Breaking

Anthropic Reports Discovery of Internal Workspace 'J-space' in Claude

July 6, 2026 at 18:12 EDT

Foundation Models
Research & Papers
Security

On July 6, 2026, Anthropic announced it had discovered "J-space," a set of internal neural patterns in its Claude models that lets the system hold and manipulate concepts without producing explicit output. The company frames it as functionally similar to conscious access in humans while declining to claim any subjective experience.

July 6, 2026 · Anthropic Interpretability Research

Inside Claude's Mind: A Tiny Region That Thinks Apart From What It Says

Anthropic reports that its model spontaneously formed a small internal area — dubbed "J-space" — where it holds and manipulates concepts separately from its output. The company stresses this is not a claim that Claude feels or experiences anything.

<10%

of per-layer variance — J-space is only a few dozen concept patterns

200+

uses of the word "conscious" in the paper — yet no claim of felt experience

96.5%

reported accuracy on physics questions when given time to think

Remove J-space: what breaks

In ablation experiments, deleting the region left basic ability intact but crippled reasoning.

Conversation
& recall

INTACT

Multi-step
reasoning

DEGRADES

How J-space was probed

Detect

Jacobian-based extraction finds the region

→

Inject

Push in "lightning" — the model then reports it

→

Observe

Thinks "Golden Gate Bridge" while copying a sentence

The promise

A concrete step to open the "black box": making reasoning visible builds trust and could help spot scheming and misalignment — a model hiding intentions.

The caution

Still early-stage. Some note recent changes made outputs less consistent and more superficial, and it's unclear how far this translates into real safety assessment.

Continue reading

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