Inside LLM 'private thoughts': J-Space isn't consciousness, it's control
Researchers spot J-Space, an internal workspace in transformers guiding reasoning. Not consciousness, but a practical clue for AI developers and safety.
New analyses suggest large language models contain a small, relatively organized internal region that temporarily holds concepts and shapes the reasoning chain — dubbed J-Space, evoking cognitive science's Global Workspace Theory. This isn't evidence of consciousness, but rather an emergent control and planning mechanism that arises spontaneously during training.
Experiments show that swapping an active concept in this workspace (say, 'spider' for 'ant') without touching the prompt or output changes the model's downstream reasoning accordingly. Similarly, relabeling a detected language can make the model misidentify it while still generating flawless text in the original language — suggesting meta-judgment and surface generation run through separate pathways.
The most striking finding: disabling this internal region lets the model keep talking fluently and confidently while losing much of its reasoning capability. That's a critical warning for builders — fluency is not a reliable signal of correctness. Teams should instrument for process, not just output: testing robustness against misleading premises and evaluating automatic fluency separately from deliberative reasoning.
The discovery also points to a new attack surface for safety and alignment: manipulating internal representations could become a threat class beyond classic prompt injection. Overall, this is less about consciousness and more about concrete tools for interpretability, alignment, and evaluation of AI systems.