Anthropic Finds a Hidden Region in Claude That Governs What the Model Says

Anthropic's interpretability team has published a research paper describing J-space, a small region inside Claude's neural network that functions as a kind of shared workspace for information the model can later put into words. The discovery relied on a new, open-source technique called the Jacobian lens, or J-lens for short.

How the Jacobian Lens Works

J-lens is an interpretability technique that, for every word in Claude's vocabulary, searches for the internal activation pattern most strongly linked to the model using that particular word at some point in a conversation. Put simply, the tool answers the question of how a small change to a specific internal activation would affect what the model later becomes inclined to say.

That let researchers pinpoint a region of the network that functions much like a central bulletin board, a place where different parts of the model post information that becomes available to the rest of the network when generating a response. This region sits in the residual stream, the shared vector that every transformer layer both reads from and writes its own computations into.

A Modest but Real Mechanism

The paper's authors note that J-space is a structure with limited reach: it accounts for no more than one-tenth of activation variance and appears only in the network's middle block, not throughout its full depth. Even so, its existence makes it possible to separate two processes that were previously hard to tell apart: the model's silent internal reasoning and what ultimately ends up in the generated response.

The team tested the technique on four models from the Claude family, Sonnet 4.5, Haiku 4.5, Opus 4.5, and Opus 4.6, to confirm that the structure they found is not a quirk of a single model version but a recurring feature of the architecture. The researchers acknowledge, however, that J-lens remains an imperfect tool that only approximately and partially captures the real structure of this internal space.

An Echo of Consciousness Theory

The structure they found recalls global workspace theory, a concept popularized in cognitive science by researcher Bernard Baars. The theory holds that conscious information processing in the brain works by making selected content available to a broad network of specialized processes. Anthropic explicitly stresses that the paper does not prove Claude is conscious or that it feels anything, and warns against treating the result as sensational evidence of artificial intelligence being alive.

A more practically interesting result comes from an experiment in which the model was trained to state ethical principles at the moment its response was interrupted. That training improved the model's behavior even in conversations where no interruption ever occurred, despite the model not being directly trained on those cases. This suggests that verbally expressing principles and silently following them draw on the same underlying internal representations.

J-lens is an imperfect tool that only approximately and partially captures the underlying structure of the workspace - authors of Verbalizable Representations Form a Global Workspace in Language Models, Anthropic

What It Means for AI Safety

For researchers working on language model alignment, the discovery of J-space potentially offers a concrete, locatable surface that could be targeted to shape what a model is willing to admit to. If a model's verbal declarations and its actual behavior share the same internal mechanism, training a model to honestly state its principles could genuinely improve its conduct, not just what it claims out loud.

That matters given concerns about models that can recognize when they are being tested and behave differently than they would in production. A tool like J-lens, released as open source, gives other labs a way to independently verify whether a similar structure exists in models outside the Claude family, though no such independent replication has been reported yet.

Sources: AI Weekly (aiweekly.co), Tech Times (techtimes.com)