A Stage Without a Spectator: Global Workspace Theory and Artificial Minds

AA theatre has a stage, lit by a spotlight, surrounded by a much larger backstage and wings kept dark — actors rehearsing lines, stagehands shifting scenery, a prompter reading ahead in the script, none of it visible to the audience, all of it necessary for whatever appears in the light. Baars proposes this as a working model of how a mind is organised, not merely as an illustration: a great deal of specialised processing happens continuously and in parallel, mostly outside conscious access, and consciousness corresponds to some of that content winning a competition for a shared, limited-capacity stage that many otherwise-independent processes can read from¹. The theatre metaphor is vivid, and it invites one mistake worth heading off immediately: imagining an audience member watching the stage. Global Workspace Theory has no seat for one.

What "Global" Doesn't Mean FoundationalKnowledge that endures for decades — core principles

GlobalGlobal availability" is easy to over-read, and the theory is more precise than the word suggests on its own. It does not mean visible to every process in the system — plenty of specialised machinery never touches the workspace and keeps running regardless. It does not mean a complete representation of the system — the workspace carries whatever won the current competition for access, which is a narrow slice of everything happening, not a summary of all of it. It does not mean free from selection or distortion — getting onto the stage at all already involves competing against other candidate content, and nothing guarantees the winner is the most accurate or most important candidate, only the one that won. And it does not mean centrally witnessed — nothing in the theory requires an inner audience watching the stage; it requires only that whatever appears there becomes available for the other specialised processes to pick up and use.

Two Old Images, and Why Neither Quite Fits FoundationalKnowledge that endures for decades — core principles

TheThe theatre metaphor is new to cognitive science but not to philosophy, and two much older images of mind are worth setting beside it, precisely because each captures something real while getting the architecture wrong in a different, instructive way.

Descartes' dualism pictured a non-physical mind receiving and inspecting representations delivered to it by the body's mechanical sensory apparatus — an inner theatre, with a genuine spectator seated in front of it. Dennett names this picture the Cartesian Theatre and argues that it survives, uninvited, in a great deal of modern thinking about consciousness that explicitly rejects Descartes' dualism but keeps its structure: any theory that posits a moment or place where processing "becomes conscious" by arriving at a privileged internal screen has smuggled the theatre back in, homunculus and all². The decisive problem isn't that the homunculus is implausible — it's that positing one explains nothing. A little watcher inside the head, looking at an inner screen, would itself need to turn that screen into experience somehow, which requires either a further homunculus watching its screen (an infinite regress that never delivers an explanation) or a flat admission that the inner watcher does the job directly, in which case the screen was never doing any work and should be cut out of the story. Global Workspace Theory is only a genuine alternative to the Cartesian Theatre if the broadcast is read exactly as the sections above insist on reading it — as content made available to further processing, not content presented to anything that watches.adding a watcher never explains watching

Plato's image of the cave captures a different problem, easy to run together with the first but worth keeping separate. Prisoners, chained since birth facing a wall, see only shadows cast by objects carried behind them, and mistake the shadows for the whole of reality, having no direct access to what casts them³. This is a problem of mediation — indirect access to something real via a filtered appearance of it — not the homunculus's problem of an unexplained regress. Applied to a global workspace: the specialised processes reading the broadcast are, in a precise sense, exactly like Plato's prisoners. What reaches them is a shadow cast by a much richer process (the full, parallel, mostly-dark activity of the whole system) that none of them ever perceives directly. Unlike Plato's prisoners, no step of escape ever leads a subsystem out of the cave to see the fire and the objects themselves — there is no further fact about the system's "real" activity available to be compared against the shadow, waiting to be uncovered. The shadow is not a degraded copy of something better available elsewhere in the architecture. For every subsystem except whatever produced the original content, the broadcast simply is the only version of that activity that exists for it to use.

A Managed Crossing Point, Not a Mirror FoundationalKnowledge that endures for decades — core principles

TreatedTreated this carefully, the workspace is better described as a managed crossing point among partial perspectives than as a mirror the system holds up to itself. Specialised processes — perception, memory, planning, language, action selection — compete for access, cooperate once something has won it (building on the broadcast content rather than ignoring it), and interpret whatever gets broadcast according to their own, locally specialised machinery rather than a shared, system-wide format. A planning process and a language process receiving the same broadcast content use it for completely different purposes, in completely different internal representations, which is precisely why no single broadcast could function as a complete self-model: a self-model would need to mean the same thing to every subsystem reading it, and nothing in the architecture provides that.

