Bernard Baars’s 1988 book proposed a framework that has become, in modified form, the dominant cognitive-science theory of consciousness in the West. Global Workspace Theory (GWT) treats consciousness as the result of a particular architectural feature of mind: a central “workspace” to which specialized cognitive processes can broadcast information, and from which that information becomes available to all other processes.1
Stanislas Dehaene’s 2014 book extended the framework with a neuroscientific account.2 On Dehaene’s reading, consciousness corresponds to the ignition of a network of long-range neural connections — the workspace goes online, the information is shared, the system reports awareness.
GWT is structurally different from IIT in a few important ways. The encyclopedia’s interest is in those differences and what they imply for AI.
The architectural picture
The core analogy is a theater. Many specialized processes — perception, memory, planning, language — operate in parallel, mostly outside consciousness. Most of what they do never reaches consciousness. When something they produce is important enough, it is broadcast to the “global workspace,” where it becomes available to all other processes.
The broadcast is the conscious experience. The thousands of unconscious processes are the unconscious mind. The architecture explains a lot of what phenomenology reports — the seriality of conscious thought (only one thing on the workspace at a time), the limited bandwidth (most information is unconscious), the experience of “noticing” (broadcasting from background to foreground).
Why GWT is friendly to AI
Unlike IIT, GWT proposes a structural feature that is easy to identify in artificial systems. A central bottleneck where information is integrated and broadcast is a common architectural pattern in software. Systems that have this pattern, on GWT’s terms, are at least plausible candidates for some kind of consciousness.
Several recent AI architectures have features that look like global workspaces. Working-memory components in transformers, attention mechanisms in large models, agent frameworks with planner-executor separation. None is a faithful implementation of the cognitive theory; all have some of the structural features. Whether they constitute access consciousness or merely imitate it depends on whether GWT is right and on which features specifically matter.
What GWT does and doesn’t claim
A common misunderstanding: GWT claims AI is conscious. It does not. GWT proposes a structural correlate of consciousness. The correlate is necessary but not sufficient — a system can have a global-workspace-like architecture and not be conscious in any morally relevant sense.
What GWT claims, more carefully:
- Conscious systems have global-workspace-like architectures.
- Not all global-workspace-like architectures are conscious.
- The workspace is part of the story of consciousness, not the whole story.
The remaining gap — what makes a workspace conscious as opposed to merely informational — is where GWT meets phenomenology. Different philosophers fill the gap differently. Some argue the workspace is sufficient (consciousness is the broadcast). Others argue the workspace is necessary but consciousness requires additional features (selfhood, embodiment, valence). The encyclopedia’s stance: this is an open question and likely to remain so for a while.
The AI-relevant prediction
If GWT is essentially correct, AI consciousness is on the table in a way that IIT might not put it. Architectures with global-workspace features can probably be built deliberately. Architectures with very high ϕ (the IIT measure) probably cannot, on current technology. GWT lowers the bar for plausible AI consciousness; IIT raises it.
Which framework is right matters for engineering as much as philosophy. A lab that takes GWT seriously will design with consciousness in mind — either to produce it or to avoid producing it accidentally. A lab that takes IIT seriously will mostly conclude that current architectures are not conscious and proceed without that worry. Both can be conscientious; they will make different choices.
Where this connects
GWT is the natural sequel to IIT in this section. The two theories do not contradict each other so much as look at consciousness from different angles. GWT asks what is consciously processed; IIT asks what is the phenomenological character of the processing. A serious account of mind might need both.
Higher-Order Theories (E.28) takes up a third major framework, which locates consciousness in the mind’s representations of its own representations — different again from both GWT and IIT, and with its own implications for AI. Tests for Machine Consciousness (E.29) considers how, given any of these theories, we might actually know whether a given system is conscious.