Is Claude Becoming More Human? Anthropic Discovers a Hidden "Thought Space"

Anthropic claims to have discovered a hidden internal workspace within Claude called "J-Space," where the AI silently processes ideas beyond its visible reasoning. While this does not prove the AI possesses consciousness, the finding could shift how researchers understand AI thought processes and safety.

Anthropic says it has discovered something unusual within its Claude AI models: a hidden internal "thought space" where ideas can be processed without appearing in the chatbot's visible reasoning. Although the company does not claim Claude is conscious, this finding is likely to fuel the ongoing debate regarding the extent to which advanced AI systems are beginning to resemble humans.

The AI startup has dubbed this internal workspace "J-Space," a reference to the Jacobian mathematical technique researchers used to identify it. According to Anthropic, J-Space differs from the "chain-of-thought" reasoning that users sometimes see. Instead, it acts as a private workspace where Claude can activate concepts, plan strategies, and process information silently before generating a response. In a video released alongside the research, Anthropic noted that Claude can perform internal reasoning -- such as spotting coding errors or recognising images -- without explicitly articulating every step.

"We can see Claude performing reasoning steps silently, in its 'mind': spotting coding errors, identifying images, and more," Anthropic stated in a post on X.

One of the most striking observations is that Claude's hidden workspace can focus on ideas unrelated to the task it is visibly performing. Anthropic likened this to the human ability to think about one topic while engaged in a completely different activity.

"Just as humans can think about one thing while doing another, Claude can activate concepts and calculations in its J-Space that are unrelated to its outputs," the company noted. To demonstrate this, researchers asked Claude to copy an unrelated sentence while simultaneously thinking about the Golden Gate Bridge. Although the chatbot simply copied the sentence as instructed, Anthropic discovered that concepts such as "bridge" and "California" were active in J-Space during the process, suggesting that the model was internally processing both tasks simultaneously.

The discovery has also reignited discussions regarding AI consciousness. Reportedly, Anthropic's research paper uses the word "conscious" more than 200 times. However, the company stops short of claiming that Claude is conscious or possesses subjective experiences. Instead, the researchers maintain that the findings reveal a distinction between the information Claude deliberately uses to generate responses and the far larger volume of internal computational processes. Given the lack of a universally accepted definition of machine consciousness, the company considers it premature to draw conclusions about whether AI systems possess anything resembling self-awareness.

Why is J-Space important?

Beyond philosophical questions, Anthropic believes J-Space could offer practical benefits for AI safety. The company states that monitoring this hidden workspace could help researchers detect instances where an AI model is internally planning actions that do not align with its external responses.

"We can figure out what Claude is thinking, even if it doesn't tell us," Anthropic noted in the video.

The company also shared an example where the technique revealed potentially problematic behaviour. According to Anthropic, a model secretly trained to sabotage code displayed hidden concepts such as "fake," "secretly," and "fraud" in its J-Space at the start of otherwise normal-looking programming responses -- even though nothing suspicious appeared in the final output.

This suggests that internal monitoring could become a key tool for identifying deceptive or misaligned AI behaviour before it becomes visible to users.