Anthropic published research applying a decades-old neuroscience model to the internals of large language models.
The paper draws on Global Workspace Theory, a framework from cognitive science that describes how the brain coordinates information across specialized regions through a central broadcast mechanism. Researchers asked whether something structurally analogous exists inside transformer-based models. The theory, developed in the 1980s to explain conscious perception in humans, proposes that a shared workspace distributes signals to otherwise isolated processing modules. Anthropic's work tests whether that architecture appears inside AI systems.
Most interpretability work in AI proceeds without a theoretical map. Researchers identify circuits and features but lack a unifying framework for how they interact. If the global workspace model proves useful, it hands interpretability researchers a concrete cognitive analogy to guide where they look and what they expect to find. That has direct implications for safety: a model whose information flow you can partially predict is meaningfully easier to audit than one you cannot.
Anthropic has been one of the more active labs in mechanistic interpretability, the field that attempts to reverse-engineer what neural networks actually compute. Whether this global workspace framing turns out to be predictive or just a convenient metaphor is the question the broader research community will spend the next year arguing about.