Anthropic links Claude’s J-space to neuroscience. Does it mean AI is becoming conscious?

July 7, 2026: Anthropic’s latest interpretability research says Claude has an internal “workspace” that resembles a neuroscience theory of conscious access. The company says the findings could improve AI safety and transparency but stops short of claiming its chatbot is conscious.

Anthropic has unveiled new research suggesting that its Claude family of large language models (LLMs) has developed a small internal “workspace” where concepts can be held, manipulated and used for reasoning before they appear in the model’s responses. The researchers call this internal representation the J-space and say it emerged naturally during training rather than being explicitly programmed. 

The study, “Verbalizable Representations Form a Global Workspace in Language Models,” was published on July 6 by Anthropic’s interpretability team. Alongside the paper, the company released an open-source implementation of its Jacobian Lens (J-lens) technique and an interactive demonstration developed with Neuronpedia to allow researchers to inspect similar representations in open-weight models. 

The research has attracted attention because it borrows ideas from neuroscience to explain how modern AI systems organise some of their internal computations. However, Anthropic emphasises that the work does not demonstrate that Claude has subjective experiences or human-like consciousness. 

What is J-space?

According to Anthropic, J-space is a collection of internal neural activity patterns associated with words or concepts that the model is “thinking about” even if those words never appear in its response.

Unlike chain-of-thought reasoning, which is expressed as text, J-space exists entirely inside the model’s neural activations. Researchers say it allows Claude to silently plan, keep track of concepts and perform multi-step reasoning without writing those intermediate thoughts down.

The name comes from the Jacobian, a mathematical concept used in the J-lens technique that enabled researchers to identify these internal representations. 

Why is Anthropic comparing it to neuroscience?

The research was inspired by Global Workspace Theory, a prominent neuroscience and cognitive science framework. 

The theory suggests that the human brain performs many specialised processes unconsciously. Only a small amount of information enters a shared “workspace,” where it becomes available for deliberate reasoning, planning and verbal reporting.

Anthropic states that J-space appears to play a functionally similar role inside Claude by allowing information to be shared across different parts of the model during complex reasoning tasks. The company says this is an analogy based on function rather than evidence that language models work like human brains. 

How did researchers discover it?

Anthropic developed a technique called the Jacobian Lens, or J-lens, to analyse the model’s internal activations.

Instead of looking only at Claude’s output, the researchers examined which internal activity patterns made particular words more likely to appear later in a conversation. By applying the technique across multiple neural network layers, they say they could observe concepts appearing, changing and disappearing while the model worked through a task.

The researchers also performed intervention experiments.

In one example described in the research paper, replacing an internal representation of “France” with “China” caused Claude to consistently answer questions using Chinese capitals, languages and currency instead of French ones. According to Anthropic, this suggests the representations are not merely correlated with outputs but play a causal role in reasoning. 

Does this mean Claude is becoming conscious?

No. The company explicitly states that its findings should not be interpreted as evidence that Claude has subjective experiences or feelings.

Instead, the paper focuses on what philosophers call access consciousness, which is information that can be reported, deliberately manipulated and used for reasoning.

This differs from phenomenal consciousness, which refers to subjective experience or the feeling of being aware. Whether access consciousness implies phenomenal consciousness remains an open philosophical question, and Anthropic says it takes no position on that debate. 

Why does the research matter?

Beyond the debate over consciousness, Anthropic says the discovery could have practical implications for AI interpretability and safety.

If researchers can monitor internal reasoning rather than relying only on a model’s final output, they may eventually be able to detect hallucinations, hidden planning or unsafe behaviour earlier in the reasoning process. 

The company argues that understanding these internal representations could help make increasingly capable AI systems easier to audit and evaluate.

Researchers also found that, when Claude’s access to J-space was disrupted, the model continued speaking fluently and recalling simple facts but performed worse on higher-order reasoning tasks, suggesting the workspace plays a selective role in complex cognition rather than routine language generation. 

What remains uncertain?

While the study represents a significant advance in AI interpretability, several questions remain unresolved.

The research has not demonstrated that all language models possess an equivalent workspace, nor does it establish that such a workspace is sufficient for consciousness. Independent researchers will also need to validate Anthropic’s findings using other models and methodologies before broader conclusions can be drawn. 

For now, the study provides evidence that Claude contains an internal reasoning workspace that shares functional similarities with ideas from neuroscience. Whether those similarities reveal anything about machine consciousness remains an open scientific and philosophical question.

What the study found

• Claude appears to contain a small internal reasoning workspace called J-space. 
• The workspace emerged during training rather than being explicitly programmed. 
• J-space differs from chain-of-thought because it exists entirely within the model’s neural activations. 
• Editing some J-space representations changed Claude’s responses in controlled experiments. 
• Anthropic believes the findings could improve AI interpretability and safety.

What remains uncertain

• The study does not show that Claude has subjective experiences.
• It does not prove AI consciousness.
• It remains unclear whether similar workspaces exist in every large language model.

Mansi Hake