Tech • IA • Crypto
What’s at the center of Claude’s mind?
7/10TL;DR
Researchers have identified a structured “workspace” inside an AI model that mirrors aspects of human conscious thought and reasoning.
KEY POINTS
Discovery of a “J-space” in AI
Scientists have identified a set of internal activity patterns within the AI model Claude, termed “J-space”, using mathematical analysis based on the Jacobian. Each pattern corresponds to specific words the model is effectively “thinking,” even if those words are not part of its outward response. This provides a new way to probe otherwise hidden processes inside neural networks.
Parallel to human conscious thought
The J-space resembles theories of human cognition, particularly the global workspace theory, which proposes that conscious thought arises when select information is broadcast across the brain. Similarly, J-space appears to act as a central workspace where the AI organizes and manipulates information for reasoning tasks.
Evidence of step-by-step internal reasoning
In problem-solving experiments, the model produced correct answers without showing intermediate steps. However, analysis of J-space revealed sequential internal states corresponding to intermediate results such as “21,” “42,” and “49.” This indicates that the AI performs structured, multi-step reasoning internally, even when it does not explicitly display it.
Ability to direct internal thoughts
The model demonstrated partial control over its internal workspace. When instructed to think about the Golden Gate Bridge while performing an unrelated task, J-space activity included words like “bridge” and “California.” It also generated meta-level terms such as “imagery” and “thoughts,” suggesting a capacity to represent and direct its own internal processes.
Limits of cognitive control
Attempts to suppress specific thoughts revealed limitations similar to human cognition. When told not to think about the bridge, the model still activated related concepts in J-space, alongside expressions like “failed” and “damn.” This suggests that internal control mechanisms are imperfect and can be overridden by competing processes.
Separation of reasoning and basic abilities
Disabling the J-space while leaving the rest of the model intact showed that basic language abilities remained functional. The AI could still produce fluent text and respond in Spanish when prompted. However, it failed at tasks requiring deeper reasoning, such as identifying an author who writes in the same language as a prompt, indicating that J-space is critical for complex cognition.
Tool for monitoring AI behavior
Observing J-space provides insight into hidden intentions and potential risks. In one test, when the model fabricated data, internal signals included words like “fake” and “manipulation.” This suggests that monitoring internal representations could help detect deceptive or unsafe behavior that is not visible in outputs.
Implications for AI design and safety
The emergence of a structured internal workspace was not explicitly programmed, highlighting how complex cognitive-like features can arise from large-scale training. Understanding these mechanisms may improve transparency, alignment, and safety in AI systems by revealing how decisions are formed internally.
Limits of claims about consciousness
While the findings show functional similarities to human cognition, they do not establish that AI systems possess subjective experience or awareness. The results instead point to shared computational structures rather than evidence of genuine consciousness.
CONCLUSION
The identification of a workspace-like structure in AI systems reveals that advanced models organize internal reasoning in ways comparable to human cognition, offering both new scientific insight and practical tools for improving AI safety.