Breaking

Anthropic Finds Brain-Like 'Global Workspace' Inside Claude

July 6, 2026 at 13:35 EDT

Research & Papers
Foundation Models
Security

On July 6, 2026, Anthropic published interpretability research arguing that a privileged internal representation space functionally resembling the human brain's "global workspace" exists inside its large language model Claude. The paper "Verbalizable Representations Form a Global Workspace in Language Models" was released alongside an official blog post and the full paper.

July 6, 2026 · Anthropic Research

Inside Claude, a "Conscious-Like" Workspace Emerges

Anthropic reports a privileged set of internal representations — dubbed "J-space" — that mirrors Global Workspace Theory: a tiny region acting as a broadcast hub for the model's verbalizable thought, not designed but emerged during training.

J-space

A small "broadcast hub" evolving across layers

Defining properties of the workspace

J-lens

Jacobian technique that isolated it

A tiny slice carries the conscious-like load

Like the human brain, only a fraction of Claude's total activations is verbalizable — yet that fraction relays information across the whole model.

Total activations

purple = J-space

J-space only

verbalizable hub

J-space intact

Fluent language
Simple fact recall
Multi-step reasoning works

J-space disabled

Fluency preserved
Simple recall preserved
Higher-order reasoning degrades

Global Workspace Theory, applied to a model

Specialized modules process in parallel; access arises when information is broadcast to a limited workspace.

Parallel
modules

→

Broadcast to
J-space hub

→

Widely relayed
access

Why it matters: an alignment-auditing tool

Detect if a model treats a blackmail scenario as fictional

Capture silent recognition of prompt injections

Surface hidden goals and "silent reasoning"

Limits acknowledged: J-space can't be controlled perfectly, forbidden thoughts can backfire, and it isn't involved in automatic processing. Anthropic frames it as access consciousness only — not phenomenal, subjective experience.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.