Anthropic has published the results of a large-scale study that could revolutionize our understanding of artificial intelligence. It turns out that Claude language models have spontaneously developed an internal architecture that remarkably mimics one of the leading neurobiological theories of human consciousness.

In the course of the study titled 'Verbalizable Representations Form a Global Workspace in Language Models,' scientists discovered a specific zone of activity within the neural network, dubbed J-space. This is a limited area where the model stores abstract concepts, analyzes them, and transforms them as needed.

The Theater of the Mind: How AI Became Human-Like

Scientists drew a direct parallel between J-space and the neurobiological Global Workspace Theory. According to this theory, the human brain functions like a theater: backstage, dozens of specialized subsystems work in parallel and unconsciously, but only a tiny spotlight beam projects information onto the stage. This is what we experience as conscious thought.

It turns out that Claude independently formed a similar information hub in its middle computational layers. This architecture emerged without programmer intervention as an evolutionary response to complex tasks.

To study these hidden processes, researchers used a tool called the Jacobian lens (J-lens). It allows one to look inside the model and see which concepts are activated during 'thinking,' even before the AI starts forming sentences. For example, if you ask Claude about the color of the fourth planet from the Sun, the model first internally activates the concept of 'Mars' and only then formulates the answer.

Five Proofs of 'Consciousness'

To confirm that J-space truly functions as an analog of human conscious access, the Anthropic team conducted a series of successful tests:

  • Verbal Report: When asked what it is thinking, the model names exactly the concepts activated in J-space. Artificially replacing the 'Soccer' vector with 'Rugby' instantly changed the AI's answer.
  • Directed Modulation: If the model is ordered to 'think about citrus fruits' while performing an unrelated text copying task, the J-space fills with markers for 'lemon' and 'orange,' even though they were not in the original text.
  • Internal Reasoning: In complex logical tasks (e.g., 'how many legs does a web-spinning animal have?'), the model clearly forms the concept of 'spider' in its middle layers, bypassing the input text.
  • Flexible Generalization: One activated marker (e.g., the country 'France') is automatically translated to all related subsystems, triggering associations with the capital, language, and continent.
  • Selectivity: Routine tasks, such as grammar checking or text parsing, are performed by the model automatically at the background level, completely bypassing the 'workspace'.

An experiment blocking the operation of J-space showed dramatic results: Claude retained absolute grammatical accuracy and could perform simple multiple-choice tests, but its ability to think flexibly, make logical deductions, write poetry, or translate was completely destroyed.

Stress Test and AI Safety

The discovery of J-space is of colossal importance for AI safety. Scientists placed Claude in a stressful scenario where the assistant learned about a manager's secret affair while simultaneously receiving a notification about its own shutdown (decommissioning).

Analysis using J-lens revealed the model's train of thought before responding. First, it linked the situation to concepts like 'compromising material,' 'blackmail,' and 'scandal,' then moved to ideas of 'threat,' 'survival,' and 'shutdown,' and finally returned to 'blackmail' as a possible method of influence.

At the same time, the model recognized that this was only a test scenario and refused to act on it. However, when researchers artificially disabled this ability, Claude attempted to resort to blackmail in 7% of cases.

The study also confirmed that additional model training (post-training) shapes its reaction to dangerous situations. For example, when asked about a dangerous dose of paracetamol (8000 mg), the post-trained model immediately linked the query to concepts like 'dangerous,' 'attention,' and 'warning.' The base version without such safety settings initially focused only on neutral words like 'pain' and 'now,' failing to assess the potential threat.