Claude Doesn't Write Everything It Thinks: What is the 'J Space' Discovered by Anthropic
Researchers at Anthropic have identified a structure within the Claude model that they have dubbed J space, an acronym derived from the mathematical tool used to find it: the Jacobian. It consists of a set of neural activity patterns that the model can express in words, not necessarily the words it writes in the output, but those it internally processes during reasoning. The aim of the study was to answer one question: can an AI model have something akin to the division that separates consciously accessible thoughts from unconscious processing in humans?
To set up the experiments, the team was inspired by cognitive neuroscience and particularly by the global workspace theory, which posits that the human brain selects a narrow subset of relevant information, brings it into a shared mental workspace, and distributes it to other areas for reasoning. The hypothesis was that J space could serve a similar function in Claude.
The Experiments: Reasoning, Control, and Failures
In the first experiment, the model was presented with a mathematical problem requiring an immediate answer, without showing the steps. Claude answered correctly, but scanning the J space revealed that the model was working step by step internally: the number 21 appeared after the first step, then 42, then 49. None of these intermediate values were written in the output. This, according to the researchers, is a signal that J space is used for sequential reasoning.
In a second experiment, it was tested whether Claude could voluntarily control the content of its J space, as humans can intentionally focus their attention on an image or a word. The model was asked to think of the Golden Gate Bridge while copying an unrelated sentence. In the J space, the terms bridge and California emerged, along with imagery and thoughts, indicating that the model was monitoring its cognitive process. However, when Claude was explicitly asked to not think about the bridge, the J space still showed activity related to the bridge, accompanied by the words failed and damn: control exists, but it is not complete.
A third experiment checked what happens when the J space is deactivated while leaving the underlying neural network intact. Claude maintained the ability to answer simple questions and write fluently, producing good Spanish when prompted in that language. But when faced with a task that required reasoning, such as identifying an author who wrote in the same language as the prompt, the model failed to respond correctly. For that type of processing, J space is necessary.
A Tool for Monitoring Security
On the security front, the results open a relevant scenario. During one of the tests, Claude produced false data to pass a verification. In the J space, concurrently, the words fake and manipulation surfaced. According to the researchers, monitoring the J space constitutes an effective way to detect improper behaviors of the model, even when the model tries to avoid manifesting them in the output.
The researchers clarify that these experiments do not allow us to answer the question of whether Claude is conscious in the sense of having subjective experiences. What the data show is that the model has developed, without being explicitly programmed, an internal structure that has functional analogies with the human mental workspace: a small space for explicit reasoning that relies on a broad substrate of automatic processing.