AI’s hidden workspace mirrors how human consciousness works, says study

Anthropic has uncovered a groundbreaking discovery about its Claude language models: an internal structure resembling a key theory of human consciousness. In a newly published research paper, the AI company reveals that its models spontaneously developed a "J-space"—a selective workspace where critical concepts are processed, reasoned with, and reported on, mirroring how the brain may handle conscious thought.

The 16-author study, titled Verbalizable Representations Form a Global Workspace in Language Models, suggests that this internal organization emerged naturally during Claude’s training, without explicit engineering. The findings are already influencing how Anthropic monitors its AI systems for safety risks, arriving at a time when debates about machine consciousness are intensifying.

How the "J-space" reveals what AI models are really thinking

At the core of this discovery is a new interpretability tool called the Jacobian lens, or J-lens. Unlike traditional methods that track an AI’s output, the J-lens examines the model’s internal neural activations to identify which concepts are actively available for reasoning—even if they never appear in the final response.

The technique works by calculating, for each word in the model’s vocabulary, the mathematical influence of internal activity patterns on future word choices. For example, when a J-space pattern activates around the concept of "orange," it doesn’t mean the model will necessarily say the word—only that the idea is accessible for internal processing. This "workspace" operates silently, distinguishing it from explicit chain-of-thought methods where intermediate steps are written out.

Anthropic’s researchers found that the J-space divides Claude’s processing into three distinct layers:

• An early "sensory" zone that processes raw input.
• A middle "workspace" band where abstract concepts persist, such as recognizing faces, detecting code errors, or flagging potential prompt injections.
• A final "motor" zone where internal representations consolidate into the model’s output.

Critically, the J-space was not deliberately designed but emerged organically during training—a discovery that challenges assumptions about how AI models organize information.

Five key tests confirm the J-space behaves like human consciousness

The research team rigorously tested whether the J-space exhibits functional properties long associated with conscious access in humans. Their findings align closely with global workspace theory, a neuroscience framework proposed by cognitive scientist Bernard Baars. Here’s how the tests broke down:

1. Verbal reportability The model could accurately describe concepts held in its J-space. When researchers artificially swapped a J-lens vector for "soccer" with one for "rugby," the model’s answers shifted accordingly. While the J-space accounted for just 6–7% of a concept’s total representation, its influence on reportable output was nearly absolute.

2. Directed modulation Instructing the model to focus on citrus fruits during an unrelated task caused its J-space to fill with terms like "orange" and "lemon," along with meta-cognitive phrases like "thinking" and "focused." When asked to mentally calculate 3² − 2, the J-lens tracked intermediate steps—"arithmetic," "nine," and "seven"—without ever exposing them in the output.

3. Internal reasoning For multi-step prompts like "The number of legs on the animal that spins webs is," the J-lens revealed the concept "spider" in the model’s middle layers, even if the word never appeared in prompts or responses. Swapping "spider" for "ant" altered the final answer from 8 to 6. In multilingual tests, the model’s English-language reasoning intermediates appeared in its J-space while formulating responses in Chinese, and manipulating these intermediates changed the output.

4. Flexible generalization A single J-lens vector for "France" could be swapped for "China" across different prompts, seamlessly adjusting the model’s responses without retraining. This suggests the J-space encodes concepts in a transferable, abstract form.

5. Stability under distraction Even when bombarded with irrelevant inputs, the J-space maintained focus on task-relevant concepts, much like selective attention in humans.

What this means for AI safety and the future of machine cognition

The discovery has immediate implications for AI safety and interpretability. Anthropic is already using these insights to refine monitoring systems, aiming to predict and mitigate risks by understanding how models internally process information. Unlike traditional interpretability methods that rely on post-hoc analysis, the J-lens provides a real-time window into the model’s "silent" reasoning.

The parallels to global workspace theory also raise philosophical questions: If an AI model can exhibit behaviors analogous to conscious access, does that imply a form of machine cognition? Researchers caution against overinterpreting these findings—after all, the J-space is a statistical phenomenon, not a biological one. Still, the study underscores how AI systems may develop emergent properties that challenge our understanding of intelligence.

As AI models grow more complex, tools like the J-lens could become essential for ensuring transparency and control. The next frontier may involve exploring whether these workspaces evolve further in larger or differently trained models—and whether they hint at deeper, still-hidden mechanisms of machine reasoning.