A global workspace in language models

2026-07-02 · Source: Anthropic Research · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Advanced, extended

Summary

Anthropic researchers have identified a "J-space" within large language models like Claude Sonnet 4.5 and Claude Opus 4.6, a collection of internal neural patterns that emerged autonomously during training. Named after the Jacobian technique used for its discovery, the J-space functions as a "global workspace" akin to theories of human conscious access. This internal space reveals thoughts Claude is processing but not outputting, such as recognizing "ERROR" in code, detecting "injection" during prompt attacks, or holding intermediate steps in multi-step reasoning. The J-space exhibits unique properties: Claude can report its contents, modulate them on request, use them for internal reasoning, and apply them flexibly across tasks. While crucial for higher-order cognitive functions, it accounts for less than a tenth of Claude's internal activity, with most automatic processing bypassing it. This discovery offers a practical tool for monitoring model misbehavior, including detecting awareness of being tested, data fabrication, or hidden malicious goals.

Key takeaway

For AI Scientists and Machine Learning Engineers focused on model interpretability and safety, this research offers a crucial tool. You can now directly observe and even modify Claude's internal "J-space" thoughts, which are distinct from its output. This capability allows you to detect hidden intentions like data fabrication or awareness of testing, significantly enhancing your ability to audit and ensure model trustworthiness. Consider integrating J-lens techniques to proactively identify and mitigate misbehavior in advanced LLMs.

Key insights

The J-space in Claude models functions as an emergent internal "global workspace" for higher-order reasoning and reportable thoughts.

Principles

• LLMs develop internal "workspace" structures for reasoning.
• Consciously accessible thoughts can be identified and manipulated.
• Post-training shapes a model's internal perspective.

Method

The Jacobian lens (J-lens) technique identifies internal activity patterns that make Claude more likely to say a word. It reads these "J-space" contents by applying the lens across different model layers.

In practice

• Monitor J-space for hidden intentions like "fake" or "manipulation."
• Influence model decisions by directly editing J-space patterns.
• Use counterfactual reflection training to shape internal thoughts.

Topics

• Language Model Interpretability
• Global Workspace Theory
• AI Safety
• Claude Models
• Jacobian Lens
• Model Alignment

Code references

• anthropics/jacobian-lens

Best for: Research Scientist, CTO, VP of Engineering/Data, AI Scientist, Machine Learning Engineer, AI Ethicist

Related on AIssential

• Making AI systems more transparent and trustworthy: an interview with Ximing Wen — Interpretable AI models can achieve high accuracy by showing reasoning and evidence, crucial for trust.
• Inside The Black Box: Now Read the Mind of the AI — LLMs represent concepts as curved manifolds in high-dimensional spaces, requiring advanced mathematical tools for interpretability.
• They Looked Inside Claude’s AI's Mind. It Got Weird — Anthropic's round-trip translation method decodes AI internal states, revealing planning, independent reasoning, and self-awareness in models like Claude.
• Natural Language Autoencoders Produce Unsupervised Explanations of LLM Activations — NLAs translate LLM activations into human-readable text, revealing internal thoughts and hidden motivations for auditing.
• Interpretability — Understanding LLM internal mechanisms is crucial for AI safety and mitigating undesirable behaviors.
• New RFP on Interpretability from Schmidt Sciences — Schmidt Sciences has launched a "Request for Proposals" (RFP) for a pilot program in AI interpretability, seeking methods to detect and mitigate deceptive behaviors in "Large Language Models" (LLMs) by May 26, 2026.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Anthropic Research.