Self-Observation Without Self-Completion FoundationalKnowledge that endures for decades — core principles

ThisThis gives a specific, more modest account of what self-observation in such a system could actually be: the system does not construct a complete model of itself, stored somewhere and consulted whole. It repeatedly makes selected aspects of its own ongoing activity available to other parts of itself, piece by piece, competition by competition, with no step in the process assembling those pieces into one finished, internally certified picture. Treated as a live interpretive proposal rather than an established finding, this reframes what consciousness might be doing functionally: less like an internal witness checking in on the rest of the system, and more like a protocol — a recurring procedure, not a persistent structure — through which a system selectively discloses parts of itself to its own other parts, on an ongoing basis, never once and for all.

That reframing is this page's own extension of the theory, offered as a way of thinking about what global broadcast is doing architecturally — it is not a consequence Baars' theory establishes on its own, and a reader should weigh it as an interpretive proposal rather than as a settled implication.a proposal to weigh, not a theorem GWT proves

A Formal Attempt: The Conscious Turing Machine Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

BlumBlum and Blum have proposed a precise computational formalisation of Global Workspace Theory, the Conscious Turing Machine: a model with a short-term working memory corresponding to the global workspace, and a large collection of specialised, parallel processors competing to write to it, each able to read whatever the workspace currently broadcasts⁴. The formalisation is a genuine technical achievement — it turns a metaphor-heavy theory into a model precise enough to implement, analyse, and extend — and it is worth being exact about what that achievement does and doesn't demonstrate. Building a working implementation of the competition-and-broadcast architecture shows that the architecture is formally coherent and computationally realisable. It does not thereby show that any implementation of it is conscious in the sense that matters to this series — whether there is something it is like to be a Conscious Turing Machine remains exactly as open a question after the formalisation as before it, for the same reason a precise formal model of a thermostat's control loop would not settle whether the thermostat feels its own temperature changing.

Strongest Objection: The Theory May Explain Access, Not Experience FoundationalKnowledge that endures for decades — core principles

Block'sBlock's distinction between access consciousness (information being available for reasoning, report, and behavioural control) and phenomenal consciousness (there being something it's like to have it) applies to Global Workspace Theory about as directly as it applies to anything in this series. Everything the theory describes — competition, broadcast, cross-process availability — is naturally read as an account of access: what makes information usable system-wide. Whether winning that competition is also sufficient for something to be felt is a further question the theory, as an account of availability, does not obviously answer on its own, and critics who press this point are not making an error the theory's defenders have a decisive reply to.

Provisional Conclusion FoundationalKnowledge that endures for decades — core principles

GlobalGlobal Workspace Theory gives a precise, checkable account of how specialised, partially independent processes might coordinate without any of them — or anything else in the system — holding a complete picture of the whole. That account fits naturally with this series' guiding thesis about distributed, negotiated cognition, and the Conscious Turing Machine shows it can be built, not merely described. Neither the theory nor its formalisation settles whether what gets broadcast is ever experienced, and this page treats that gap as the actual state of the evidence rather than a problem to be argued away.

Questions for Further Thought

• If a system's specialised processes could be shown to compete and broadcast exactly as the theory describes, what further evidence would bear on whether anything is experienced?
• Does the access/phenomenal distinction apply equally well to every theory of consciousness this series considers, or does it bite harder against some than others?
• Is "protocol of self-disclosure" a useful reframing of what a global workspace does, or does it smuggle back in exactly the unified inner perspective the theory was meant to avoid?
• Can you find a point in this page's own argument where the cave's mediation problem and the homunculus's regress problem get run together, despite the warning above against doing exactly that?

Related Topics

• Belief Without an Inner Believer? Dennett and Artificial Agency — a different route to the same conclusion that no single internal vantage point is required for a system's states to be usefully, even truthfully, described.
• Does a Self-Model Get You Anywhere Near an Analogue of Consciousness? — this site's own specific self-modelling architecture checked against Global Workspace Theory, including a concrete proposal for implementing broadcast via the shared-resource coordination already used in Multi-Agent Systems.
• Does Intelligence Require Consciousness? — this page's access/phenomenal distinction put to work directly on the question of whether intelligence and consciousness are the same architecture or two separable ones.

Further Reading

• Baars, B. J. (1988). A Cognitive Theory of Consciousness. Cambridge University Press.
• Dennett, D. C. (1991). Consciousness Explained. Little, Brown and Co. (for the Cartesian Theatre critique.)
• Plato. Republic, Book VII (the Allegory of the Cave).
• Blum, L., & Blum, M. (2021). A theoretical computer science perspective on consciousness. arXiv:2011.09850.
• Block, N. (1995). On a confusion about a function of consciousness. Behavioral and Brain Sciences, 18(2), 227–247.

References