KurisutinaCitation: Seth, A. K., and Bayne, T. Theories of consciousness. Nature Reviews Neuroscience 23, 439–452. Published 3 May 2022. DOI: 10.1038/s41583-022-00587-4.
Complete reading and version
Read on 29 September 2026: the complete 37-page accepted manuscript, downloaded from the authors' University of Sussex repository, file SethBayne_NRN_accepted.pdf. Its PDF creation metadata is 24 March 2022. The repository labels it an accepted version; it is not silently treated as the publisher's typeset version. The publisher's public record confirms the citation and publication date but restricts the full article text. No separate scientific supplement is listed on the publisher record or among the repository's files.
All 1,814 extracted text lines were read: abstract, every main section, acknowledgements, declarations, glossary, the complete two-page Table 1, four figures and captions, Boxes 1–3, and all 178 bibliographic entries and annotations. Figures 1–4 and both table pages were additionally inspected visually on PDF pages 19–24. Bibliography entries were read as references; their underlying papers were not opened or counted as independent evidence. There is no separate empirical Methods section, dataset or original experiment to reproduce in this narrative review.
Source files and hashes are in papers/consciousness_connectomics/seth_bayne2022_theories.provenance.json. This reading establishes coverage of the complete accepted manuscript, not a line-by-line comparison with the inaccessible final full text. It is a 2022 theory map, preceding the 2025 Cogitate results already summarized in this library.
Question and type of contribution
Why do consciousness theories proliferate without experiments clearly distinguishing them? The authors argue that theories often differ in their explanatory targets, level of description and assumptions about how consciousness can be measured. They compare four broad families: higher-order accounts, global workspace accounts, integrated information theory, and reentry/predictive-processing approaches.
This is conceptual synthesis with selected empirical examples. It is not a systematic review with an exhaustive search, a meta-analysis, a new test of the theories, or a ranking by pooled evidential strength. Its usefulness here is specifying what an experiment would have to discriminate. The authors' descriptions of cited empirical findings remain secondary evidence unless the corresponding original paper has separately been fully read.
The target needs to be specified before theories can be compared
The review distinguishes three questions that are easy to merge:
- Global state and local content. Wakefulness, dreaming and sedation concern an organism's overall conscious condition. Seeing a face, feeling pain or entertaining a thought concern particular experiences. The authors prefer a potentially multidimensional description of global states over assuming that every state lies on one ordered consciousness scale.
- Experiential character and functional role. What an experience feels like differs as an explanatory target from how information guides flexible action, memory or report. Distinguishing the targets does not prove that their underlying mechanisms are independent.
- Why this content is conscious and why it feels this way. Explaining why a house rather than a face dominates binocular rivalry is different from explaining the experiential difference between seeing a house and hearing a sound.
Selfhood includes several potentially separable contents: bodily ownership, volition, affect, autobiographical experience and others. Merely labeling a model a model of selfhood leaves these targets unspecified. The review also discusses how local contents form a unified experience, while acknowledging disagreement about the precise nature and necessity of that unity.
The theory map
| Approach | Central proposal as presented in the review | Distinction or unresolved commitment |
|---|---|---|
| Higher-order theories | A mental representation becomes conscious through an appropriate representation of that representation | Being later or deeper in a processing hierarchy does not by itself constitute a meta-representation. Variants disagree about its form and neural implementation. |
| Global workspace theories | Information becomes conscious through availability to many consuming processes; neuronal versions emphasize recurrent ignition and widespread broadcasting | What makes a workspace sufficiently global needs specification. Explaining broad access is not automatically an explanation of differences in experiential character. |
| Integrated information theory | Experience is identified with a physical system's irreducible intrinsic cause-and-effect structure | This is a claim about causal organization in a state, not merely observed activity, external decoding or a large connectivity statistic. Exact quantities are infeasible for realistic brains. |
| Local reentry/recurrency accounts | Recurrent processing within perceptual systems can support experience, conditional on enabling processes such as arousal | Reporting and executive use may require additional regions. The relevant form of recurrence must be distinguished from recurrence occurring without consciousness. |
| Predictive-processing approaches | Perceptual content is related to inferred causes of sensory input, commonly represented through interacting predictions and prediction errors | This is a broad account of brain function, with several proposed connections to consciousness rather than a single agreed consciousness theory. |
Higher-order accounts need not require a conscious confidence judgment accompanying every conscious perception: their relevant higher-order state can itself be unconscious. Global workspace access and meta-representation are therefore not interchangeable mechanisms merely because both accounts often emphasize prefrontal cortex.
The IIT discussion associates the quantity of integrated information with global state and the structure of intrinsic causal relations with experiential character. It explicitly separates those theoretical quantities from practical proxies such as perturbational complexity. Success of a proxy does not uniquely support IIT when other theories also accommodate the observations.
