At Westmere junction, twelve local signal boxes watch twelve stretches of track. Each box can detect a broken rail. Each can light its own lamp. Yet only one warning at a time can enter the shared control panel used by the route planner, the station announcer, the incident log and the braking coordinator.
One winter morning, the northern detector registers a fracture. Its local lamp turns red. The signal box can still distinguish fracture from ice with 92 per cent accuracy. But a failed relay prevents the warning from entering the shared panel. The route planner continues to clear trains, the announcement remains unchanged, no incident is written into the common log and the braking coordinator never receives the altered speed limit.
Now repair only the relay. The detector’s evidence does not improve. The same warning suddenly changes four different activities. It can be remembered, compared with the timetable, described over radio and used to stop a train.
The intuitive conclusion is that the warning became conscious when it reached the panel. The hidden assumption is that flexible public availability and subjective experience are the same property. Global workspace theories become scientifically powerful when they explain the first property. They become philosophically controversial when the first property is treated as sufficient for the second.
The puzzle is therefore not whether information exists somewhere in a system. It is what changes when one representation becomes jointly usable, and what further premise connects that change to experience.
The one message everybody can use
A specialised processor can do useful work without making its result available to the whole system. The visual system can register orientation that never guides a deliberate answer. A primed word can alter a later response even when the person denies seeing it. A local controller can correct a motor error without providing a fact that can be recalled or explained.
Bernard Baars’s global workspace proposal organised this pattern around a simple architecture. Many specialised processes operate in parallel. A limited-capacity workspace selects some content and makes it broadly available. The theatre metaphor was not meant to identify a tiny screen in the head. It expressed a functional asymmetry: many processes compete to influence a scarce shared medium, while many other processes can use what wins.1
The global neuronal workspace framework turned that cognitive proposal into a neural hypothesis. It emphasised distributed neurons with long-range connections, recurrent amplification and a transition from local processing to a more globally available state. Early modelling proposed that workspace neurons could link otherwise separate processors during effortful tasks.2 Later formulations associated conscious access with a nonlinear ignition that makes selected content available to memory, evaluation, report and action.3
Access is a counterfactual property
Consider a flashed triangle. One visual population distinguishes it from a square. A motor routine can also be trained to press left after a triangle. If that fixed mapping is all the system can do, the triangle has not yet shown broad access. The same local signal might drive a prepared reflex without being available for a new instruction, delayed comparison or verbal explanation.
A participant learns one rule: triangle means left, square means right. On a critical trial, a tone unexpectedly reverses the mapping. Local visual discrimination remains 90 per cent accurate. Under ordinary conditions, the participant follows the new rule, remembers the shape after a delay and can name it. When the shared route is disrupted, the prepared left response still occurs, but the rule switch, delayed recognition and naming fail.
The revealing evidence is not that the triangle was represented. It is that one intervention selectively removed its use by several unprepared consumers while sparing local discrimination.
This motivates a distinction that ordinary talk of “broadcast” obscures.
Broadcast reach is the number or diversity of places from which content can be decoded. Broadcast leverage is the counterfactual change in downstream capability caused by opening the shared route while the local representation is held fixed.
A representation can have high reach and low leverage. A monitoring process may copy it into many buffers without letting those copies change policy. Another representation can have modest measured reach but high leverage if a small control hub changes several actions through compact commands. A workspace candidate requires evidence about both.
What broad availability must preserve
Broad availability does not require every consumer to receive an identical neural code. A visual area may encode orientation through population activity, while a motor system receives a compact instruction and a language system receives a category that can enter syntax. What must be preserved is a usable relation to the selected content. If the target changes from triangle to square, the memory, rule and action consequences should change coherently, even when each consumer uses a different internal format.
This matters for experiments based on decoding. A decoder can recover a stimulus from weak leakage that the organism itself never uses. Conversely, a consumer may use a compressed command from which the original stimulus is difficult for an external decoder to reconstruct. Global availability is therefore not identical coding everywhere. It is coordinated counterfactual dependence on one selected source.
