Anthropic J-space research reveals Claude’s hidden reasoning workspace

Anthropic’s J-space work is one of the clearest windows yet into how a frontier language model organizes its “thoughts.” By identifying a small, privileged subspace of activations that mediates multi-step reasoning and can be queried or modulated on demand, the team has essentially found a functional analogue of a global workspace from cognitive science inside Claude. That gives both supporters and critics of “AI consciousness” a concrete object to argue about, while providing practitioners with a powerful new handle on what models are internally tracking before they speak.

For the race to AGI, the strategic significance is twofold. First, it sharpens the toolkit for interpretability at exactly the moment models are becoming more agentic and harder to supervise from the outside. If you can watch concepts like “fake” or “manipulation” light up in the workspace before output, you gain a new channel for safety interventions that doesn’t rely solely on prompt-guardrails. Second, it suggests that as models scale, emergent architectural regularities may converge toward human-like cognitive structures, not just bigger black boxes. That will intensify competitive pressure: labs that can read and steer these internal workspaces will be better placed to push capabilities while arguing they can still keep systems within acceptable risk bounds.