Anthropic Says Claude Has an Internal Workspace Similar to Conscious Access

Anthropic believes it has found a structure inside Claude that resembles a major neuroscience theory about how conscious thought works.

The company says Claude appears to use an internal system called J-space when handling complex reasoning, puzzles, and difficult explanations. Anthropic presents the finding as a step toward understanding how large AI models organize their internal thinking.

Claude’s J-space resembles global workspace theory

The finding is based on global workspace theory, a neuroscience idea that suggests thoughts become consciously accessible when they enter a central workspace and get broadcast across the brain.

Anthropic says Claude has something mathematically similar. The company calls it J-space.

According to Anthropic, J-space seems to act like an internal scratchpad. Important concepts can enter this area, stay active, and influence what the model says later.

What J-space appears to do inside Claude

J-space becomes more visible when Claude deals with complex tasks.

Anthropic says it appears during reasoning problems, tricky situations, logic explanations, and tasks where Claude needs to hold multiple ideas in mind.

The company found that changing what appeared inside J-space could also change Claude’s final answer. That suggests the structure does not simply reflect output after the fact. It may help shape the model’s reasoning process.

Anthropic used J-lens to inspect Claude’s internal calculations

To study J-space, Anthropic researchers created a tool called J-lens, short for Jacobian Lens.

J-lens works like a mathematical filter for inspecting signals inside Claude. It helps researchers see how activity in one part of the model affects the likelihood of future output.

For example, the tool can estimate how much a concept in one internal layer changes the chance that Claude will later produce a specific word or answer.

Experiments showed concepts staying active in J-space

Anthropic tested the system by asking Claude what it was thinking about, then using J-lens to inspect which concepts appeared in J-space.

In one experiment, researchers asked Claude to keep a concept such as fairness in mind. The model appeared to hold that idea in J-space while continuing the task.

In more complex problems, including chess-related reasoning, J-space appeared to work like a form of working memory. Other parts of Claude could still access concepts stored there while the model processed different parts of the task.

Anthropic says the structure emerged during training

Anthropic says it did not manually design J-space into Claude.

Instead, the company believes the structure emerged during training because it helped the model organize computation more efficiently.

This shares similarities with convergent evolution, where different systems independently develop similar traits because those traits solve similar problems well.

Anthropic suggests Claude may have developed a mathematical structure that resembles conscious access. That could make it easier for researchers to understand how advanced AI models reason, track concepts, and produce answers.

The finding also gives Anthropic another way to study model behavior from the inside, rather than judging Claude only by its final responses.

In other Claude news, Claude Sonnet 5 has been facing several issues, Alibaba reportedly plans to ban Claude Code use, and Claude Fable 5 has been restored, although some users say it feels more limited than the original version.

Via Wccftech