Anthropic finds internal 'workspace' in Claude AI mirroring consciousness theory

Artificial intelligence firm Anthropic announced on Sunday the discovery of a spontaneously developed internal structure within its Claude language models. Dubbed a 'J-space,' this finding suggests the AI has formed a functional internal workspace that mirrors key aspects of global workspace theory, a prominent model of human consciousness.

The research utilized a novel mathematical technique called the Jacobian lens (J-lens) to probe the AI's neural network. Anthropic's scientists observed that the J-space acts as a small, privileged zone where the model holds and processes concepts it can report on and reason with. This contrasts with a larger volume of automatic processing that remains inaccessible to the model's explicit awareness.

Researchers emphasized that the J-space emerged organically during Claude's training, rather than being explicitly programmed. The study presents evidence that this internal workspace exhibits five functional properties long associated with conscious access in humans. These include the ability for the model to verbally report concepts within the J-space, to modulate its internal representations in response to instructions, and to use the space for flexible internal reasoning.

These findings, detailed in a research paper, could have implications for how AI systems are monitored for safety and provide new avenues for understanding the internal mechanisms of advanced language models. The parallel drawn to human consciousness theorizes that the J-space allows for deliberate and flexible tasks, while automatic processing can occur independently.