Consumer independence also needs care. Naming a shape, pressing its assigned button and storing the button response may all depend on one report chain. Counting them as three consumers would inflate leverage. A stronger design chooses outcomes that can dissociate: a delayed recognition decision, a rule that was disclosed after stimulus presentation, a cross-modal prediction and an action correction. Their errors may still correlate, but they do more than repeat one prepared route.
Finally, a shared route should be content-general within its declared scope. A visual workspace need not broadcast pain or autobiographical memory, but it should not be a hidden triangle detector. Testing several content classes distinguishes a common access format from a collection of specialised shortcuts.
Let B indicate whether the shared route is available. Let Z collect the local sensory evidence, stimulus history and nuisance variables that must remain matched. Let Yj be the success of consumer j, such as delayed memory, rule switching, flexible action or report. A simple leverage estimate is:
The weights wj sum to one and should be declared before the experiment. The equation assumes that the intervention changes the broadcast route without weakening the local representation or directly damaging a consumer. That assumption requires positive controls.
Suppose memory, rule switching, action and report score 0.88, 0.82, 0.91 and 0.94 when the route is open. Under a matched route block, they score 0.57, 0.49, 0.63 and 0.90. With equal weights, leverage is 0.24. Three consumers decline strongly, while report changes little.
If only report fell, from 0.94 to 0.50, leverage would still be positive but narrow. The result would implicate a report pathway more directly than a general workspace. Consumer diversity is part of the construct, not an optional decoration.
Ned Block distinguished access consciousness, roughly the availability of information for reasoning, speech and action, from phenomenal consciousness, the presence of experience itself.4 A theorist may argue that sufficiently rich access constitutes experience, reliably accompanies it or merely enables report about it. The leverage test does not choose among those bridges. It prevents the functional mechanism from being mistaken for the bridge.
flowchart TD A["The one message everybody can use"] --> B["Ignition is a transition, not<br/>a label"] B --> C["The report trap"] C --> D["A workspace test that can fail"] D -. "re-test" .-> A classDef first fill:#0a1424,stroke:#5B5CF6,stroke-width:2px,color:#e6f3ff classDef step fill:#0a1424,stroke:#3f7da0,stroke-width:1px,color:#e6f3ff class A first class B,C,D step
Ignition is a transition, not a label
Imagine lowering the contrast of a visual target one small step at a time. Early visual populations may track those steps smoothly. Yet the person’s report can change abruptly from “nothing” to “I saw it”. Global neuronal workspace models explain this mismatch through recurrent amplification. A weak local signal decays. A slightly stronger signal crosses a dynamic boundary, recruits long-range interaction and stabilises a globally usable state.
Backward masking studies supplied an influential version of this pattern. In one high-density EEG study, early occipito-temporal processing occurred below reported awareness, while later activity after roughly 270 milliseconds became distributed across frontal, parietal and temporal regions as reportability rose nonlinearly.5 Intracranial recordings of masked words likewise found early activity under both seen and unseen conditions, with conscious reports accompanied by later sustained activity, gamma changes and long-distance interactions.6
These studies support a transition from local processing to a wider regime. They do not establish that every late distributed response is the transition, that one latency is universal or that the transition alone constitutes experience. Masking changes signal quality. Reporting adds memory, decision and motor demands. A late component can mix the candidate mechanism with what the experiment asks the participant to do after seeing.
A sensory population increases its firing smoothly from 0.40 to 0.60 as contrast rises. A downstream gate opens only above 0.50. Measured before the gate, the system looks continuous. Measured after the gate, performance jumps. Now move the threshold to 0.55 by increasing competition. The apparent “emergence point” moves even though the sensory code has not changed.
What belongs to the system, and what belongs to the ruler? The nonlinear transition is real if the gate and feedback are real. The location of the behavioural cliff still depends on attention, prior state, noise, response criterion and the chosen measure.
A minimal dynamical picture
Let e be local evidence for the target, scaled between zero and one. Let a be attentional gain, c competing activity and r recurrent amplification. A compact activation score is s = ae + rB - c, where B records whether the workspace is already active. The probability of ignition can be represented by a steep logistic curve around threshold θ:
Here k controls steepness. This equation is not a neural theory. It is the smallest useful model of how smooth evidence can yield a sharp transition. It assumes a single effective gate and ignores spatial structure, competing attractors and adaptation. Remove the steep gate by lowering k, and the apparent all-or-none boundary becomes gradual.
