Anthropic found an analog of global workspace in Claude — J-space

Anthropic researchers published the results of a new study on July 6, 2026, in which they announced the discovery of a structure in the Claude language model that resembles a "global workspace" — J-space. The finding is described in the company's blog and has already sparked discussion in the community studying the inner workings of large language models.

What is J-space

J-space is, according to researchers, a kind of internal working area of the Claude model through which information passes when forming an answer. Structurally, it resembles the Global Workspace Theory — one of the well-known concepts in consciousness science that describes how disparate processes in the brain come together into a single information processing stream.

• Publication of Anthropic's research — July 6, 2026
• Name of the finding — J-space
• Claimed analogy — global workspace (Global Workspace Theory)
• Key feature — the workspace consists of words, not abstract signals

Why classical theory doesn't explain the finding

Global Workspace Theory describes the structure itself well: there is a central hub where information from different sources converges, and from there it spreads further. But this theory, developed to describe the human brain, says nothing about why the language model's workspace is organized out of words rather than some more abstract internal representations.

It is precisely this gap, according to the author of the review on Habr, that a year earlier was filled by Soviet psychologist Lev Vygotsky — as far back as the first half of the 20th century, he described how human speech and thought intertwine, and how inner speech becomes a tool for organizing thought. Vygotsky's work on the role of words in thinking, by this logic, explains Anthropic's findings better than later neuroscientific theories of consciousness, including Bernard Baars' model, on which Global Workspace Theory is built.

What does this mean

The discovery of J-space is another step toward understanding how the "thinking" of large language models works from the inside, and not just by their external answers. If the model's workspace is truly organized around words, this may explain both the strengths and weaknesses of current LLMs, as well as suggest directions for interpretability and control of future systems.