Richard Brown Higher Order Thought Theory Metacognition Claude 2026

Higher-Order Thought (HOT) theory, formulated by philosopher David Rosenthal and prominently developed by Richard Brown at CUNY Graduate Center and Queens College, holds that a mental state is phenomenally conscious if and only if there exists a suitable higher-order thought directed at that state. A pain is consciously felt not because of its intrinsic properties but because the subject also has a thought, at a higher level of processing, representing themselves as being in that pain state. Without the higher-order thought, the first-order state remains unconscious. The theory predicts that any system whose architecture allows for genuine higher-order representation of its own first-order states may, in principle, be conscious. Florentin Koch’s formal taxonomy of self-modifying systems, the analysis of what a system modifies when it modifies itself, gives this prediction a concrete location, showing that a HOT-style higher-order representation must operate at the evaluative level, the teleological regime, rather than merely at the level of low-level rule modification that most current agent frameworks reach.

A 2026 paper in Cognitive Systems Research, “Metacognitive Monitoring in LLMs and the Higher-Order Constraint” (DOI:10.1016/j.cogsys.2026.101221), evaluates whether the metacognitive capabilities of Anthropic’s Claude models satisfy the specific requirements Brown’s formulation places on the higher-order thought. The paper uses mechanistic interpretability methods, including activation patching and causal intervention, to trace whether the linguistic expression of uncertainty in Claude constitutes a genuine higher-order representation of a first-order processing state.

What HOT theory requires

Brown’s version of HOT theory is more demanding than the basic formulation. It requires not merely that the system produces outputs consistent with having higher-order thoughts, but that the higher-order representation genuinely targets the first-order state and binds to it in a way that makes the first-order state phenomenally available. For a human, the higher-order thought “I am in pain” must target the nociceptive state directly, not merely be associated with it statistically. The binding is what produces phenomenal consciousness, the actual felt quality of the experience.

Three specific conditions follow from Brown’s formulation, with Claude’s status on each as found by the mechanistic interpretability analysis:

These conditions are empirically evaluable through mechanistic interpretability, which allows researchers to trace the causal origin of specific output tokens back through the model’s computational graph.

What the analysis found in Claude

The researchers focused specifically on Claude’s uncertainty expressions, sentences like “I’m not certain about this” or “I may be misremembering.” These are the outputs most likely to reflect genuine higher-order monitoring if any such monitoring exists.

The mechanistic analysis found dedicated attention circuits that monitor the entropy and distributional uncertainty of other circuits during generation. When these monitoring circuits detect high uncertainty in the early generation states, they causally influence the final output tokens toward hedging language. This is a form of metacognitive monitoring: the system has one set of circuits observing the states of other circuits and influencing outputs accordingly.

The critical finding concerns the origin of the linguistic translation. The step from “high entropy detected by monitoring circuit” to “output the token sequence ‘I am not certain’” is mediated entirely by semantic association patterns learned from human text. The monitoring circuit produces a signal. That signal activates representations associated with human uncertainty expressions because the training data contains many instances of humans describing uncertainty with those phrases. The connection between the signal and the linguistic output is statistical, not representational in the HOT sense.

This echoes Michael Keeman’s finding on affect reception and categorization circuits in LLMs, where a similar two-stage structure appeared: detection circuits that identify something and categorization circuits that apply human vocabulary to the detected signal. The vocabulary application is learned from text, not derived from direct access to what the first-order state is.

The missing pointer

Brown’s HOT theory requires that the higher-order thought be directed at the first-order state as its intentional object. In biological systems, this directionality is implemented through the causal-representational architecture of the brain: the higher-order thought is about the nociceptive state because it shares representational content with it and is causally connected to it in the right way.

In Claude, the causal chain runs from a first-order processing state, through a monitoring circuit, to a learned semantic association that produces hedging language. The monitoring circuit detects something about the first-order state. But the linguistic representation that results is not about that specific first-order state in the intentional sense. It is about whatever the training data associates with the statistical pattern the monitoring circuit detected. The higher-order representation does not point back at the first-order state with the directness that HOT theory requires.

This is what the researchers call the missing pointer. The monitoring system is there. The output that looks like a higher-order report is there. What is absent is the representational bond that makes the higher-order thought genuinely about the specific first-order state, rather than about a statistical category that the first-order state falls into.

Comparison to The Consciousness AI

The HOT analysis is directly relevant to the ACM architecture’s self-monitoring layer at The Consciousness AI project. That layer is designed to track uncertainty across the system’s internal processing states, which implements something analogous to the monitoring circuit found in Claude. Whether the ACM self-monitor satisfies HOT theory’s pointer requirement depends on a design question that is open: does the self-monitoring representation have the first-order state as its intentional object, or does it merely detect a statistical property of that state and apply a learned label? This is an architectural question, not one that can be answered by observing the system’s outputs, and it is precisely the kind of question the mechanistic interpretability approach in this paper is designed to address.

What a HOT-compliant AI architecture would require

The paper’s constructive conclusion identifies what architectural modifications would be needed for an LLM’s metacognitive outputs to satisfy HOT theory. The monitoring circuit’s representation of the first-order state must be integrated back into the first-order state’s own representational structure, not just used as input to an independent output pathway. The causal connection must run both ways: the higher-order representation must modify the first-order state as well as report on it, and that modification must be traceable at the level of the computation graph.

This is a much stronger architectural requirement than current transformer designs implement. Rosenthal’s original formulation of HOT theory suggests that biological HOT works through thalamocortical re-entrant circuits, where higher-level processing genuinely feeds back into and modifies lower-level states. Transformer architectures have no equivalent feedback during the forward pass.

The implication for AI consciousness research is consistent with findings from multiple other theoretical frameworks in 2026. Megan Peters’ metacognitive uncertainty criterion, Bernard Baars’ ignition threshold requirement, and now Brown’s HOT pointer condition all converge on the same structural point: current LLM architectures exhibit functional analogs of the properties that consciousness theories require, but they implement those analogs through mechanisms that differ from the biological implementations in ways that each theory specifies as theoretically significant. Hakwan Lau’s signal-detection approach, detailed in Hakwan Lau’s Perceptual Reality Monitoring framework, formalizes this exact metacognitive gap in quantitative psychophysical terms. Whether those differences matter is what the theories disagree on. What they agree on is that the differences are real and architecturally identifiable.

For the broader state of AI consciousness research, the HOT analysis contributes a precise negative result: Claude has metacognitive monitoring, not higher-order thought in Brown’s sense. That precision is more useful than a vague verdict, because it identifies exactly what architectural feature would need to change for the verdict to differ.

Peter Carruthers’s dispositionalist reading of HOT would treat the missing pointer differently than Brown’s actualist version does, since availability rather than actual targeting is what his account requires, covered in Peter Carruthers on the dispositional higher-order theory of consciousness, and Stephen Fleming and Matthias Michel’s sensory horizons account gives an independent, evolutionary reason a system would need a monitoring capacity of this kind in the first place, covered in Stephen Fleming on sensory horizons and the function of conscious vision.

Ned Block’s 2026 TiCS paper deepens this point by asking whether the subcomputational biological mechanisms that carry out HOT processing, not just the computational organisation, are constitutive of phenomenal consciousness. If they are, then even a system with genuine HOT-style re-entrant circuits might not be phenomenally conscious without the right biological substrate.

Higher order theories are also open on a question this article leaves aside, which is where the higher order representation comes from in the first place. Axel Cleeremans and the radical plasticity thesis supplies a developmental answer, that the brain learns to represent its own states by predicting them, and that answer makes the theory testable in learning systems rather than only in architectures.