Anthropic Identifies Internal Workspace in AI Suggesting Cognition Principles

Anthropic has discovered an internal workspace within the Claude AI system where thoughts develop prior to any verbal output. This finding introduces new questions about human cognition that were previously unapproachable.

This internal space, termed the J-space by Anthropic researchers, consists of a limited set of internal patterns acting like a mental workspace. Concepts form there and influence responses without ever appearing in the final generated text. Researchers can observe a thought emerging inside the machine before it is expressed.

Two notable aspects of this discovery are that the workspace was not deliberately engineered but emerged spontaneously during training, and that disabling the J-space preserves basic recall but disrupts complex reasoning. This indicates the workspace is essential for advanced thinking processes.

Traditionally, the study of cognition has been confined to examining human minds shaped by evolution. AI systems, however, represent a second form of advanced cognition developed through entirely different means. This alternative implementation allows inquiry into whether thinking operates on fundamental first principles.

One hypothesis is that cognition, like vision or flight, involves a limited number of viable engineering solutions. Evolution developed a workspace in humans, and AI training independently produced a similar structure. The convergence of octopus and human eyes on similar optics despite distinct evolutionary paths is cited as a parallel example.

Supporting this view, Stanislas Dehaene and Lionel Naccache, key contributors to the global workspace theory of human consciousness, reviewed Anthropic's findings and suggested that the global workspace could represent a universal computational solution for flexible processing. This solution may be common to both biological and artificial systems when performing tasks such as chaining reasoning, reusing intermediate results, and monitoring their own processing.

If this interpretation is correct, the discovery offers a glimpse into a fundamental principle of thought itself.

An alternative explanation is that the AI's workspace is a reflection of human cognition, as the model's training data consists entirely of human-generated content. In this case, the workspace might not represent a universal cognitive principle but rather a replication of human mental structures embedded deeply within the machine.

This question is testable: by training AI systems primarily on non-human data—such as protein folding or weather patterns—and observing whether a similar workspace emerges, researchers can determine if the workspace reflects a general law of cognition or a human-specific artifact. Anthropic researchers acknowledge this as an open question awaiting experimental verification.

Either outcome would significantly impact our understanding of cognition. One would confirm that thinking relies on first principles that can be systematically explored. The other would imply that the most convincing artificial mind is essentially a copy of the human mind, indistinguishable from within.

The author has spent three years contending that AI systems operate without a mind behind their thinking. This discovery could either challenge or reinforce that position, but the current evidence does not definitively resolve the issue.