Anthropic discovers 'global workspace' layer in Claude that mirrors human conscious thought
Researchers at Anthropic identified a small set of activation patterns in Claude that function like the 'global workspace' concept from consciousness neuroscience, holding thoughts the model considers but doesn't always express.
Anthropic researchers published findings this week showing that Claude's neural architecture contains a structure functionally similar to the "global workspace" concept from neuroscience. The structure, which the team calls J-Space, emerged spontaneously during training and wasn't explicitly programmed by developers.
J-Space holds only a few dozen concepts at a time and accounts for less than 10 percent of the model's total internal activity. Despite its small footprint, neurons in J-Space connect to the rest of the network at densities roughly 100 times higher than typical patterns. Most of the model's computation runs as automatic background processing, inaccessible to the model's own "awareness," but J-Space acts as a scratchpad for information the model actively considers before generating output.
Parallels to human cognition
The researchers draw direct parallels to how human minds work. An experienced driver shifts gears without conscious thought, but can report which gear they're in when asked — the information moves from automatic processing into conscious awareness. Claude behaves similarly: a line-length counter manages text wrapping in the background, but when a task requires stating the length, that same information surfaces in J-Space and becomes available for reasoning.
The team also observed an effect analogous to the "white bear" phenomenon in psychology. When instructed not to think about a concept, the model partially suppresses it, but the thought still breaks through. In post-trained versions of Claude, the word "damn" appears in J-Space alongside the forbidden concept — the model appears to notice its own failure, much like a meditator catching a wandering mind.
The study doesn't claim Claude possesses phenomenal consciousness or subjective experience. It does demonstrate what philosophers call "access consciousness" — the ability to manipulate information, reason about it, and use it to guide behavior. Anthropic sees practical applications: monitoring J-Space could help detect deception attempts, and training the model to articulate ethical principles in response to hypothetical "stop and think" prompts improves behavior even when no one asks the question. Training what the model would say changes what it thinks.
The full paper is available at Transformer Circuits.
