Anthropic Finds a 'Global Workspace' Inside Claude

Anthropic has published research proposing that Claude's internal architecture bears a functional resemblance to the 'global workspace' model long used in cognitive science to explain how human consciousness works. The theory holds that only a limited subset of information processed by the brain becomes broadly accessible at any given moment, effectively acting as a bottleneck that filters vast unconscious processing down to a small, shareable slice of content. Anthropic's researchers say they have found something structurally similar happening inside Claude. To investigate this, the team built a new interpretability tool called J-lens, designed to trace which parts of the model's internal activity actually become accessible to the model itself during processing. The findings indicate that, much like the human brain, only a small fraction of Claude's total internal computation is ever available for the model to draw on when producing an output. The rest remains, in effect, background processing that never surfaces into anything resembling a usable or reportable state. Anthropic is deliberately cautious about what this means. The company is explicit that identifying a global workspace-like structure is not the same as claiming Claude has subjective experience or consciousness in any meaningful sense. Instead, the value of the finding is framed in practical terms: understanding how information is filtered, prioritised and made accessible within the model is directly useful for interpretability and safety work. If researchers can map which internal states become part of this accessible workspace, they are better placed to understand how Claude arrives at its outputs, and potentially to spot when something in that process is going wrong. This research fits into Anthropic's broader interpretability programme, which has increasingly focused on opening up the internal mechanics of large language models rather than treating them purely as black boxes. Tools like J-lens add to a growing toolkit aimed at tracing internal representations, features and now something closer to attention-like bottlenecks that determine what a model can 'see' of its own processing at any given step. Whether or not the global workspace comparison holds up under further scrutiny, Anthropic's framing suggests this line of work is aimed squarely at practical safety outcomes rather than philosophical claims about machine minds.