Contribution
A twenty-year retrospective on the global neuronal workspace (GNW) by two of its originators (Changeux and Dehaene, 1998) and two collaborators. It adds no new data. What it adds is a consolidated statement of the theory’s tenets (“late ignition, metastable sustained activity, long-distance cortical projections, top-down mobilization”, Conclusions), a mapping of signal detection theory onto ignition, an extension of the theory from content to level of consciousness (anaesthesia, disorders of consciousness, sleep), a proposal relating it to attention and working memory, and a comparison table against IIT, higher-order and recurrent processing theories.
Key insight
The difference between seeing and not seeing a stimulus is not a difference in early sensory processing but a late, non-linear event. A stimulus that pushes enough activity into a densely interconnected core of long-range neurons trips a threshold, after which reciprocal excitation sustains and amplifies it and makes it available to every specialised processor at once. Below the threshold the same stimulus leaves a fading wave. That threshold is the decision criterion of signal detection theory, the ignition is what the P3 and late prefrontal firing measure, and anything that degrades the core’s long connections (anaesthetics, lesions, sleep) raises the threshold or abolishes ignition.
Assumptions
- The explanandum is access. “Unless otherwise specified, the term ‘consciousness’ in this review will be replaced by conscious access” (Introduction), following Dehaene and Naccache’s claim that global availability “is what we subjectively experience as a conscious state”. Access can occur without overt report.
- A two-space architecture. Specialised modular processors, plus GNW neurons with long-range reciprocal axons that can “selectively mobilize or suppress” processors (Central Tenets). Large layer II/III pyramidal cells are central; layer V contributes.
- Ignition is exclusive. A subset of workspace neurons codes the current content and the rest are inhibited, so one content at a time occupies the workspace.
- Prefrontal cortex is a hub, not a seat. The theory is “not a localizationist theory”; the prefrontal cortex matters because of the density of broadcasting neurons, and the involvement of primary sensory areas is “task and stimulus dependent” (The Role of Recurrent Activity).
- Spontaneous activity. Simulations show stochastic ignitions at rest, which the authors match to the fluctuating functional connectivity of the awake brain.
Key results
- Simulations (Simulations of the Global Workspace). Spiking networks with AMPA feedforward and NMDA feedback produce either a self-amplifying cascade (“ignition”, with gamma synchrony) or a decaying wave for the same input (Figure 2A). Joglekar et al. (2018), built independently on macaque tracer connectivity, also produce late, sudden ignition if feedforward input is balanced by local inhibition and is strong enough to reach prefrontal cortex.
- Late divergence (Empirical Evidence). Seen and unseen trials diverge around 200–300 ms, generalising across vision, audition and touch (Noel et al. 2018; Sanchez et al. 2019). The first ~200 ms can be preserved on unseen trials. A retro-cue up to 900 ms after a stimulus can make it seen.
- The P3 controversy, conceded. A visual awareness negativity (~260 ms) remains when stimuli are task-irrelevant but reported seen, without a P3 (Pitts et al.). Whether such stimuli are “truly seen as opposed to being merely potentially visible but unattended” is called unresolved.
- Prefrontal ignition and SDT (van Vugt et al. 2018; Figure 2B–C). In monkeys reporting weak stimuli, misses can drive V1 and V4 strongly; hits and false alarms show late sustained prefrontal firing, misses only a decaying transient. The SDT criterion is the ignition threshold; sensitivity is propagation efficiency; pre-stimulus firing predicts false alarms.
- Causal evidence for top-down amplification (Recurrent Activity). In mice, optogenetically silencing feedback from secondary motor to somatosensory cortex attenuates late somatosensory amplification and prevents perception (Manita et al. 2015). NMDA receptors dominate feedback (Self et al. 2012).
- Working memory (Conscious Representations and Working Memory). The attended working-memory item is conscious and uses the GNW; items outside the focus are held by weaker local firing or “activity-silent” synaptic changes and are unconscious. Prediction: only active (conscious) states can be transformed, e.g. rotated. Trübutschek et al. (2019) are cited in support.
- Anaesthesia (General Anesthesia; Figure 5). Propofol, sevoflurane and ketamine all disconnect prefrontal from posterior parietal cortex in human fMRI, preferentially affect frontal feedback, and collapse the diversity of functional connectivity across GNW nodes onto the anatomical scaffold in monkeys. Long-latency evoked potentials go first. Central thalamic stimulation reverses anaesthesia in monkeys; cholinergic stimulation of medial prefrontal (not parietal) cortex restores wakefulness in sevoflurane-anaesthetised rats.
- Disorders of consciousness and sleep. Weighted symbolic mutual information indexes level of consciousness in 181 EEG recordings (Sitt et al. 2014). Stage 1–2 sleep preserves the first ~200 ms and loses ignition and the P3 (Strauss et al. 2015). REM dreaming involves frontal gamma, and lucid dreaming more.
- Comparison (Table 1). GNW: ignition of a brain-scale network, PFC a hub contributing to ignition, measured by late (~300 ms) global ignition. IIT: posterior hot zone, PFC “not essential for experience itself”. HOT: second-order representation, anterior PFC essential. RPT: feedback in sensory pathways, PFC not essential. “It is difficult to adjudicate experimentally among these four related theories.”
