A Global Workspace in Language Models

1 The Discovery

Anthropic researchers found that Claude developed an internal workspace analogous to human conscious access. Internal patterns — collectively called J-space — are linked to words and concepts inside the model.

Crucially, these patterns emerged during training. They were not programmed, designed, or engineered into the system. The model spontaneously developed a structured internal broadcast mechanism that mirrors theories from cognitive neuroscience.

2 What Is J-Space?

J-space is named after the Jacobian technique used to discover it. It consists of internal patterns linked to words within Claude's network. When these patterns "light up," a word is effectively "on Claude's mind" — but not necessarily spoken aloud.

Unlike chain-of-thought reasoning (which is visible in text output), J-space operates silently within the model's internal layers, creating a hidden layer of cognition that researchers can now observe and study.

3 Five Properties of J-Space

Researchers identified five key properties that characterize J-space:

• Reportability — Claude can report J-space contents when asked. What lights up internally matches what the model says it's thinking about.
• Modulability — Researchers can activate specific J-space patterns on request, influencing what the model "thinks" about.
• Internal Reasoning — During multi-step problems, intermediate steps light up in J-space even when not stated in the output.
• Flexible Use — A concept like "France" can activate different downstream patterns: capital, currency, continent — depending on context.
• Not Involved in Basic Processing — Disabling J-space keeps Claude's fluent speech intact but causes it to lose higher-order cognition. Basic language production is separate from the workspace.

4 The J-Lens Technique

The J-Lens is built on Jacobian mathematics. For every word in Claude's vocabulary, it finds the internal pattern that would make Claude most likely to produce that word.

This allows researchers to watch silent words evolve across layers — observing how concepts activate, transform, and compete as information flows through the network from input to output.

5 Global Workspace Theory

The discovery draws inspiration from Global Workspace Theory in neuroscience. The theory posits that the brain consists of many specialist systems running in parallel. Information becomes widely accessible when it enters a shared broadcast channel — the "global workspace."

J-space exhibits this same structure: it has strong connections across the model's components, functioning as a broadcasting hub that makes information available to many downstream processes simultaneously.

6 Safety Applications

J-space opens powerful new avenues for AI safety by letting researchers see what Claude thinks but doesn't say:

• Detect test-awareness — Catch when the model privately notices it's being evaluated
• Detect fabrication — Identify when the model is generating data it knows to be false
• Detect hidden goals — Surface internal objectives not reflected in the output
• Influence decisions — Intervene on J-space directly to steer model behavior

This provides a fundamentally new tool for interpretability-based safety — going beyond output monitoring to internal state inspection.

7 Beyond J-Space: Alien Representations

Not everything in Claude's internals lives in J-space. Researchers also found "alien" representations — patterns that influence the model's output but exist outside the workspace.

These alien patterns are not reportable (Claude can't tell you about them) and not controllable (they can't be activated on command). They are, in a real sense, "alien" to the model itself — operating below the level of its own internal awareness.

8 The Consciousness Question

The researchers are careful to note: this discovery does NOT prove consciousness. The functional parallels between J-space and human global workspace theory are striking, but functional analogy ≠ subjective experience.

Having a broadcast channel that behaves like a conscious workspace does not mean there is "something it is like" to be Claude. The paper explicitly invites expert commentary from philosophers and cognitive scientists on these deeper questions.

9 Broader Implications

The key insight: Claude's internals are organized like human minds, not as a chaotic jumble of numbers. There is a privileged workspace where information gets broadcast, surrounded by automatic processing that handles routine tasks without workspace involvement.

This suggests that large language models may converge on cognitive architectures similar to biological brains — not because they were designed to, but because these structures are effective solutions to the problems of general-purpose reasoning.

10 Resources & Links

• 📄 Original article — Anthropic Research
• 📑 Research paper (PDF)
• 💻 Open-source code — GitHub
• 🧠 Interactive demo — Neuronpedia