Home / Theories of Consciousness

Global Workspace Theory and AI: Could a Machine Qualify?

Glowing pods in a ring, linked by threads of light to a central pod

Global workspace theory is the theory of consciousness that most clearly leaves room for AI. It says a state becomes conscious when its information is broadcast across a whole system, so that many separate processes can use it at once. Nothing in that description requires neurons, which is why the theory sits at the centre of the AI consciousness debate.

In July 2026 that debate moved from argument to evidence. Anthropic researchers reported a workspace-like structure inside Claude with many of the functional properties the theory describes. The same researchers take no position on whether Claude experiences anything, and a large brain study published in 2025 had already challenged parts of the theory itself.

What global workspace theory says

Bernard Baars proposed the theory in his 1988 book A Cognitive Theory of Consciousness. In the Stanford Encyclopedia of Philosophy's summary, processes in the mind compete for a resource with limited capacity, and whatever wins is broadcast to the rest of the system. That capacity limit matches the familiar limits on attention and working memory: you can only hold a few things in mind at once.

Broadcasting is the key move. Information that stays inside one specialised process is unconscious on this view; information that is broadcast becomes available to memory, reasoning, speech and action together.

Streams of light flowing into a bowl and one golden stream rising out

How the brain version fared in 2025

Stanislas Dehaene and colleagues turned the idea into a neural model in 2000, the global neuronal workspace. In this version, conscious perception begins with the "ignition" of a wider network including frontal and parietal areas. A 2022 paper describes the workspace as taking information from one specialised processor and making it globally available.

In April 2025 the theory faced its rival in a head-to-head test. The Cogitate Consortium published the results in Nature. They matched some predictions of both global neuronal workspace theory and integrated information theory, but challenged key tenets of each. For the workspace theory, the problems were a general lack of ignition when a stimulus disappeared and limited representation of some conscious content in the prefrontal cortex.

Glowing ripples spreading across a fountain basin from one drop of light

Global workspace theory and AI in 2026

Because the theory is stated in terms of what a system does with information, it can be applied to machines, and positions on it have shifted with the evidence:

  • The case for. In 2024 Simon Goldstein and Cameron Domenico Kirk-Giannini argued in A Case for AI Consciousness that if the theory is correct, language agents, language models wrapped with memory and planning tools, might easily be made phenomenally conscious if they are not already.
  • The case against. In 2023 David Chalmers listed the lack of a global workspace among the obstacles to consciousness in language models, as explained in can AI be conscious.
  • The 2026 evidence. Wes Gurnee, Nicholas Sofroniew, Jack Lindsey and colleagues at Anthropic found a "J-space" in Claude Sonnet 4.5: a small, changing set of concepts the model is poised to put into words. It supports verbal report, holding an idea in mind while doing something else, and reasoning, and it has limited capacity, as the theory requires. Removing it left simple tasks such as judging sentiment largely intact but sharply weakened multi-step reasoning and translation, and made the model's descriptions of experience more mechanical.

A closer look: does a workspace in Claude make Claude conscious?

Not on the evidence so far, and not according to its discoverers. They study access, the availability of information for use, and say they take no position on phenomenal consciousness, adding that the philosophical implications are unclear and likely controversial. They also note that the J-space shares only some architectural features with the brain's workspace.

A second limit comes from theory. The September 2026 framework From cacophony to hierarchy points out that global workspace theory can be read in two ways: as a claim about computation, which a model could satisfy, or as a claim about a system's physical causal structure, which a model running on ordinary chips may not. The strongest structural view is integrated information theory, which holds that function alone is never enough.

Checklist: reading a claim that an AI has a global workspace

  • Is there real broadcast? The theory needs one limited stage whose contents reach many separate processes.
  • Is there competition and a capacity limit? Selection is part of the theory, not a detail.
  • Is the theory itself being assumed? A system that meets the conditions is conscious only if the theory is true, and the 2025 brain results put parts of it in doubt.
  • Access or experience? The 2026 findings concern access; experience is a further question.

More in Theories of Consciousness · Report a correction