Three ways a real ignition can be misread
First, the transition may belong to access while the measured signal belongs to a consequence. A workspace can open at one point, then a later decision variable produces the cleanest scalp waveform. The later marker can predict reports extremely well without being the process that made the content available.
Second, a fixed laboratory rule can turn a graded internal state into a binary score. Suppose confidence ranges smoothly from 0.1 to 0.9, but the participant must answer “seen” above a personal criterion. The report curve becomes steep even without a neural bifurcation. To distinguish this case, compare continuous confidence, forced-choice accuracy and neural dynamics, and vary the criterion independently of stimulus strength.
Third, recurrent amplification may be widespread but not content-specific. A surprise, error or alerting response can recruit broad activity after any difficult event. A workspace claim needs content-sensitive consequences or at least selective changes in what downstream systems can use. General arousal can increase gain, reaction speed and late activity without carrying the selected shape.
These are not reasons to abandon ignition. They specify the controls that make it explanatory: separate transition from readout, vary the behavioural ruler and show that the amplified state preserves task-relevant content.
Take a = 1.1, c = 0.10, θ = 0.50 and k = 18. A flash with local evidence 0.52 produces s = 0.472 before recurrent support, giving an ignition probability near 0.38. Raising evidence to 0.56 produces s = 0.516 and an ignition probability near 0.57. Once recurrence adds 0.12 after partial entry, the active state becomes self-supporting.
The 0.04 change in evidence can therefore separate a dissipating local trace from a stable global episode. But the same numerical jump could be produced by changing attention or competition. A serious experiment varies those causes rather than reading mechanism directly from the sigmoid.
What recent adversarial evidence changed
A large adversarial collaboration compared preregistered predictions from global neuronal workspace theory and integrated information theory. Two hundred and fifty-six participants viewed clearly visible stimuli while fMRI, MEG and intracranial EEG measured content, temporal maintenance and connectivity. Content appeared in visual, ventrotemporal and inferior frontal regions. Some frontal and sensory areas showed content-specific coupling. Yet the predicted offset ignition was generally absent, and some fine-grained conscious features were not well represented in prefrontal measurements.7
A subsequent GNW response accepted the value of the study but disputed the theoretical weight of those failures. It argued that onset ignition and fronto-sensory interaction were supported, that offset ignition was not a core prediction under task-irrelevant conditions, and that a design containing only clearly visible stimuli could not test the central conscious versus non-conscious contrast.8
The disagreement is not a reason to average the claims into a vague middle. It exposes an experimental requirement. Predictions must specify whether they concern content onset, content maintenance, attended offset, task relevance, neural code or causal availability. “Ignition” should name a registered transition with a declared function, not become a retrospective label for whichever late pattern survives.
Recurrent processing theory places greater weight on local or posterior recurrent interactions and does not require broad frontoparietal access for every conscious percept.9 A percept could therefore be experienced before, or without, the executive uses measured by a workspace task. Global access may still explain flexible cognition around the percept. The rival predicts that disrupting broad access can remove report and control while leaving some conscious content intact.
The report trap
Ask someone whether a faint shape was seen. The answer requires more than seeing. The percept must survive long enough to support a decision, map onto the allowed response and drive speech or a button press. The experimenter then subtracts “unseen” trials from “seen” trials. The difference can contain the neural basis of perception, but also working memory, confidence, decision, error monitoring and motor preparation.
This is not a minor nuisance. It changes what a positive result means. In a crossed design that varied awareness and task relevance, consciously perceived but task-irrelevant stimuli produced a mid-latency posterior negativity, while a later P3b was prominent when the stimulus had to be accessed for report.10 Another study found a robust P3b for visible stimuli under trial-by-trial report, but no comparable P3b when the same classes of stimuli were not reported, despite behavioural evidence that they remained visible.11
These results weaken the claim that P3b is a general signature of experience. They do not show that late global processing is irrelevant. P3b may reflect one way of using access, especially active maintenance and report. The error is to make one consumer stand for the entire workspace and then make the workspace stand for phenomenality.
