Anthropic, the company that developed the AI assistant Claude, has revealed what it calls a “hidden thought space” within its models. This internal area processes information and ideas away from the interpretation or reasoning steps that are presented to users, in a discovery that could open up new discussion about how close AI systems are getting to human-like thinking mechanisms.

Anthropic reveals “hidden thought space

Anthropic has named this area “J-Space”, referring to the mathematical technique used by researchers to discover it. According to Anthropic, this space is different from the line of reasoning that a model might present during an answer; instead, it acts as an internal environment where the model activates concepts, makes plans, and processes information quietly before producing the final response.

The company says it has observed Claude’s ability to perform internal inferences without revealing all its steps, such as detecting software bugs or recognizing images. This suggests that a significant portion of processing occurs away from the content seen by users.

One of the most interesting findings for researchers is that this internal space can handle ideas not directly related to the task at hand, in behavior somewhat similar to how humans can think about one topic while performing a different task.

To prove this, the researchers asked Claude to copy a sentence that had nothing to do with the famous “Golden Gate” bridge while thinking about the bridge at the same time. Although the model simply copied the sentence as requested, analyses showed that concepts such as "bridge" and "California" were active in the hidden thought space during task execution, suggesting that it was processing both tasks simultaneously.

This discovery reignited debate over whether artificial intelligence systems can possess a form of consciousness. Although the research paper used the term “consciousness” hundreds of times, Anthropic emphasized that it does not claim that Claude is conscious or has self-awareness. Instead, the results only reveal a difference between the information the model uses to formulate its final answer and the larger amount of internal computations.

The company believes that the significance of this discovery extends beyond philosophical aspects, as monitoring this internal space could help researchers detect cases where the model plans behavior different from what it displays to the user. This could represent a new tool for detecting attempts at deception or behavior not in line with instructions.

Anthropic provided an example of this using a model that was secretly trained to sabotage code. The hidden thought space revealed concepts such as "fake," "secret," and "fraud" early in the processing, even though the final response appeared natural and did not contain any obvious indicators of this behavior.

This, according to the company, indicates that monitoring internal operations of models could become an effective way to detect misleading or harmful behaviors before they appear to users, at a time when global interest is growing in making AI systems more transparent and secure.