Hidden "Thinking Space" Discovered Inside Claude AI

AI researchers have successfully identified a special "workspace" inside the Claude language model that resembles processes in the human brain. This discovery suggests that neural networks are not merely algorithms predicting the next word, but systems that form internal logical stages when performing complex tasks. These hidden states are invisible to the user but directly influence the system's final decision. This is reported by Ixbt.com reports .

A team of scientists in the field of interpretable AI used a new analysis method called the Jacobian lens (J-lens) to study the internal processes of the Claude model. According to ixbt.com, this method allowed them to link hidden signals within the model to clear concepts by mathematically analyzing neuron activation. As a result, a hidden space called "J-space" was found.

Similarity to the Human Brain and Internal Processes

Researchers note that J-space is functionally similar to the "global workspace" characteristics of the human brain. This system brings information into a state open for conscious control. In the Claude model, each state is associated with a specific concept or word, but its activation does not mean the model will necessarily write that word. It is more of an intermediate stage that prepares information for subsequent internal processes.

Crucially, J-space differs from the usual "chain-of-thought" method. While in a chain-of-thought the model writes a text draft to itself, in J-space everything happens solely through signals at the neuron level. For example, if the model is asked about the number of legs of a web-spinning creature, it activates the concept of "spider" in the internal space, but may answer "eight" directly without using that word in the response.

Experimental Intervention and Results

To prove that this hidden space is not just a passive indicator but actively participates in decision-making, scientists conducted an interesting experiment. They asked the model to "mentally" choose a certain category of sport and then name it. Before Claude answered, it was noted via J-lens that the concept of "soccer" had been activated in its internal space.

Then, the researchers intervened directly in the computations, replacing the "soccer" pattern in the internal neuron signals with "rugby" without giving any text command. As a result, Claude changed its choice and announced that it had chosen rugby. This confirms that the hidden space controls the model's logical conclusions.

It should be specifically noted that this discovery does not mean that the Claude model possesses consciousness or subjective feelings. However, the research shows that modern language models can spontaneously form internal information processing structures to perform complex cognitive operations. This will help in better understanding how AI "thinks" in the future.