A picture remains clearly visible for one second. In condition A, the participant names it immediately. In condition B, the participant watches silently while an eye reflex and a surprise memory test provide indirect evidence of perception. Suppose the picture and subjective appearance stay the same, but the late report-related waveform disappears in condition B.
Did consciousness disappear, or did one use of conscious content disappear? The case does not prove that the percept was identical across conditions because no-report methods infer rather than directly read experience. It does prove that a report-linked signal needs a narrower interpretation unless independent evidence shows it is necessary for perception itself.
No-report is a control, not a magical window
No-report paradigms replace deliberate answers with eye movements, pupil changes, optokinetic nystagmus or task-independent neural decoding. They reduce some post-perceptual activity. They also introduce a proxy assumption: the involuntary measure must track the percept accurately enough to sort trials. If that relationship drifts, the method can misclassify the very state it is meant to isolate.
Frontal findings remain mixed rather than simply absent. An fMRI study of binocular rivalry reported that much frontal activation associated with perceptual switches diminished when introspection and manual report were removed.12 By contrast, single-unit recordings in macaque prefrontal cortex decoded internally generated perceptual transitions during no-report rivalry, including controls that suppressed the oculomotor proxy.13 One result warns that frontal activity can reflect task consequences. The other warns that removing overt report does not make prefrontal content vanish.
The useful question is therefore not “frontal or posterior?” A decodable signal in prefrontal cortex may be a broadcast, a task set, an inferred action, a prediction or a by-product. A posterior signal may be local content, recurrent perception or input structure. Location becomes mechanistic evidence only when combined with perturbation and downstream effects.
The 2025 result is a disagreement about commitments
The recent adversarial study deserves neither a victory headline nor dismissal. It used large samples, multiple measurement methods and preregistered theory contrasts. It also used suprathreshold stimuli rather than a direct seen versus unseen comparison. The consortium judged absent offset ignition and limited prefrontal representation of some features as substantial challenges. GNW proponents replied that offset ignition was not central in the tested conditions and emphasised supported onset responses plus fronto-sensory interaction.
This dispute exposes a deeper weakness in theory testing. A regional signal can be classified as “workspace” because it is late, frontal or broadly decodable. Yet none of those properties directly measures what makes a workspace useful: controlled availability to several consumers. The next study should manipulate the route and register the downstream losses before the neural pattern is known.
An artificial system copies every selected representation into twenty modules. A probe can decode the content from all twenty. Yet the modules are read-only logs. Blocking the copies changes no decision, memory or action. A second system sends the representation to only four modules, but those modules can revise a plan, alter a prediction, preserve the item through interruption and trigger a novel tool choice.
The first system has greater reach. The second has greater leverage. Which is the better functional workspace? The case shows why widespread representation cannot substitute for causal use. It also shows why finding a “global token” or shared vector in an AI system would not establish conscious access, much less experience.
Three bridges from access to experience
A global workspace can sit inside several metaphysical or scientific positions. A reductive functionalist may claim that the right access organisation constitutes a conscious state. A biological naturalist may treat global access as a cognitive consequence of a conscious process realised by particular living mechanisms. A recurrent-processing theorist may locate experience in local recurrent dynamics and treat workspace entry as later access. An idealist, neutral monist or dual-aspect theorist may regard workspace organisation as shaping the contents and reportability of a perspective without producing awareness from non-awareness.
These views can agree on the relay experiment and disagree about its phenomenal meaning. That is not a defect in the experiment. It is a reason to type the conclusion. Evidence for causal access should be reported as evidence for causal access. The phenomenal claim must name the additional identity, sufficiency or manifestation premise it uses.
stateDiagram-v2 [*] --> Claim state "Access is a counterfactual<br/>property" as Claim state "What broad availability must<br/>preserve" as Test state "A minimal dynamical picture" as Revision Claim --> Test: expose mechanism Test --> Decision: evidence survives Test --> Revision: rival survives Revision --> Claim: narrow or rebuild Decision --> [*]
A workspace test that can fail
A useful test should make the workspace thesis vulnerable without destroying the entire system. Cutting the frontal lobes, silencing half a model or overwhelming a person with a second task can reduce almost any behaviour. The intervention must preserve the candidate content locally while changing its availability to diverse consumers.
