Anthropic Discovers Internal Structure Resembling 'Consciousness' in Claude

Anthropic has revealed that a functional structure common to a leading neuroscience theory of human consciousness naturally emerged within its AI model 'Claude'. This achievement was compiled in a paper titled 'Verifiable Representations Form a Global Workspace in Language Models' by 16 researchers and was publicly released by the company on Sunday.

The question of whether AI is 'thinking' has been a long-debated topic among researchers. In conventional AI research, the internal processing of models has often been treated as a black box, leaving few clues as to what constitutes 'conscious processing'. This research is positioned as the first attempt to tackle such questions with concrete observational means.

The research team focused on the 'Global Workspace Theory' proposed by neuroscientist Bernard Baars. According to this theory, the brain functions like a theater. While many specialized processes occur in parallel backstage, only a small portion of that information is spotlit and shared globally, which is then experienced as 'conscious thinking'. Anthropic's research team observed a structure similar to this existing within Claude.

To discover this structure, the research team developed a new mathematical technique called the 'Jacobian Lens (J-lens)'. This technique calculates how the internal activation patterns of a model influence which words it will output in the future. What was identified through this method is a region called 'J-space (J-space)', where a small privileged zone containing concepts the model can reference and manipulate is clearly distinguished from a vast amount of automatic processing that the model cannot access or verbalize.

What is crucial is that this J-space was not intentionally incorporated at the design stage. According to the research team, this structure naturally emerged within Claude's learning process. Additionally, unlike the chain-of-thought that the model outputs as text, J-space quietly functions as internal neural activation and can retain concepts without being written out. After analyzing Claude's computational layers using J-lens, the research team also confirmed that processing is divided into three different regions, including an 'sensory' initial zone.

This discovery is directly linked to Anthropic's AI safety efforts. The company explains that this research has already begun to change how it monitors safety risks in AI systems. By becoming able to observe the model's internal states with greater precision than before, there is an expanding possibility of detecting unexpected behaviors and potential risks at an earlier stage.

The significance of this research extends beyond questions of AI capability. The question 'Does AI possess consciousness?' has transcended philosophical domains and begun to hold practical importance from perspectives of safety, control, and ethics. If large-scale language models like Claude spontaneously develop functional structures corresponding to human consciousness theory, the question of how to interpret and handle AI's internal states will occupy an increasingly important position in future research and discussion.