Predictive-processing variants can address properties such as perceptual presence, emotion and bodily selfhood; active-inference variants connect perception with action and physiological regulation. The review does not establish that any predictive model is conscious. Different proposals still need to identify which predictive operations explain the presence of experience rather than ordinary unconscious processing.
Table 1 and Box 2 widen the landscape to attention, learning, affect, bodily regulation, sensorimotor, cellular and other accounts. This breadth prevents a two-theory contest from being mistaken for an exhaustive contest. It does not provide equal evidential assessment of every listed theory.
What would discriminate the accounts?
A shared observation can have several explanations. The review uses the cerebellum as an example: different theories can each offer reasons why its architecture would not supply their proposed consciousness mechanism. Compatibility with this observation alone cannot select the best account. These are explanations discussed by the authors, not new lesion or intervention results in this article.
Anatomical labels are weaker than specified processes. The frontal-versus-posterior debate mixes neural evidence with assumptions about report and cognitive access. Decoding stimulus information from a region does not show that the brain uses that region to make a meta-representation or broadcast the information. Conversely, diminished activity in a no-report experiment does not automatically identify which part of the difference reflects consciousness, reporting, attention or task engagement.
Timing requires a defined observable. Earlier sensory negativities and later ignition/access signatures can motivate different accounts, but the time at which consciousness begins is distinct from perceived duration or the duration of a subjective moment. The reviewed timing evidence is selective and contested, not a settled universal latency.
Novel predictions are especially valuable when competitors differ. One distinctive IIT-related proposal discussed is that changing a system's causal possibilities could change experience without changing its observed activity trajectory—for example, an inactive neuron and a neuron rendered incapable of activation need not have the same causal role. The review proposes that precise interventions might test such claims. It does not report that this prediction was experimentally confirmed or supply a generally accepted measurement of the accompanying experience.
Theory comparison must specify the competing predictions and auxiliary assumptions. A result that fits one theory deserves more weight when alternatives predicted a different result, rather than when they can all accommodate it after the fact. The authors favor cumulative evaluation across diverse observations rather than declaring a winner from one supportive finding.
The measurement problem
Reports provide essential evidence about experience but depend on introspection, memory and the ability to communicate. Removing overt reports addresses some of those requirements while introducing a validation question for the replacement marker. An eye movement or neural statistic still needs evidence of sensitivity and specificity for the claimed target.
The review separates detecting a particular conscious content from identifying whether a creature or system has the capacity for experience. A marker validated for perceptual transitions in communicating adults does not automatically answer whether an infant, injured patient, nonhuman animal, organoid or artificial system is conscious.
There is a potential circularity: a contentious theory should not validate its own preferred marker and then treat that marker as independent confirmation of the theory. The authors argue for validation drawing on more than either introspective reports alone or theoretical commitments alone, while acknowledging that no uncontested solution is available. This is a methodological challenge, not a declaration that all existing measures are worthless.
Authors' proposed route forward
The authors call for greater precision, coverage and measurement validity. Computational models can force broad ideas into explicit mechanisms and predictions, while clarifying whether apparently rival accounts share processes. Models must still be connected to appropriate phenomenological evidence; explaining a functional task cannot silently replace the intended experiential target.
Coverage should extend beyond visual perception in awake human adults to affect, thought, volition, dreaming, altered states, development and other organisms. The review also discusses different attitudes toward the philosophical explanatory gap: attempting to solve it, explaining specific experiential properties, or questioning the intuitions that motivate it. It does not resolve that debate empirically.
Critical assessment and implications for this project
The review supplies a useful comparison framework, not a canonical final taxonomy. Theory families have internal disagreements, and the evidence selection is narrative. Broad claims about a family must therefore be checked against a specific formulation before designing a discriminating test. Its 2022 assessments cannot stand in for a current survey of later evidence.
For the present research, the following are our inferences:
- A connectome-based model needs a declared target: a sensory computation, a state transition, access to information, experiential content or a claim about a subject. A score on one target should not be relabeled as another.
- Tests should compare explicitly implemented mechanisms, including their intervention predictions, rather than equating a brain region or network diagram with a theory.
- Matching recorded activity does not necessarily identify all causal dispositions of the source circuit. The review's inactive-versus-inactivated example makes that conceptual distinction especially clear; it does not prove which unobserved dispositions matter in a real reconstruction.
- Model selection and consciousness-marker validation are separate problems. A model may predict a marker well while the marker's relationship to experience remains uncertain outside its original population and task.
- The Cogitate study provides a later, concrete attempt to operationalize selected theory commitments. Its results should be interpreted at that level, with this review's broader landscape in view.
Nothing in this review establishes that a connectome reconstruction, predictive model or language-model persona has experience. Nothing establishes the impossibility of artificial consciousness or resolves personal continuity through copying. Its contribution is a more precise account of what those claims would need to explain and how their evidence could be distinguished.