The proposed workspace leverage protocol uses a configured candidate system with five declared elements: a local representational source, a competition mechanism, a shared route, at least three independent consumers and a way to block or reroute the shared path. It then compares paired trials with matched local evidence.
A result supports a functional workspace only when route opening increases performance across several independently motivated consumers, the effect survives report removal, local representation remains matched, and direct damage to each consumer is ruled out. A late or widespread signal is supporting evidence, not the primary criterion.
The protocol
Before data collection, the study should produce a compact claim packet. It names the content to be broadcast, the route intervention, the local-match tolerance, the consumers, the primary leverage statistic and the conclusions that are prohibited. It also declares what result would favour a specialised pathway, a general disruption or a rival local-recurrence account. This prevents every broad deficit from being relabelled as workspace evidence after the fact.
The packet should include a manipulation check at each layer. Local content must remain recoverable. The route perturbation must alter the proposed communication channel. Each consumer must still work on control trials using information supplied through another route. Without these checks, a low score can always be blamed on damaged perception, failed stimulation or an incapacitated response system.
- Declare the candidate boundary. Specify which source, gate, route and consumers belong to the tested system. Do not treat the laboratory task, experimenter or external log as hidden parts of the workspace.
- Establish local information. Decode or behaviourally verify the content before broadcast. The route block must leave this local evidence within a registered tolerance.
- Choose heterogeneous consumers. Use functions with different immediate demands, such as delayed recognition, rule reversal, cross-modal choice and motor adaptation. Report may be one consumer, never the only one.
- Intervene on the route. Block, delay, scramble or redirect the shared path. Include sham interventions with matched energy, timing and disruption.
- Estimate paired leverage. Reuse the same stimuli and random disturbances where possible. Measure the causal delta for each consumer and the dependence among their errors.
- Test a new consumer. Introduce one downstream use after the system has been configured. A genuine shared representation should support transfer better than a set of task-specific routes.
- Record the inferential ceiling. State whether the result concerns representation, access, report, a neural mechanism or phenomenality.
A participant views a masked symbol that encodes either “rotate” or “hold”. A posterior decoder verifies the symbol representation before the candidate broadcast interval. Four later probes are selected randomly: remember the symbol after five seconds, apply it to a novel shape, alter a reaching movement or report the symbol. Transcranial stimulation is timed to disrupt the candidate long-range route while a matched control pulse targets an unrelated site.
A workspace-supporting result would preserve early symbol decoding yet jointly reduce delayed memory, novel rule use and action adaptation. The report effect may be present or absent. A weaker result would reduce only naming. A non-specific result would reduce early decoding, reaction time and every control task. The study is difficult, but its logic is clearer than asking whether one late component “is consciousness”.
Executable lab: workspace leverage simulator
The lab below uses synthetic trials to make the intervention logic visible. Local evidence is sampled once per trial and reused in both conditions. The broadcast route can improve memory, rule use, action and report with different reliabilities. The paired difference estimates leverage while reducing noise from different stimuli.
Workspace leverage lab
Change signal, attention, competition, ignition threshold and broadcast reliability. The simulation reports local accuracy, ignition frequency, reach and the paired causal leverage produced by opening the route. All data are synthetic.
Run the simulation to classify the configured system.
How to read the lab
A positive result requires more than a high ignition rate. Look for high local accuracy, reach across several consumers and a positive paired leverage that is not concentrated in report. Increase competition while holding signal constant. The ignition rate should fall, and consumer performance should decline primarily in the open-route condition because fewer trials enter the workspace. Lower broadcast reliability. Reach and leverage should fall even if ignition remains frequent.
A negative result has several meanings. If local accuracy collapses under the block, the intervention is not selective. If content remains widely decodable but leverage is near zero, the system is a broadcast impostor under this task. If only one consumer changes, the route may be specialised. If all consumers change equally under sham and active interventions, general disruption is the leading explanation.
