Proposed by Bernard Baars and developed into the global neuronal workspace model by Stanislas Dehaene and Lionel Naccache, the theory pictures the mind as many specialized processors running unconsciously in parallel, with a limited-capacity workspace on top. Information that enters the workspace is “broadcast” widely, which is what makes it reportable, holdable, and usable for flexible reasoning; everything else does its work without ever becoming accessible. It is a theory of access consciousness, deliberately agnostic about subjective experience.
The theory entered machine learning discourse when Anthropic’s 2026 J-space work tested its functional signatures (reportability, top-down control, use in reasoning, cross-task broadcast, selectivity) against a language model’s internals and found a workspace-like structure that emerged from training without being designed.