A third major theory of consciousness — distinct from both Integrated Information Theory (E.26) and Global Workspace Theory (E.27) — locates consciousness in what the mind does about its own states.
The position is higher-order theory. Its leading articulator is David Rosenthal, whose 2005 book Consciousness and Mind is the canonical statement.1 The basic claim, in one sentence: a mental state is conscious when, and only when, the mind has another mental state about it — a higher-order representation that the first state is occurring.
This sounds strange. It works as follows.
The proposal
A first-order mental state is about something in the world: the perception of red, the desire for coffee, the belief that it is raining. A higher-order mental state is about a first-order mental state: the representation that one is perceiving red, the awareness that one wants coffee, the noticing that one believes it is raining.
On Rosenthal’s view, a first-order state is unconscious by default. It becomes conscious only when a higher-order state represents it. Conscious experience is, structurally, the representation of one’s own mental activity to oneself.
The view explains some otherwise puzzling phenomena. Subliminal perception: information enters the system, influences behavior, but is not consciously experienced because no higher-order state forms. The phenomenology of “noticing”: the sudden conscious access to something that was already in the mind. The structure of introspection: it is exactly the higher-order representation, made explicit.
What this implies for AI
Higher-order theory has a clear implication for AI systems: consciousness requires self-modeling. A system that represents its own internal states — that has, in some functional sense, beliefs about its beliefs, awareness of its awareness — is, on this theory, a candidate for consciousness in a way that a system with no self-modeling is not.
This matters because recent AI systems are building toward self-modeling. Reflective LLMs that critique their own outputs, agent architectures with metacognitive layers, models that report uncertainty about their reasoning — these are all systems with at least proto-higher-order structure. The theory does not say they are conscious; it says they have one of the structural features it considers necessary for consciousness.
The trouble — and the deep trouble for higher-order theories generally — is that trained linguistic competence makes self-modeling cheap. An LLM can produce fluent first-person reports of its internal states without having those states. The model is trained on a corpus saturated with human reports of mental life; producing such reports is, for the model, ordinary text completion. The reports are not evidence of higher-order representation in any deep sense.
Three problems
Higher-order theories face several philosophical objections that are worth knowing.
The infinite regress. If a state is conscious because a higher-order state represents it, what makes the higher-order state conscious? If it needs a third-order state, the regress goes upward indefinitely. Most defenders argue the higher-order state need not itself be conscious — the relation is asymmetric — but this answer feels unsatisfying to many.
The mismatch problem. A higher-order state can misrepresent the first-order state it is about. Rosenthal accepts this; in such cases, on his theory, the misrepresented version is what is consciously experienced. Critics find this counterintuitive — am I really wrong about my own pain? — but the theory bites the bullet.
The training problem (the AI version). A system can produce higher-order reports (text describing its internal states) without having higher-order representations (actual second-order mental states). The theory needs a way to distinguish, and current versions don’t have a clean one. Telling a real higher-order representation from a learned first-person verbal style is, in present LLMs, very hard.
Why this matters
If higher-order theories are right, then AI consciousness is closer than GWT or IIT would suggest, because recent AI architectures already include features that look like higher-order representation. The labs are building toward self-modeling for instrumental reasons — reflective agents are more reliable, calibrated outputs are more useful — and may inadvertently be building toward consciousness as a side effect.
If higher-order theories are wrong, the resemblance between AI self-reports and conscious self-knowledge is misleading. The systems are producing higher-order outputs without having higher-order states; the correlation is in the training data, not in the architecture.
The encyclopedia’s stance: the question is open, the answer matters, and the current methods of telling cannot settle it. The next article — Tests for Machine Consciousness (E.29) — takes up the question of what, in principle, would settle it.
Footnotes
-
Rosenthal, 2005. ↩