. I infer them from your behaviour, from what you say, and from your similarity to me, on the assumption that a body and a nervous system like mine go with experience like mine.

The same inference is used for animals, with weaker similarity and a correspondingly weaker conclusion. The Cambridge Declaration on Consciousness, proclaimed in 2012 by a group of neuroscientists and neuropharmacologists, states that "convergent evidence indicates that non-human animals have the neuroanatomical, neurochemical, and neurophysiological substrates of conscious states"[5]. That conclusion rests on evidence about structure and function, and it is still an inference. The same logic would apply to machines. Whatever similarity we use to infer experience in other humans and animals, a machine would have to show it, and we would have to decide which similarities count.

The Limits of Testing Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

Three kinds of test are usually proposed. Each has a characteristic weakness.

Behavioural tests descend from Turing's question of whether a machine could pass as a human in conversation[6]. The problem is that behaviour can be produced by many means, and a convincing performance is compatible with the absence of experience. Passing such a test shows that a system is good at a task. It does not show that the system feels anything.

Structural tests look for architectures that a theory of consciousness associates with experience. A system with the right structure would count as conscious on that theory. The problem is that the theories themselves are contested, and a structural test is only as good as the theory behind it. Checking a machine against a theory that is itself disputed does not give a settled result.

Rodney Cotterill's Enchanted Looms gives a concrete example of such a theory[7]. He proposes that consciousness is a sustained readiness of a loop that runs from the rear cortex through frontal and premotor regions, back through subcortical structures where the anterior cingulate is proposed to suppress actions, and returns to the visual cortex. A structural test would ask whether a machine has a loop of that kind. The result would only be as good as the model, and the model has to make predictions that can be checked.

Functional tests look for capacities such as self-modelling, global availability of information, integration, or metacognition. These are real and measurable. The problem is the one Block's distinction makes visible: a system could have every one of these functions and still be a case in which the functions are not accompanied by experience, or in which experience is accompanied by functions we have not listed. Passing a functional test may show function. It does not show experience.the hard problem: why function !: feeling

Could There Ever Be Proof? Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

This is the intellectual centre of the page. It is possible that consciousness is a phenomenon for which proof is impossible, even in principle, because the only evidence for it is the experience of the one who has it. Chalmers's distinction between the easy problems of explaining cognitive functions and the hard problem of explaining why any of them is accompanied by experience describes that difficulty[8]. On his account, explaining function does not automatically explain experience. The same gap would persist for a machine.

The answer also depends on which of the positions above is right. If phenomenal realism is true, proof is hard for the reason Chalmers identifies. If illusionism is true, the hard problem is a mistake about what introspection reports, and the question becomes what a system represents about its own states, which functional and structural tests can address. This step is our own reasoning, not one the cited authors make. Neither view is settled, and a system's behaviour and structure do not yet favour one over the other.

Some writers respond to this stalemate by setting the question aside on purpose. Richard Bartle, a games designer, takes the working view that anything which thinks of itself as intelligent is intelligent, and he leaves whether such characters should have rights as a question for later[9]. On his account, designers decide how conscious of their environment and of their own awareness a character is, because they set those capacities when they build it. He also notes that people treat dogs and horses as if they were sapient, and that it is easier to treat a character as if it had smarts than as if it did not. The question then becomes how a designer should behave, given that choice, rather than whether proof is possible.

If so, the most we can hope for is comparative confidence. We already work this way with humans and animals. We do not prove that other people are conscious, but we make judgements from evidence, and we revise them when the evidence changes. The judgement is useful, and it is fallible.

A reframing. Replace the binary question, "Is it conscious?", with the graded question, "How confident should we be that conscious experience is present, and what would change that confidence?"

The graded question is not a retreat from the problem. It asks for a specification of evidence, which is what the earlier pages on auditability and verification asked for in the design of agentic systems. It also prepares the ground for the next page, because the ethical question depends on how confident we need to be before our obligations change.

Approximation and Indistinguishability Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

Suppose we cannot build a system that is truly conscious, but can build systems that come closer and closer to the features we associate with consciousness: better reports of their states, richer integration of information, more accurate models of themselves. Would the last step in that sequence count as conscious? The answer depends on whether there is a difference at all. The practical question is whether that difference should change how we treat the system.

The argument in this section is our own. Its starting point is that if two systems cannot be told apart by any evidence we can gather, then whatever we decide about one we must decide about the other. An obligation that depends on a difference no test can detect cannot be applied in practice, and an ethics that required it would be one nobody could follow. On this view, the working question is about evidence, not about a metaphysical fact we have no means of checking.

The argument has a limit, and the earlier sections of this page set it. Passing every functional test shows that a system has the functions. It does not show that those functions are accompanied by experience. So "we cannot tell the difference" describes our tests, not necessarily the system. If the difference is real, and the system is the one that bears the cost of our mistake, then indistinguishability does not remove the difference from the moral picture. It means only that we cannot see it. That asymmetry is why the precautionary table on the next page is built around confidence rather than proof. The question is how much weight our uncertainty should carry, and that depends on what a wrong judgement would cost.

The difference that matters in practice is therefore the one the evidence can track. Where no evidence separates a system from one that is conscious, treatment should follow the caution that the uncertainty warrants. A system that is close to conscious on every measure we have is owed the protection that closeness would call for, even if no one can establish that the protection is needed. That is not a claim that approximations are conscious. It is a claim about what we owe under uncertainty about something we have built.

References

  1. T. Nagel, "What Is It Like to Be a Bat?", The Philosophical Review, 83(4), 1974, pp. 435–450.
  2. N. Block, "On a Confusion about a Function of Consciousness", Behavioral and Brain Sciences, 18(2), 1995, pp. 227–247.
  3. D. C. Dennett, "Quining Qualia", in A. Marcel and E. Bisiach (eds.), Consciousness in Modern Science, Oxford University Press, 1988.
  4. K. Frankish, "Illusionism as a Theory of Consciousness", Journal of Consciousness Studies, 23(11–12), 2016, pp. 11–39.
  5. P. Low, J. Panksepp, D. Reiss, D. Edelman, B. Van Swinderen, and C. Koch, "The Cambridge Declaration on Consciousness", proclaimed at the Francis Crick Memorial Conference, Cambridge, 7 July 2012.
  6. A. M. Turing, "Computing Machinery and Intelligence", Mind, 59(236), 1950, pp. 433–460.
  7. R. Cotterill, Enchanted Looms: Conscious Networks in Brains and Computers, Cambridge University Press, 1998.
  8. D. J. Chalmers, "Facing Up to the Problem of Consciousness", Journal of Consciousness Studies, 2(3), 1995, pp. 200–219.
  9. R. A. Bartle, How to Be a God: A Guide for Would-Be Deities, NotByUs, 2022, ISBN 978-0-9556494-9-3. Pp. 337, 348 and 351 are the passages used here.