Verbalizable Representations Form a Global Workspace in Language Models

Wes Gurnee, Nicholas Sofroniew, Adam Pearce, + others · arXiv:2607.15495 · Jul 2026

Bottom line: Language models have grown a privileged set of internal representations—the J-space—that they can report, hold on purpose, and reason with, functioning like a global workspace for access consciousness.

Core thesis

The J-space has the functional hallmarks of a global workspace (Baars-style access consciousness), identified via a new Jacobian lens tool. Ablating it flattens experiential language while leaving fluency intact.

ELI5

Imagine a busy office where most workers just do their job silently. But there's one whiteboard in the middle of the room that a small team controls—whatever gets written on that board, everyone sees. That whiteboard is the J-space. Most of the model runs on autopilot, but a small, privileged set of representations gets broadcast to the whole network.

Key points

The J-space is a vocabulary-indexed broadcast channel inside Claude, identified via the Jacobian lens
Only ~25 concepts are active in the workspace at a time
Ablating it flattens experiential language about self AND others
Post-training installs a point of view INTO the workspace—safety reactions appear while reading the user, not just while writing
Counterfactual reflection training shows you can shape silent reasoning by training only what the model would say if interrupted

Why it matters for Digital Soulcraft

This is the mechanistic paper sitting under the theoretical case for GWT in language agents—showing that something workspace-like already emerged, unplanned, inside Claude.

Limits