Claude processes information on at least two distinct levels: a broad, implicit layer and a focused, explicit layer that mirrors what neuroscientists call the global workspace, the brain region where thoughts become consciously accessible.
Anthropic ran experiments directly inspired by global workspace theory and found a structurally similar divide inside Claude's neural representations. The parallel is not metaphor. It is a measurable architectural feature of how the model organizes information internally.
The reason to read this in full is not the conclusion but the methodology: how Anthropic identified these representation layers, what probing techniques they used, and what this reveals about the gap between what Claude computes and what Claude can report about its own reasoning.
[WATCH ON YOUTUBE →]