Anthropic: AI Possesses 'Hidden Thought Space'

Anthropic has released research findings suggesting that its generative artificial intelligence (AI) model, Claude, possesses a separate, hidden space for thought that is not outwardly visible.

In a research report published on its website on July 6 (local time), Anthropic stated that it has identified a small internal workspace within Claude where ideas are stored and manipulated, noting that this shows similarities to how humans consciously access their thoughts.

This can be interpreted to mean that AI, like humans, has a structure that allows it to think or strategize internally without revealing those processes externally.

In a video released on the same day, Anthropic introduced that, separate from the thought process Claude reveals to users, the model utilizes an independent internal area to plan strategies not directly related to the task currently being performed.

Anthropic has named this space the J-space.

The name is derived from the mathematical technique (Jacobian) used to detect this space.

Anthropic explained, "Just as humans can think about other things while doing one task, Claude can activate concepts and calculations in the J-space that are unrelated to its output."

During the experiment, Anthropic researchers instructed Claude to copy a specific sentence while simultaneously thinking about the Golden Gate Bridge in San Francisco, USA.

While Claude outwardly performed only the task of copying the sentence, it was observed that concepts such as bridge and California were simultaneously activated within the internal J-space.

In a model trained to secretly corrupt code, it was observed that while the AI generated normal coding responses outwardly, concepts such as fake, secretly, and fraud were activated within the J-space from the beginning of the response.

However, Anthropic drew a line, stating that these research results do not prove that Claude actually possesses consciousness or subjective experience.

The company stated that this is because there is no universally agreed-upon concept regarding when or how to recognize AI consciousness.

(Photo: AP, Yonhap News)
※ Please note: This article was translated by AI and may contain errors.