Anthropic finds neural patterns in Claude that resemble a 'global workspace'

On July 6, 2026, Anthropic published the study "Verbalizable Representations Form a Global Workspace in Language Models" on its transformer-circuits.pub site. The paper argues that language models like Claude develop a structure that resembles what neuroscientists call the global workspace of consciousness.

What Anthropic found

The authors start from an analogy with the human brain. While you read a sentence, your brain runs dozens of processes in parallel without you noticing: it holds your posture, regulates your breathing, and turns lines on a screen into recognizable words. But some of this activity does reach awareness — an image that pops into your head, a decision you think through, where you direct your attention. Neuroscientists call this activity consciously accessible and distinguish it from automatic processes. It has special properties: you can describe it in words, control it, and reason with it.

• Publication date — July 6, 2026
• Publisher — Anthropic, developer of the Claude model family
• Title — Verbalizable Representations Form a Global Workspace in Language Models
• Venue — transformer-circuits.pub, Anthropic's research site for model interpretability

According to the study, language models show a similar split. The researchers found a small set of internal neural patterns that play a special role relative to everything else the model does: they can be expressed in words (verbalizable). The far larger volume of internal computation stays entirely "unspeakable" and runs on its own, much like the brain's automatic processes.

Why it matters for AI interpretability

Anthropic has long studied the internals of its models through its mechanistic interpretability program — an effort to understand what actually happens inside a neural network between the input text and the output, rather than judging the model only by its answers from the outside. Finding a structure analogous to a global workspace gives researchers a new tool. If part of a model's internal representations really works as a "reportable" subsystem, that opens a more precise way to check what the model actually "knows" and can report about its own internal states, and what stays hidden even from the model itself.

The distinction matters more as language models are increasingly used in settings where they must not just give an answer but explain their reasoning — for example, in chain-of-thought reasoning or when a user asks the model why it made a particular decision.

What it means

The study does not claim that language models are conscious in the human sense. It shows a structural analogy: a small, verifiable set of internal representations separated from the rest of the automatic processing. That is a meaningful step for AI interpretability. Knowing which parts of a model's internal processes can, in principle, be checked and verbally reported — and which cannot — directly affects how far we can trust the explanations a model gives about its own reasoning.

Frequently asked questions

When did

Anthropic publish its study on the global workspace in LLMs?

Anthropic published "Verbalizable Representations Form a Global Workspace in Language Models" on July 6, 2026, on transformer-circuits.pub.

What is a global workspace in the context of language models?

It is a small set of a model's internal neural patterns that can be expressed in words and controlled. The analogy is to the part of the human brain's activity that is consciously accessible, as opposed to automatic processes that run without attention.