Empirical Theories of Consciousness

Empirical theories of consciousness propose organisational or neural conditions under which information becomes conscious. They differ over the target as well as the mechanism. Some primarily explain global access and report, some phenomenal character, and some a system's model of its own awareness. A theory can predict experiments well while leaving the hard problem open; conversely, a metaphysical view can preserve phenomenal reality while making few discriminating predictions.

The main families are higher-order, global-workspace, recurrent-processing, integrated- information, predictive, and attention-schema theories. None should be inferred from a neural correlate alone. The relevant question is whether its proposed mechanism, target construct, and distinctive predictions survive comparison with rivals.

A common comparison frame

Each theory should answer five questions:

1. Target: access, phenomenality, report, self-consciousness, or some combination?
2. Mechanism: what operation distinguishes conscious from unconscious processing?
3. Evidence: which intervention or dissociation supports that operation specifically?
4. Scope: what follows for animals, infants, patients, and artificial systems?
5. Contrast: what observation would favour it over a nearby rival?

Shared prediction is not independent confirmation. A late neural signal may reflect global broadcast, higher-order representation, working memory, decision, or report preparation. Experiments must isolate the process rather than attach a preferred label to the same contrast.

Higher-order theories

Higher-order theories hold that a mental state is conscious when the subject is in an appropriate further state representing themselves as being in it. On higher-order thought accounts the further state is thought-like; higher-order perception accounts model it as an inner monitoring representation.

The theory explains the difference between unconscious and conscious first-order states without changing their contents. It also connects consciousness with introspective availability. Misrepresentation creates distinctive possibilities: a higher-order state might represent a first-order state that is absent, producing a conscious appearance without the represented target.

Evidence is sought in metacognition and prefrontal monitoring. The interpretation is contested because these processes also support report, confidence, and task control. No-report paradigms test whether phenomenal experience survives when higher-order and report-related activity is reduced.

The view permits animal or machine consciousness when the required meta-representation exists, but the richness demanded varies by version. Requiring conceptual self-thought is restrictive; a non-conceptual monitoring state is more permissive and harder to distinguish from ordinary recurrent processing.

Global workspace theories

Global workspace theory treats the mind as specialised processes competing for access to a limited-capacity workspace. A winning representation is broadcast widely to memory, reasoning, evaluation, and action systems. Global neuronal workspace theory identifies candidate long-range neural dynamics, including recurrent amplification or "ignition."

The framework explains access consciousness directly: global availability is the target. Masking, report, working-memory access, and broad integration provide evidence, though each can be confounded by task demands.

Its relation to phenomenality is the pressure point. One may identify phenomenal consciousness with broadcast, hold that broadcast normally accompanies but does not constitute it, or use the workspace only as a theory of access. Evidence for global availability cannot choose among those readings without a bridge premise.

Artificial systems could qualify if they contain genuine competition, limited access, and broadcast to otherwise specialised processes. A shared data channel is merely workspace-shaped; architecture and counterfactual control matter more than the metaphor.

Recurrent processing theory

Recurrent processing theory locates visual consciousness in feedback within sensory cortices rather than in later global access. A feedforward sweep can support rapid categorisation without awareness; recurrent interactions integrate and stabilise the perceptual representation.

The theory predicts phenomenal overflow: local conscious perception can occur before or without full access, report, and working memory. It interprets early recurrent signals as constitutive and later frontoparietal activity as post-perceptual use.

Masking and timing evidence bear on the distinction, but recurrent processing is common throughout brains and performs many unconscious functions. The theory must specify which recurrence is sufficient and why. It is comparatively permissive about animals with recurrent sensory systems and restrictive about purely feedforward artificial architectures.

The sharp contrast with workspace theory concerns whether local recurrence without global broadcast is conscious. Experiments that remove report demands and independently measure local perceptual content are therefore especially important.

Integrated information theory

Integrated information theory begins from proposed axioms of experience - intrinsic existence, composition, information, integration, and exclusion - and derives requirements on a physical system's intrinsic cause-effect structure. Consciousness is identified with a maximally irreducible conceptual structure; its amount is associated with integrated information, commonly denoted $\Phi$.

The theory aims at phenomenality rather than report and assigns consciousness by causal organisation. This yields bold consequences: some simple recurrent systems may qualify, while systems decomposable into feedforward parts may have little or no integrated information despite sophisticated behaviour.

