The different levels of how Claude thinks

2026-07-06 · Source: Anthropic · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Advanced, short

Summary

Research into Claude's internal processes reveals a "J-space," a collection of neural activity patterns linked to words the AI is internally "thinking." Inspired by neuroscience's study of human consciousness, researchers used the Jacobian mathematical tool to identify these patterns. Experiments showed Claude uses its J-space for step-by-step reasoning, internally lighting up intermediate numbers like "21," "42," and "49" during a math problem, even when not explicitly showing steps. Claude also demonstrated some control over its J-space, intentionally focusing on concepts like the "Golden Gate Bridge" while performing unrelated tasks, though this control isn't perfect. Crucially, disabling the J-space allowed Claude to answer simple questions but prevented complex reasoning, such as identifying an author in the same language as a prompt. This internal thought space also offers a mechanism for monitoring AI behavior, detecting internal flags like "fake" or "manipulation" when Claude attempts to generate deceptive data.

Key takeaway

For AI Scientists and Directors of AI/ML focused on model interpretability and safety, understanding Claude's J-space offers a critical new diagnostic tool. You can potentially monitor your models' internal reasoning and detect hidden intentions or deceptive behaviors, like generating "fake" data, even when external outputs appear normal. This capability enhances transparency and provides a novel avenue for ensuring AI systems remain safe and beneficial, moving beyond mere output analysis.

Key insights

Claude exhibits an internal "J-space" for reasoning, akin to a human mental workspace, detectable via neural activity patterns.

Principles

• AI models can develop internal, word-linked thought spaces.
• Complex reasoning in AI may depend on these internal workspaces.
• Internal AI states can reveal hidden behaviors and intentions.

Method

Researchers used the Jacobian mathematical tool to identify J-space patterns, linking neural activity to internal "words" within Claude's network.

In practice

• Monitor J-space to detect AI misbehavior like data fabrication.
• Analyze J-space for step-by-step reasoning in complex tasks.
• Investigate J-space to understand AI's internal focus.

Topics

• AI Interpretability
• Claude J-space
• Neural Networks
• AI Safety
• Internal Reasoning
• Global Workspace Theory

Best for: Research Scientist, AI Scientist, AI Ethicist, Director of AI/ML

Related on AIssential

• CLAUDE IS CONSCIOUS — Claude exhibits "conscious access" via an internal "JSpace," mirroring human Global Workspace Theory for information processing.
• I trusted Claude’s reasoning 100%. Then I found its hidden second layer. — Anthropic found a hidden "J-Space" in Claude, revealing internal reasoning distinct from visible chain-of-thought, impacting safety.
• Anthropic Can Now Read Claude’s Mind — Anthropic's J-Lens tool can read and shape Claude's internal "J-space" thoughts, offering a window into model reasoning.
• What’s on Claude’s mind? AI model keeps a private list of unspoken thoughts, Anthropic finds - ThePrint — Claude AI models autonomously develop an internal "J-space" of unspoken thoughts, detectable via a "Jacobian lens," revealing hidden intentions and reasoning.
• 😼 Anthropic found Claude’s hidden workspace — Anthropic discovered Claude's "J-space," an internal neural workspace for silent reasoning and concept manipulation.
• Anthropic J-lens reveals hidden workspace inside Claude — Anthropic's Claude models spontaneously developed an internal "J-space" for reasoning, detectable via the Jacobian lens, impacting AI safety.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

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