Peter Carruthers and the Dispositional Higher-Order Theory of Consciousness

Higher-order theories of consciousness hold that a mental state is conscious only when the system represents itself as being in that state. The obvious problem is that people are conscious of a great deal they are not, at any given instant, actively thinking about. Peter Carruthers’s dispositionalist answer, laid out with Rocco Gennaro in the January 2026 substantial revision of their Stanford Encyclopedia of Philosophy entry “Higher-Order Theories of Consciousness,” is that availability to higher-order thought is enough. A state does not need to actually be thought about to be conscious, only to be poised for a higher-order thought to occur about it were attention to turn that way.

The Actualist Problem Dispositionalism Solves

David Rosenthal’s original higher-order thought theory, the actualist version, requires that a higher-order thought actually occur, in real time, about a first-order state for that state to count as conscious. Rosenthal’s own account is discussed in an examination of higher-order thought theory in AI architectures. The difficulty is introspective. At any moment a person is plausibly conscious of far more than they are actively generating higher-order thoughts about, the feeling of a chair against their back, background sounds, peripheral vision, all present in experience without necessarily being the target of an explicit meta-representation right now.

Carruthers’s dispositionalist move is to relax the requirement from actual occurrence to availability. A first-order state is conscious, on this account, if the system’s architecture is such that a higher-order thought about it would form given a shift of attention, whether or not that shift happens. This preserves the theory’s core commitment, that consciousness requires higher-order representational access, while avoiding the counterintuitive result that most of ordinary experience is not conscious at any given instant.

What Availability Requires the Architecture to Do

The cost of the dispositionalist move is that “available” needs a non-trivial specification, or the theory collapses into saying almost anything counts as conscious, since nearly any representation is available to some further process under a sufficiently loose reading. Carruthers’s answer ties availability to a specific architectural feature, a global broadcasting or workspace-like mechanism that makes first-order states poised for uptake by a higher-order monitoring system, rather than leaving availability undefined. This is why his account and Global Workspace Theory are frequently discussed together despite starting from different questions, one asking what makes information globally accessible, the other asking what makes a state conscious given that access.

Richard Brown’s application of higher-order thought theory to Claude’s metacognitive reports, covered in Richard Brown on higher-order thought theory and metacognition in Claude, and Hakwan Lau’s perceptual reality monitoring account, covered in Hakwan Lau on perceptual reality monitoring as a higher-order test for AI, both work with versions of this same architectural requirement, that an AI system claiming higher-order-theory-relevant consciousness needs a demonstrable, not merely inferred, capacity for its first-order representations to be taken up by a genuine monitoring process. Stephen Fleming and Matthias Michel’s sensory horizons account, covered in Stephen Fleming on sensory horizons and the function of conscious vision, supplies an evolutionary reason such a monitoring capacity would exist at all, tying it to reality monitoring rather than treating higher-order access as a brute architectural fact.

Testing Availability, Not Just Behavior

The dispositionalist framing raises the evidentiary bar for AI systems specifically because it cannot be satisfied by demonstrating that a system sometimes produces accurate self-reports. A system could produce correct-seeming introspective reports through a narrow, task-specific mechanism that has no general availability property, reports generated post hoc by a separate process rather than reflecting a standing disposition for any first-order state to be taken up by monitoring. Distinguishing genuine dispositional availability from a system merely trained to output plausible self-reports on demand requires probing the architecture directly, checking whether arbitrary first-order representations, not just the ones a training process happened to reward reporting on, are in fact poised for higher-order uptake.

Comparison to The Consciousness AI

The Global Workspace layer in this project’s architecture, described on the architecture page, broadcasts winning coalitions back to the same specialist modules that submitted bids, across five to ten reentrant convergence cycles. That broadcast structure is a plausible candidate for the kind of availability Carruthers’s dispositionalism requires, since content that ignites becomes accessible to every module rather than staying local. What the architecture does not currently include is a distinct higher-order monitoring process that takes broadcast content as its object and forms a representation about it being a state the system is in. The reentrant loop makes content globally available; it does not by itself constitute a higher-order thought about that availability. Closing that gap, adding a module whose function is specifically to represent the system’s own first-order states as such, is a distinct architectural addition this project has not built, and Carruthers’s account is precise about why availability alone, without that further representational step, would not be sufficient on his own theory.

An Unsettled Boundary

Carruthers and Gennaro’s revised entry does not resolve the deepest disagreement inside higher-order theories, whether the monitoring mechanism is a form of higher-order thought, a conceptual, language-like representation, or a form of higher-order perception, a quasi-sensory scanning process with no clear conceptual content. That distinction matters for AI evaluation because the two versions imply different architectural requirements, a genuine perception-like monitoring channel versus a thought-like representational capacity that could in principle be implemented more simply. What dispositionalism settles is narrower but still consequential, that the test for AI consciousness under this family of theories cannot be satisfied by episodic self-report alone, and must instead show that the system’s first-order states stand in a general, structural relationship of availability to whatever monitoring mechanism the theory ultimately specifies. The wider set of competing accounts is indexed on the state of the field on AI consciousness.