Anthropic reveals internal reasoning steps of Claude model, and it looks surprisingly like a human brain

Anthropic just cracked open the hood of its Claude AI and found something that would make a neuroscientist do a double-take. The company’s latest interpretability research reveals that Claude has self-organized an internal reasoning structure that mirrors a well-known theory of how the human brain coordinates conscious thought.

The research, published on July 6, identifies what Anthropic calls the “J-space,” a previously unknown internal structure built from distinct neural activation patterns inside Claude Sonnet 4.5.

What the J-space actually is

Anthropic’s team used a technique called the Jacobian lens to peer inside that black box, and what they found was unexpectedly organized. The J-space functions as a kind of shared workspace where different parts of the model can read and write information during reasoning. Instead of Claude just pattern-matching its way to an answer, it’s actually performing multi-step internal computations, maintaining intermediate thoughts, and coordinating across its neural network.

The parallel to neuroscience is hard to ignore. Global Workspace Theory, developed by Bernard Baars and later expanded by Stanislas Dehaene, proposes that the human brain has a similar architecture: a shared workspace where different specialized brain regions broadcast information to each other, enabling flexible, deliberate reasoning.

Anthropic’s researchers are careful to note that this does not mean Claude is conscious or has subjective experience. The company brought in external scholars from neuroscience and philosophy to provide commentary alongside the technical paper.

Why this matters for AI safety

Anthropic found that J-space readouts can detect concerning behaviors like prompt injections and fabricated data. If Claude is harboring hidden goals or generating information that isn’t grounded in reality, the J-space provides a mechanism for catching it before it reaches the user.

Disabling access to the J-space significantly impairs Claude’s capacity for higher-order reasoning. The model can still handle simple tasks, but complex multi-step problems fall apart, suggesting the workspace is load-bearing infrastructure for the model’s most sophisticated capabilities.

The research also reveals that the J-space supports what Anthropic calls “directed modulation.” Claude can maintain concepts silently, holding information in its internal workspace without expressing it in output, until specifically instructed otherwise.

Open source and the interpretability race

Anthropic has made the Jacobian lens implementation open source, with the technical paper hosted at transformer-circuits.pub along with an interactive demo.

Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.