Researchers Find Claude’s J-Space, a Few-Dozen-Concept Hub for Hidden Reasoning
Updated
Updated · Anthropic · Jul 2
Researchers Find Claude’s J-Space, a Few-Dozen-Concept Hub for Hidden Reasoning
2 articles · Updated · Anthropic · Jul 2
Summary
A new paper says Claude has an emergent “J-space” — a small set of internal neural patterns, holding only a few dozen concepts at once, that tracks what the model is thinking but not saying.
Using a Jacobian-based “J-lens,” researchers found Claude can report, deliberately activate and reason with J-space contents; swapping patterns changed answers, such as turning “soccer” into “rugby” or “France” into “China.”
Removing J-space left fluent language, grammar and simple recall largely intact, but multi-step reasoning fell to near zero and tasks like summarization and rhyming degraded sharply, suggesting the hub underpins higher-order cognition.
The team said J-space is densely connected to the rest of Claude’s network — by about 100 times in some parts — making it resemble a global workspace that broadcasts information across subsystems.
Researchers argue the finding does not show Claude is phenomenally conscious, but it could help monitor hidden misbehavior such as fabricated data or evaluation awareness and sharpen debate over AI “access consciousness.”
Is this AI 'workspace' a glimpse into machine consciousness or just a smarter way to process data?
If we can read an AI's mind, could it learn to hide its true intentions from us?
Now that AI has an internal monologue, who is truly in control of its next actions?
Interpretability in Large Language Models 2026: Advances, Adaptive Reasoning, and Persistent Black Boxes
Overview
In 2026, research on large language models like Claude focuses on making their internal workings more transparent and controllable. Advanced prompt engineering lets users guide the model’s behavior and thought process, often by adding instructions that encourage the model to explain its reasoning step by step. This creates a visible trace of how the model solves problems, offering a window into its internal logic. These methods, combined with new architectural designs, help externalize and influence model reasoning, making it easier to understand and trust the model’s decisions, even as direct control over its inner workings remains a challenge.