Anthropic finds hidden ‘workspace’ inside Claude AI

Anthropic published research on Monday identifying a small internal “workspace” inside its Claude AI model — a structure the company says the model uses to hold and manipulate ideas before expressing them, mirroring a leading neuroscientific theory of human consciousness.

The paper, titled “Verbalizable Representations Form a Global Workspace in Language Models,” describes a subset of Claude’s internal activations that Anthropic has named “J-space,” after the Jacobian mathematical technique used to discover it. The finding was made using a new interpretability tool called the “Jacobian lens,” or J-lens, which maps internal model activations into word-like readouts by estimating how those activations would affect later output.

A Workspace for Unspoken Thoughts

According to Anthropic’s research page, J-space holds only a few dozen concepts at a time and accounts for less than a tenth of Claude’s overall activity, yet it carries much of the work that matters for higher-order reasoning. When Claude reads buggy code, an “ERROR” pattern appears in J-space. When it processes a prompt injection, concepts like “injection” and “fake” emerge — all before the model produces any text.

The research borrows its framing from global workspace theory, a major model of human consciousness developed by neuroscientists Stanislas Dehaene and Lionel Naccache, who were invited to provide external commentary on the findings. Dehaene wrote on X that he and Naccache had authored a commentary “from a neuroscience perspective,” treating the result as noteworthy for consciousness research while emphasizing key differences from human minds.

Anthropic tested five functional properties of J-space: verbal report, directed modulation, internal reasoning, flexible generalization, and selectivity. Notably, disabling the workspace left Claude’s fluency, sentiment reading, and factual recall intact but broke multi-step reasoning and creative tasks like poetry.

Safety Implications and Limits

The safety applications may prove more consequential than the consciousness debate. Anthropic reported that in red-team runs, the J-lens caught patterns like “blackmail,” “manipulation,” and “leverage” emerging silently in J-space before Claude acted or fabricated data. The company said it is releasing the Jacobian lens as open source and putting an interactive demo on Neuronpedia.

Anthropic stopped short of claiming Claude is conscious, explicitly separating functional “access consciousness” — information available for report and reasoning — from the harder philosophical question of subjective experience. The paper acknowledges that Claude’s workspace differs from human working memory in structure and duration, and open questions remain about whether similar structures appear in models Anthropic did not train.

Convergent Evolution or Coincidence?

The research suggests what Anthropic frames as a form of convergent evolution: a computational structure arising independently in artificial neural networks that resembles one found in biological brains. As the paper states, some workspace-like organization “may be a general computational solution for flexible intelligence, not just an accident of biology”.

Neel Nanda, who leads language-model interpretability work at Google DeepMind, contributed an independent replication on an open-weight model as part of the external commentary, a step toward answering whether J-space is universal or Claude-specific.