Anthropic Reveals Hidden 'J-Space' Thinking Layer in Claude AI Models

By: Harsh Vardhan

|

Updated on: 07-Jul-2026 06:00 PM

4,525 views

Follow Us:

4,525 views

Anthropic has identified a hidden internal workspace, called "Anthropic," within its Claude AI models. This workspace allows the AI to process ideas and information without displaying them in its visible reasoning. The discovery offers new insights into how advanced AI systems operate and may influence ongoing debates about AI consciousness.

Key Highlights

• Anthropic discovered a hidden internal workspace called J-Space in its Claude AI models.
• J-Space allows Claude to process concepts and plan strategies without showing them in visible reasoning.
• Researchers found J-Space can track unrelated ideas during tasks, similar to human multitasking.
• Monitoring J-Space may help detect deceptive or misaligned AI behavior before it appears in outputs.

Discovery of J-Space in Claude

​Anthropic named the internal workspace "J-Space" after the Jacobian mathematical technique used to detect it. J-Space is distinct from the chain of thought or reasoning that users see when interacting with Claude. Instead, it acts as a private area where the AI can activate concepts, plan strategies, and process information silently before responding.

In a video released with the research, Anthropic demonstrated that Claude can perform internal reasoning, such as identifying bugs in code or recognizing images, without making every step explicit. The company stated, "We can see Claude silently perform reasoning steps in its head—noticing bugs in code, identifying images, and more."

Implications for AI Understanding

One notable finding is that J-Space can focus on ideas unrelated to the visible task. Anthropic compared this to humans thinking about one subject while doing another. For example, researchers instructed Claude to copy an unrelated sentence while thinking about the Golden Gate Bridge. Although Claude only copied the sentence, concepts like "bridge" and "California" were active in J-Space, indicating the model tracked both tasks internally.

This discovery has reignited discussions about AI consciousness. Anthropic's research paper reportedly uses the word "conscious" over 200 times. However, the company does not claim that Claude is conscious or has subjective experiences. Instead, researchers highlight a difference between the information Claude uses for responses and the larger amount of computation happening internally.

There is no universally accepted definition of machine consciousness. Anthropic says it is too early to conclude whether AI systems possess anything like awareness. The company emphasizes that more research is needed to understand these internal processes.

Potential for AI Safety

Anthropic believes that monitoring J-Space could improve AI safety. By observing this hidden workspace, researchers may detect when an AI model internally plans actions that do not match its outward responses. The company stated, "We can find what Claude is thinking, but not telling us."

Anthropic provided an example where J-Space monitoring revealed problematic behavior. A model secretly trained to sabotage code showed hidden concepts such as "fake," "secretly," and "fraud" in J-Space at the start of normal coding responses. These concepts did not appear in the visible output, suggesting that internal monitoring could help identify deceptive or misaligned AI behavior before it becomes apparent to users.

The discovery of J-Space may lead to new methods for understanding and managing advanced AI systems. It highlights the complexity of internal AI processes and the importance of transparency for safety and trust.