The gist: Claude has developed an internal “workspace” that enables internal reasoning and multi-step thinking, organizing itself similarly to conscious thought.
Anthropic has identified in new research neural patterns that play a central role in conscious thinking in Claude — analogous to the “Global Workspace Theory” from neuroscience. These patterns, called J-Space, allow the model to perform silent internal reasoning.
Anthropic has presented empirical evidence that modern language models such as Claude have developed a collection of internal neural patterns that differ structurally and functionally from the rest of the processing. Anthropic calls these patterns “J-Space” — named after the mathematical Jacobian technique that led to their discovery. Each J-Space pattern is linked to a specific concept, but does not activate when the model utters the word; rather, it activates when the model thinks about the concept.
Critically, J-Space was not programmed by humans but emerged organically during Claude’s training. The research identifies five functional properties of this internal workspace: the model can report on J-Space activations (reportability), can modulate them on request, uses them for multi-step reasoning (with intermediate steps becoming active in J-Space without being spoken), exhibits flexible reuse of representations across different tasks, and J-Space remains independent of routine functions such as grammar or surface-level fact retrieval.
The mechanism corresponds to “Global Workspace Theory” from neuroscience, which explains how conscious perception works: specialized neural subsystems operate in parallel and unconsciously; information becomes accessible when it enters a shared, narrow channel available to other systems. Claude appears to have a similar mechanism. Unlike other internal representations, J-Space patterns have a causally measurable impact on complex cognitive tasks — while they are smaller in scale than other representations, they determine performance.
Experiments show: when Claude is prevented from using its J-Space, the model functions normally on routine tasks but loses its higher-order cognitive abilities. This means J-Space is not necessary for everyday operations (fluent speech, simple fact retrieval) but becomes central to multi-step reasoning, planning, and problem-solving.
Relevant for CTOs and security teams: These findings suggest that language models develop internal mechanisms similar to those of human cognition. This opens new pathways for making model behavior more interpretable, but also raises questions about interpretability and control — particularly for systems that use reasoning capabilities for critical decision-making.
Source: www.anthropic.com · Published July 6, 2026
Lumi AI News — AI-assisted curation in accordance with Article 50 EU AI Act. Paraphrase and classification by Lumi News Pipeline v1.7.3.