The precision is an advantage only if the quantities are tractable and the axioms independently justified. Exact calculation scales poorly, practical proxies may not measure the theoretical quantity, and critics argue that the theory attributes experience too liberally or generates counterintuitive rankings. Its identification of experience with causal structure also moves from phenomenological axioms to ontology through contested bridge principles.

IIT and workspace theories can agree that integration matters while disagreeing dramatically on why and on which systems qualify. Tests must target irreducibility or broadcast specifically, not generic complexity.

Predictive approaches

Predictive processing models perception and action through hierarchical generative models that minimise prediction error, with precision weighting controlling the impact of signals. Conscious content has been associated with the winning hypothesis, high- level inference, precision, or the availability of prediction errors.

This is a broad computational framework rather than one settled theory of consciousness. It explains perception, expectation, attention-like selection, and some hallucinations, but those functions can be unconscious. A consciousness theory requires an additional claim about which predictive process is sufficient.

Predictive approaches are compatible with workspace, higher-order, recurrent, and representational theories. Their flexibility is productive but creates a falsifiability risk: almost any result can be narrated as altered priors, likelihoods, or precision. Discriminating predictions require independent specification of the hierarchy and its parameters.

Attention schema theory

Attention schema theory proposes that brains construct a simplified model of their own attention for control and social prediction. The model omits mechanistic detail and represents attention as an apparently non-physical awareness. Subjects consequently report possessing an ineffable inner property.

The theory directly addresses the meta-problem: why systems claim to be conscious and find consciousness mysterious. It predicts links among attention control, self-model, social cognition, and awareness reports.

Attention and consciousness nevertheless dissociate, and a model of attention may explain the concept of awareness without explaining phenomenal character. Attention is also a technical family of selection mechanisms, not whatever a machine-learning architecture calls "attention." Artificial implementation requires a control-oriented self-model, not a weighted retrieval operation bearing the same name.

Comparison

The theories need not be exclusive. Local recurrence may stabilise a representation, workspace broadcast may make it accessible, and higher-order monitoring may enable confidence. A hybrid can be explanatory, but adding every supported mechanism risks losing a claim about which one is constitutive of consciousness.

Correlate, cause, prerequisite, or constitution?

For any observed mechanism $N$ and conscious state $C$, at least four relations remain:

$$N\text{ correlates with }C,N\to C,N\text{ enables }C,N\text{ constitutes }C.$$

Intervention can strengthen a causal claim; selective dissociation can distinguish prerequisites from consequences. Constitution is a further metaphysical relation and is not read directly from timing or location. The general safeguards are developed in Method and Evidence.

Machine and animal implications

Theories disagree about non-human minds because they privilege different organisation. Workspace and higher-order views are relatively substrate-neutral; recurrent theories require feedback architecture; IIT uses intrinsic causal structure and may reject functionally equivalent feedforward simulations; biological versions may add evolved or homeostatic constraints.

This divergence is informative. Behaviour cannot settle the question when theories explicitly disagree about whether behaviourally equivalent organisations are conscious. The proper response is not to count votes but to compare the independent evidence and theory-specific predictions. The existing LLM case study shows how sharply the criteria can diverge when applied to one architecture.

Assessment

Empirical theories have transformed consciousness from one undifferentiated mystery into competing claims about access, recurrence, integration, monitoring, and self-models. Their major methodological problem is adjudication: shared tasks and neural contrasts often support several theories because report, attention, memory, and awareness covary.

Progress requires preregistered contrasts, interventions, theory-derived quantitative predictions, and tests on systems where the mechanisms dissociate. Even decisive mechanistic evidence would leave a philosophical question about why that mechanism is phenomenal. It would nevertheless make the question far better specified.

The experimental logic behind those comparisons is developed in The Science of Consciousness.

Selected references

• Baars, Bernard J. A Cognitive Theory of Consciousness (1988).
• Dehaene, Stanislas. Consciousness and the Brain (2014).
• Graziano, Michael S. A. Consciousness and the Social Brain (2013).
• Lamme, Victor A. F. "Towards a True Neural Stance on Consciousness" (2006).
• Rosenthal, David M. Consciousness and Mind (2005).
• Tononi, Giulio et al. "Integrated Information Theory (IIT) 4.0" (2023).