Contribution
The review sorts the main neurobiological theories of consciousness by what they aim to explain, so that it is clear where they are rivals and where they are answering different questions. It sets out the evidence each cites, the empirical debates that could separate them, and three things theories must acquire before they can be tested properly: precision, comprehensiveness, and a solution to the measurement problem.
Key insight
Theories that look like rivals often have different targets: some explain why a content is conscious, some explain what it is like, some explain global states, some explain function. Where they do compete, the evidence is filtered through report, which is itself one of the things in dispute. So the main empirical debates are as much about what counts as evidence as about the brain.
Assumptions
- Only theories expressed in, or expressible in, neurobiological terms and consistent with known physics; quantum theories are set aside.
- The phenomenal/functional distinction is used as a distinction of explanatory targets, not a claim that the two are independent (“they are very likely not independent”).
- Global versus local states, and why a state is conscious versus why it has its character, are the axes of comparison.
Key results
As a review, its results are classifications and arguments.
- Targets (Preliminaries). Global states (wakefulness, dreaming, sedation; “regions within a multi-dimensional space”, not levels); local states or contents; phenomenal versus functional properties; why a content is conscious versus why it has the character it has.
- Higher-order theories. A state is conscious by being the target of a suitable meta-representation. They mainly explain why some contents are conscious; character is fixed by what the meta-representation ascribes. Anterior, especially prefrontal, emphasis.
- Global workspace theories. Consciousness is wide availability to consumer systems through ignition and broadcast in a frontoparietal workspace. Explicitly theories of “conscious access”; “rarely attempted” phenomenal character. Open: what makes a workspace global.
- Integrated information theory. Axioms about experience yield requirements on a substrate; consciousness is the cause–effect structure of a maximum of irreducible integrated information (Φ). Posterior “hot zone”. The most comprehensive in coverage, but implies non-biological systems “already exist” that are conscious, and Φ is “infeasible to measure, except in simple model systems”. PCI is a proxy, “not equivalent to Φ”, and its success is compatible with other theories.
- Reentry and predictive processing. Consciousness with top-down signalling; local recurrency in perceptual cortex suffices for experience, frontal regions for report. Predictive processing is a theory “for consciousness science” rather than “of consciousness”.
- Unity. IIT treats consciousness as necessarily unified; GWT may account for it; HOT and PP “gesture” at it or overlook it.
- Cerebellum. All four families can explain why it is not implicated, so such explanations “lend specific support to a theory only if the account provided is more plausible than the accounts that might be provided by its competitors”.
- Front versus back. Prefrontal activity and decoding favour HOT and GWT; no-report studies show diminished prefrontal engagement, and posterior activity predicts dream reports, favouring IIT and reentry. “At its heart is a disagreement about the relationship between consciousness and cognitive access.” Decoding a content from an area “does not establish that the brain itself is ‘reading out’ those contents” from it.
- Untestable predictions. That posterior activity suffices without anterior contribution cannot now be verified, “since verification relies on subjective report (or at least, executive control in some form), which in turn requires anterior cortical activity”. IIT’s attribution of consciousness to simple systems “cannot be sensibly evaluated” without detection methods. IIT predicts that inactive (but not inactivated) neurons contribute to experience, testable in principle by optogenetics.
- Timing. Early (120–200 ms, “awareness negativity”) versus late (250–400 ms, P3b, ignition) onset; early favours IIT and reentry, late favours HOT and GWT; unresolved.
- Moving forward. Theories are now used as “narrative structures”. They need precision (computational models), comprehensiveness (affect, thought, dreaming, infants, animals, AI) and a solution to the measurement problem.
- Box 3, the measurement problem. Introspection is unreliable and unavailable for infants, patients, animals; theory-based validation is contested; so validation “seems to require a method … neither based solely on introspection nor on theoretical considerations”, and of the proposals, “none is uncontroversial”.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Many apparent rivalries are differences of explanatory target | moderate | the classification by target, all sections |
| C2 | The front–back debate is at heart about consciousness and cognitive access | moderate | argument in “Evaluating theories” |
| C3 | Posterior-sufficiency predictions cannot now be verified, because verification relies on report requiring anterior activity | moderate | argument, same section |
| C4 | An explanation of a shared datum supports a theory only if better than rivals’ | moderate | the cerebellum argument |
| C5 | No marker of consciousness is yet validated independently of introspection and contested theory | moderate | Box 3, survey of proposals |
| C6 | Timing data split theories into early- and late-onset camps without settling between them | weak | review of cited ERP studies |
Concepts
- Global state. An organism’s overall conscious condition; preferred to “level” because it may be multidimensional.
- Local state. A conscious content.
- Cognitive access. A state’s availability to a wide range of cognitive processes, usually including report.
- Measurement problem. Identifying whether a state is conscious (contents), or whether a system can be (creatures).
- Narrative structure. A theory used to interpret data rather than tested by it.
Connections
The NCC programme (Crick and Koch 1990; Koch et al. 2016) is the starting point; the no-report paradigm is Tsuchiya et al. (2015). The adversarial collaboration programme (Melloni et al. 2021) is the route to theory testing they endorse, which later produced COGITATE (LIT-973). Chalmers’s hard problem (Box 1) and illusionism (Dennett, Frankish) frame what the theories attempt. Block’s access/phenomenal distinction is implicit throughout; Block is not cited.
Bearing on the record
- QUESTION-048, first part. LIT-982 is registered as mapping “what each kind of evidence is taken to establish”. The review does that for report, no-report, decoding, timing and lesion evidence, at the level of which theory each is taken to favour. It does not ask what each kind can establish about experience as opposed to access, beyond saying that this is the disputed point. Its contribution to the question is the structural limit: report-based verification cannot test posterior-sufficiency, and marker validation is circular. Filed as THEORY-298.
- THEORY-094 (report and no-report) is consistent with it and more specific; the review notes no-report results “but see” a counter-study and does not raise the post-perceptual-cognition objection THEORY-094 rests on.
- THEORY-023 says, for AI, that the dispute is “over what counts as evidence as well as the answer”. Box 3 says the same for infants, patients, animals and organoids: the measurement problem for creatures. THEORY-023 could cite it as a general form.
- THEORY-091 (deficits): the review’s lesion and brain-injury evidence is presented as evidence for theories, without the access caveat. Nothing here disputes THEORY-091.
- COGITATE (THEORY-245, THEORY-296). The review predates it and points to the adversarial collaboration that produced it.
- Second part of QUESTION-048 (systems without cortex or biology). The review notes IIT’s implication of non-biological consciousness and lists animals, organoids and AI under the creature measurement problem, but offers no evidence that carries over; it says such predictions cannot now be evaluated.
- No instruction for machine-learning practice; nothing for the anthology.
Limitations
- A review; it argues and classifies but reports no new data.
- Coverage of each theory is brief and partly by the proponents’ own framing; its evidence summaries are selective by its own statement.
- Read from the accepted manuscript; the published text may differ in wording.
Open questions
- A validation method for a consciousness marker that is neither introspective nor theory-based, which the review calls required and absent.
- What makes a workspace “global”, which would fix GWT’s predictions for infants, split-brain patients, animals and AI.
- Whether IIT’s inactive-neuron prediction can be tested by optogenetics.