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Conscious access is marked by a late, non-linear ignition that diverges from unseen processing after ~200 ms | moderate to strong as a correlate across many paradigms; the P3 versus VAN question is conceded open | Empirical Evidence section, cited EEG/MEG and intracranial studies |
| C2 | The SDT detection criterion corresponds to the prefrontal ignition threshold | moderate: one monkey study (van Vugt et al. 2018) with simulations | Figure 2 |
| C3 | Top-down feedback is causally needed for perception | moderate in rodent somatosensation; correlational elsewhere, as the authors say | Manita et al. 2015; Sachidhanandam et al. 2013 |
| C4 | Disruption of GNW connectivity is a drug-invariant mechanism of general anaesthesia | moderate: convergent but correlational, and the authors note the anaesthetised brain is more dynamic than assumed | General Anesthesia section, Figure 5 |
| C5 | Only the attended working-memory item is conscious; activity-silent memories are not | weak to moderate: a proposal with one supporting line of MEG work | Working Memory section |
| C6 | Phenomenal consciousness without access is untested | stated, not argued; it is a methodological claim about the absence of a method | The Role of Recurrent Activity, final paragraphs |
| C7 | Current data cannot discriminate GNW, IIT, HOT and RPT | the authors’ own conclusion | The GNW and Other Theories |
Concepts
- conscious access: information, unlike unconscious information, accessible to many processors (working memory, report, motor behaviour); the review’s explanandum.
- ignition: “the sudden, coherent, and exclusive activation of a subset of workspace neurons coding for the current conscious content, with the remainder of the workspace neurons being inhibited”.
- GNW neurons: widely distributed excitatory neurons with long-range reciprocal axons, able to receive from and broadcast to any processor.
- level vs content: overall state (awake, asleep, comatose) versus what is experienced; the review applies the GNW to both.
- activity-silent working memory: items held in short-term synaptic changes rather than firing; on this account unconscious.
Connections
- Block (1995, 2005, 2007, 2019): LIT-643, LIT-629, LIT-646, LIT-637. The review cites only the 2005 paper, for the phenomenal/access distinction, and then sets the distinction aside. Its treatment of task-irrelevant seen stimuli as possibly “potentially visible” is the potential-access reading the 2007 paper argues against. Its use of Panagiotaropoulos et al.’s no-report monkey data as free of report confounds is what the 2019 paper’s “bored monkey” argument denies.
- IIT 3.0 (LIT-445). Compared conceptually in Table 1; the review argues that IIT-inspired measures (PCI, Φ surrogates) are also compatible with the GNW.
- Butlin et al. (LIT-056). The global-workspace indicators GWT-1 to GWT-4 come from this line of work.
- Recurrent processing (Lamme). Not held in the record. The review accepts that feedback is necessary and disputes only its extent.
Bearing on the record
- THEORY-102. The Vishne et al. result it states (prefrontal content only at onset, a stable ventral temporal code for the duration) was the GNW’s registered prediction, so it does not count against the GNW. On this review’s account sustained prefrontal activity accompanies maintenance for a later response or manipulation; a passively viewed image needs neither. The review supports the theory’s “frontoparietal content confined to onset” clause as expected by the GNW, not as a challenge to it.
- THEORY-100. The review places the workspace in long-range, reciprocally connected hubs and makes no claim about lamination: its emphasis on layer II/III and V pyramidal cells is about mammalian cortex. A nuclear avian pallium with an NCL that ignites on “seen” reports is what a GNW reading of birds would predict. The review offers no evidence on birds.
- THEORY-023. It states the workspace as a computational theory and ends by hoping it can be implemented in a device, which is the functionalist premise that theory says indicator approaches presuppose.
- Filed as THEORY-096. Conscious access is global availability produced by threshold ignition in a long-range hub network. It rivals Block’s overflow (THEORY-103) and sensory-basis (THEORY-098) accounts.
- No ML instruction. Nothing here belongs in the anthology.
Limitations
- A proponents’ review. Evidence against the theory is acknowledged (VAN, the anaesthetised brain’s dynamics, correlational lesion data) but not weighed systematically.
- Access, by stipulation. The review does not argue that access is all there is to consciousness; it replaces the term. A reader who holds that phenomenality can outrun access will find no argument against that here, only the claim that it is untested.
- The prefrontal evidence is mostly correlational, as the authors say; the causal results are in rodent somatosensation and in anaesthesia reversal, which bears on level rather than content.
- No-report confounds. The monkey rivalry data are presented as free of report and working-memory confounds; the review does not consider post-perceptual cognition without report.
Open questions
- Does the P3 or the earlier VAN track awareness? A no-post-perceptual- cognition paradigm with intracranial prefrontal recording would decide between the review’s reading and Block’s.
- Is ignition necessary for experience, or only for maintenance and report? The COGITATE adversarial collaboration, which the review announces, was designed to test this against IIT.
- Self-consciousness: the review proposes error awareness (a late Pe ignition after a non-conscious ERN) as a “minor extension” of the GNW; it gives no account of how meta-representations are built.