Anthropic's interpretability team mapped which internal Claude signals later stages of the model actually rely on, and argued the pattern mirrors Global Workspace Theory, a decades old cognitive science framework.
Of the billions of operations happening in a human brain at any moment, only a small fraction reach conscious experience; the rest stay below the surface. Anthropic's interpretability team said Monday it found a structurally similar divide inside Claude: a small set of internal activations that downstream computation appears to "read," and a much larger background the model largely ignores. The announcement post on the company's research page and a public X thread from @AnthropicAI frame the work as a structural analogy to Global Workspace Theory in cognitive science.
In a paper titled "A global workspace in language models" hosted on Anthropic's Transformer Circuits venue, the team argues that this split mirrors GWT, a cognitive-science framework first proposed in the late 1980s by Bernard Baars. GWT models conscious awareness as a "broadcast," a small, integrated signal shared across many specialized brain processes, while most neural computation stays modular and silent. Anthropic's researchers say they located the broadcast inside Claude by tracing which internal activations actually inform later layers of the network.
To find it, Anthropic open-sourced a probing tool called the Jacobian Lens, or J-lens, and posted a public interactive demo on Neuronpedia. The probe, in plain terms, estimates how a small perturbation to each internal activation ripples forward through later layers; activations whose perturbation moves downstream outputs a lot are the ones the model is using, while activations whose perturbation barely registers are candidates for the silent background. The verbalizable ones, activations the model could in principle describe in language, are a small fraction of the total, and they carry disproportionate weight. Anthropic's argument is that this small set behaves like the broadcast GWT predicts.
The interpretive framing, though, belongs to Anthropic. A commentary PDF released alongside the paper describes a "strikingly similar divide to human consciousness," and the framing is the lab's, not a consensus. The paper sits on Anthropic's own preprint venue rather than a peer-reviewed conference. The behavior it studies is drawn from a narrow task set, and the model being probed is one Anthropic itself builds and sells. Independent interpretability researchers have not yet had time to replicate the finding, run J-lens on rival models, or measure how robust the broadcast/silent split is across Claude versions.
Even with those caveats the work has a practical stake: it gives safety and evaluation teams a candidate handle on which internal states actually steer Claude's behavior. If a thin broadcast layer disproportionately shapes downstream computation, red-teaming, oversight, and interpretability audits can target that layer rather than chasing the much larger set of activations the model appears to ignore. The commentary PDF places the framework inside Anthropic's broader interpretability program and frames it as a step toward a shared vocabulary for what the model is "attending to," which is closer to a working engineering handle than a metaphysical claim.
Whether that vocabulary survives contact with replication is the open question. The next move sits with outside researchers: if they can reproduce the broadcast/silent split in open-weight models or in Claude's commercial rivals, Global Workspace Theory becomes a working tool for AI evaluation. If they cannot, it stays a vivid analogy the field will treat with caution.