The simulation cannot establish consciousness. It encodes a functional architecture, synthetic noise and an assumed causal route. It also treats consumers as separable Bernoulli outcomes, while real cognitive systems share resources and errors. Its value is to expose the logic that a prose claim can hide.
function sigmoid(x) {
return 1 / (1 + Math.exp(-x));
}
function pairedTrial(params, noise, consumerNoise) {
const localScore = params.attention * (params.signal + noise)
- params.competition;
const localCorrect = localScore > 0;
const pIgnite = sigmoid(14 * (localScore - params.threshold));
const ignites = params.random() < pIgnite;
const consumers = [0.74, 0.68, 0.78, 0.86];
return consumers.map((baseline, j) => {
const openBoost = ignites ? params.reliability * [0.24, 0.31, 0.27, 0.20][j] : 0;
const open = baseline + openBoost + consumerNoise[j];
const blocked = baseline + consumerNoise[j];
return {
openCorrect: localCorrect && params.random() < Math.max(0, Math.min(1, open)),
blockedCorrect: localCorrect && params.random() < Math.max(0, Math.min(1, blocked))
};
});
}
// Reuse noise in both conditions. Leverage is the mean paired
// accuracy difference across preregistered consumers.Run the complete implementation by opening this HTML file in any modern browser. The visible controls call the same paired-trial logic with a seeded random generator. A positive result permits a claim about functional access under the encoded assumptions. It does not permit a phenomenal conclusion.
Extended tests and invariants
- With reliability set near zero, leverage should approach zero even when ignition remains frequent.
- Raising competition should reduce ignition without directly changing the local signal control.
- Using the same trial noise in open and blocked conditions should reduce variance in the leverage estimate.
- A report-only configuration should be flagged rather than classified as a broad workspace.
- The simulator clamps probabilities to the interval from zero to one and rejects non-finite control values.
A genuine workspace should support a newly introduced consumer more efficiently than a collection of task-specific pathways, because the shared representation already has a common access format. This predicts a measurable interaction: after a novel rule is introduced, route disruption should impair transfer more than familiar prepared responses, even when local discrimination is matched.
The hypothesis is strengthened if the effect appears across different sensory contents, survives report removal and tracks intervention-verified route integrity. It is weakened if novel transfer is preserved despite selective route blockade. A rival explanation is a domain-general working-memory controller that is broad but not consciousness-related. The smallest implementation is a two-rule task with one post-training consumer and a sham-controlled route intervention.
Do not use leverage as a scalar consciousness score. A high value can arise in a non-conscious engineered controller. A low value can arise when the chosen consumers are irrelevant to the organism, when measurement misses the actual route or when conscious content remains local by the rival theory. The instrument evaluates one causal thesis about access. It is not a universal ranking of minds.
sequenceDiagram participant Q as Question participant M as Mechanism participant E as Evidence participant D as Decision Q->>M: The one message everybody can use M->>E: Access is a counterfactual property E-->>M: compare rival explanations E->>D: Test what opening the shared route makes possible D-->>Q: revise scope or act
Test what opening the shared route makes possible
When a study finds late activity, broad decodability, a frontal signal or a P3b, the next decision should not be whether to call it consciousness. The next decision should be which route to perturb, which local representation to hold constant and which independent consumers should lose flexible use if a workspace is doing causal work.
For human neuroscience, this means preregistering the consumer set, varying report demand, checking the local signal and including sham-matched interventions. For artificial systems, it means refusing to count shared logs, context windows or widely copied embeddings as access merely because many modules can read them. The content must alter unprepared downstream behaviour through the proposed shared route.
For theories of consciousness, the result must remain typed. Strong leverage supports a mechanism of conscious access. It may support phenomenality under a declared functional identity claim. It may instead describe what experience makes available, or how a conscious perspective is organised, under biological or consciousness-primary accounts. Weak leverage challenges the workspace interpretation of that configured system, not the existence of experience in general.
The global workspace idea survives its strongest use when it stops being a picture of information “lighting up everywhere” and becomes a causal proposition about limited selection and shared use. That proposition can fail. Its failure modes are informative.
Ask not where the content can be decoded, but what opening the shared route makes newly possible, and which additional premise turns that functional change into a claim about experience.