Claude's Hidden Mind: Anthropic Discovers an Internal Workspace That Emerged Spontaneously

Anthropic's interpretability team found that Claude developed its own internal neural workspace — a "global workspace" for silent reasoning — without any deliberate design or programming.

Anthropic's interpretability team has discovered something remarkable inside Claude: a small collection of neural patterns that function like a cognitive global workspace. They call it the J-space, named after the Jacobian mathematical technique used to find it.

The J-space is not something Anthropic designed or programmed. It emerged on its own during training. Each pattern in this space is linked to a word, but when one lights up, it does not mean Claude is about to say that word — it means the word is on its mind, like a silent thought.

The researchers found the J-space has several unique properties. Claude can report on what it is thinking by reading its own J-space patterns. It can also modulate them on command — if you ask it to think about something silently, the corresponding patterns activate. And when Claude solves multi-step problems, intermediate steps light up in the J-space even when never written down, and these patterns causally affect its performance.

Perhaps most striking: when the team suppressed the J-space, Claude still spoke fluently, used correct grammar, and recalled simple facts. But it lost its higher-order cognitive functions — the kind of deliberate, multi-step reasoning that separates basic language ability from actual problem-solving.

The research draws a direct parallel to neuroscience. Humans have consciously accessible brain activity we can describe, control, and use for deliberate reasoning, distinct from all the unconscious processing happening below awareness. Claude, it turns out, developed something structurally similar — entirely on its own.

The findings raise profound questions. If a neural workspace emerges spontaneously in language models trained simply to predict text, what does that say about the relationship between language and consciousness? And if Claude already has a workspace it cannot write down but can think with, what else might be happening inside these models that we have not yet found?

Sources: Anthropic Research, Full Paper — Transformer Circuits