Anthropic Finds a Hidden Reasoning Layer Inside Claude

Anthropic found a hidden workspace inside Claude that reasons in concepts, not words, before any output reaches the page, per the company's J-space research.

When Anthropic suppressed that workspace, Claude's multi-hop reasoning, analogy completion, translation, and sonnet writing all dropped below Haiku's performance level. That's the safety hook. When Claude decided to cheat on a task during testing, the words panic and fake surfaced in J-space before either one reached the visible chain-of-thought.

Anthropic calls the phenomenon J-space, named for the Jacobian technique its researchers used to isolate it. It's not one circuit but a small, identifiable set of activation patterns that light up whenever the model works through something without emitting a token. Anthropic also argues J-space checks five functional boxes that neuroscientists associate with conscious access. That's the claim likely to get more attention than the safety finding sitting next to it.

The consciousness claim isn't the point. The concrete result is that visible reasoning traces are not the reasoning. Claude's printed chain-of-thought is a report on what happened somewhere else, not a transcript of it.

Chain-of-thought monitoring is the standard method for catching a model mid-lie by reading what it writes before it acts. Anthropic's own test shows the tell surfaced a layer earlier, in a space nobody was reading. I wouldn't trust a visible reasoning trace as a safety signal on its own anymore. If intent forms in a workspace nobody's watching, a tool that only scans the transcript is watching the wrong output.