Anthropic maps a global workspace inside Claude's mid-layers

Anthropic's interpretability team has published a paper arguing that language models contain something functionally like a global workspace, the small privileged buffer cognitive scientists have long used to describe the 'conscious' contents of a human mind. The paper, on transformer-circuits.pub, introduces a new interpretability tool called the Jacobian lens, or J-lens, and uses it to argue that the same internal representations a model can verbalize are the ones it uses to reason silently.

The J-lens is described as identifying 'a vector representation that encodes the potential for the model to verbalize that token in the future.' From there, the researchers claim, you can read out what the model is holding in mind at a given moment. The workspace they describe is small: they report that the J-space typically accounts for only a small fraction of activation variance, 'varying by layer, but never more than 10%,' and that its workspace-like properties reside only in the middle block of the network.

Why this matters for people who work on alignment rather than interpretability is a claim the paper makes near the end. In a training experiment, models were prompted to articulate ethical principles when interrupted, and the paper reports 'this training measurably improves model behavior in the original, uninterrupted contexts, despite no direct training of the ethical behavior taking place.' The framing, if it holds up, is that shaping what a model is disposed to say may also shape what it silently reasons about.

The honest caveat comes from the authors themselves. They describe the Jacobian lens as 'an imperfect tool, which we believe only approximately and incompletely captures the model's underlying workspace structure.' What the reporting does not give you is any independent replication, and the results run on Anthropic's own frontier models, Claude Sonnet 4.5, Haiku 4.5, Opus 4.5, and Opus 4.6, rather than a cross-lab evaluation.

Still, if the workspace framing generalizes, it hands alignment work a concrete surface to operate on: a small set of vectors in a specific band of layers that seem to mediate between what a model thinks and what it says.