Anthropic Found Neural Patterns in Claude Similar to a \'Global Workspace\'

Anthropic published a study on July 6, 2026 on its transformer-circuits.pub website titled "Verbalizable Representations Form a Global Workspace in Language Models" — about how language models like Claude form a structure structurally similar to what neuroscientists call the global workspace of consciousness.

What Anthropic discovered

The authors of the study draw from an analogy with the human brain. While a human reads a sentence, the brain simultaneously performs dozens of processes that the person doesn't notice: maintains posture, regulates breathing, converts lines on a screen into recognizable words. But some of this activity is still accessible to consciousness — a surfaced image, a considered decision, direction of attention. Neuroscientists call such activity consciously accessible and distinguish it from automatic processes: it has special properties — it can be described in words, it can be managed, and reasoning can be based on it.

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

According to the study, similar division was found in language models: the model found a small set of internal neural patterns that, against the backdrop of all other processing, play a special role — they can be expressed in words (verbalizable), unlike a much larger volume of internal computations that remain completely "non-verbalizable" and flow on their own, analogously to automatic processes in the brain.

Why this matters for AI interpretability

Anthropichas long studied the internal structure of its models as part of a mechanistic interpretability program — attempts to understand what exactly happens inside a neural network between input text and output response, rather than only evaluating the model by its responses from outside. Discovery of a structure analogous to a global workspace gives researchers a new tool: if part of the model's internal representations is indeed structured as a "reportable" subsystem, this opens a more precise way to verify what the model actually "knows" and can report about its own internal states, and what remains hidden even from the model itself. Such a distinction is especially important given that language models are increasingly used in scenarios where they need not just provide an answer, but explain the course of their reasoning — for example, in chains of thought (chain-of-thought) or when asking a model to explain why it made a particular decision.

What this means

The study does not claim that language models have consciousness in the human sense — it shows a structural analogy: a small verifiable set of internal representations separated from the rest of automatic processing. For AI interpretability, this is a significant step: understanding which part of the model's internal processes are in principle accessible for verification and verbal report, and which are not, directly affects how much you can trust the explanations the model gives about its own reasoning.

Frequently Asked Questions

When was

Anthropic's study on global workspace in LLMs released?

The study "Verbalizable Representations Form a Global Workspace in Language Models" was published by Anthropic on July 6, 2026 on the website transformer-circuits.pub.

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

This is a small set of internal neural patterns of a model that can be expressed in words and that can be managed — by analogy with the part of human brain function that is consciously accessible, as opposed to automatic processes that occur without attention.