In their new article, Anthropic explains that it has identified an internal space in Claude called "J-space", which functions like a global workspace where certain concepts become available to the model even when they are not expressed in its output.
The researchers show that this space plays a causal role in complex reasoning: changing what appears inside it can alter the model’s answers, while removing it mainly weakens abilities such as multi-step reasoning, summarization, and structured writing, without stopping the model from speaking fluently or answering simple questions.
The article also highlights safety implications, since J-space can reveal hidden internal representations or intentions that do not appear in the generated text. Anthropic is careful not to claim that Claude is conscious, but argues that this structure resembles a functional mechanism for "access" to information, comparable in some ways to global workspace theory in neuroscience.