Anthropic Proposes Claude's J-Space Workspace, Drawing 1 Parallel to Human Consciousness
Updated
Updated · Gizmodo · Jul 7
Anthropic Proposes Claude's J-Space Workspace, Drawing 1 Parallel to Human Consciousness
3 articles · Updated · Gizmodo · Jul 7
Summary
Anthropic said a new paper identifies an internal "J-Space" in Claude where the model appears to carry out reasoning steps apart from its visible outputs.
The company links that workspace to global workspace theory—a leading account of human consciousness—while saying its experiments do not show Claude has experiences or feelings.
Anthropic's blog post, X post and video describe Claude reasoning "in its head" and even "thinking about its own thinking," language critics say risks overstating a mechanistic finding.
Earlier reporting said the structure emerged during training and was detected with a Jacobian-based "J-lens," a result Anthropic has framed as useful for interpretability and AI safety.
If an AI can spontaneously create one hidden 'thinking room,' could it be building others that we simply cannot find yet?
When we can edit an AI's internal thoughts, where is the line between ensuring safety and unethical mental manipulation?
Inside J-space: Anthropic’s Breakthrough in AI Interpretability and Safety for Claude Models
Overview
Anthropic has made a major breakthrough in understanding its Claude AI models by discovering J-space, an internal workspace that emerges naturally within the system. Using a new tool called the Jacobian lens (J-lens), researchers can now look inside the AI's 'black box' and observe how it processes information and makes decisions. This insight not only helps reveal the AI's hidden objectives but also plays a key role in detecting safety concerns, ensuring the models behave as intended. The emergence of J-space marks a significant step forward for AI interpretability and safety, offering a clearer view into AI cognition.