TECH NEWS: Anthropic identified an internal representation in Claude that resembles aspects of the neuroscientific global-workspace theory. This does not prove consciousness.
The overwhelming majority of active synapses in the human brain perform automatic and involuntary tasks. These include regulating breathing and processing visual information.
Only a small portion of these processes becomes consciously accessible. Such thoughts can be put into words, held in mind, and used for reasoning.
In neuroscience, this is often described through global-workspace theory. According to the theory, thoughts become consciously accessible when they enter a privileged workspace accessible to different parts of the brain.
Anthropic has now studied a structure in its Claude model that resembles this idea in certain respects. The researchers presented their results in a detailed paper, but an important point must be made: the research does not claim Claude is conscious, and certainly does not claim that it has a soul.
Imagine an AI model’s internal operation as a massive office. At any moment, countless background tasks are underway, but there is also a main meeting room where broad concepts and strategic decisions come to the foreground.
Anthropic says Claude contains such a privileged internal space, called J-space. The model may use it when solving complex tasks, handling complicated situations, or trying to express its own reasoning in words.
The researchers developed an interpretability tool called the Jacobian Lens, or J-lens. It uses mathematical methods to identify internal representations that are readily available for the model to verbalize later.
The J-lens is not a mind-reading device. It instead examines a particular part of Claude’s internal calculations and attempts to predict which concepts may appear in future outputs.
It can be used to study questions such as this: if a particular signal activates in layer twenty, by how much does it increase or decrease the mathematical probability that the model will use the word banana five words later?
Anthropic researchers say they demonstrated the existence of such a central workspace in Claude. They asked the model which concepts it was focusing on, then used the J-lens to inspect which representations were present in J-space at that moment.
When Claude was asked to keep the concept of fairness in mind, researchers observed that a representation of fairness entered J-space and remained there. During complex tasks such as chess problems, J-space may function like a kind of internal notebook.
When researchers experimentally modified the content of that space, Claude’s final answer also changed. Other components of the model, even those working on different subtasks, could access concepts placed inside J-space.
According to Anthropic, this structure was not directly designed into Claude. It may have emerged during training because it proved useful for organizing calculations and complex tasks.
In biology, convergent evolution describes different species developing similar traits because those traits offer an efficient answer to a particular problem. Anthropic’s result resembles that idea only in the sense that a structural similarity appears between an artificial system and one theory of human cognition.
It does not mean that Claude is conscious in the human sense, and it does not establish intention, subjective experience, or a soul. The research instead suggests that a global, shareable workspace may be a mathematically useful solution for complex information processing.
