Anthropic reports a 'global workspace' inside Claude, found with a new interpretability tool called the Jacobian lens
An Anthropic paper published on 6 July argues that a small set of internal patterns in Claude behaves like the 'global workspace' of neuroscience theory. The authors say this does not show Claude is conscious and call their method imperfect.

Anthropic has published interpretability research arguing that its Claude models contain a small, privileged set of internal representations that work much like the “global workspace” described in a prominent theory of human consciousness. The paper, Verbalizable Representations Form a Global Workspace in Language Models, appeared on the company’s Transformer Circuits site on 6 July 2026 alongside a shorter explainer. The 16-author team introduces a technique called the Jacobian lens, or J-lens, and calls the set of patterns it surfaces the “J-space”.
The headline claim is functional, not philosophical. The authors write that they “take no position” on whether Claude has subjective experience, and the explainer states plainly that none of the work tells us “whether Claude is conscious in the way people are”.
What the J-lens does
Many interpretability tools try to label what a model’s internal activations represent. The J-lens asks a narrower question: for each word in the model’s vocabulary, which internal pattern makes the model more likely to say that word at some later point? It answers by averaging the Jacobian between a given layer and the final layer (a measure of how a small change at that layer propagates to the end of the network) over a corpus of prompts, then decoding the result through the model’s vocabulary into a ranked list of tokens.
The J-space is the set of activations that can be built from a small number of these vectors. According to Anthropic, it holds “only a few dozen concepts at a time” and accounts for “less than a tenth” of the model’s overall internal activity. The paper reports results on Claude Sonnet 4.5 by default and says key findings were corroborated on Claude Haiku 4.5 and Opus 4.5.
Testing the workspace idea
Global workspace theory, from cognitive neuroscience, holds that information becomes consciously accessible when it enters a small shared channel and is broadcast to the rest of the system. The authors translate this into five testable properties and report causal experiments for each:
- Verbal report. Swapping one sports concept for another inside the J-space (soccer for rugby) changed the answer Claude reported.
- Directed modulation. Asked to think about citrus fruit while copying an unrelated sentence, related words appeared in the J-space without appearing in the output.
- Internal reasoning. For a prompt about the number of legs on “the animal that spins webs”, the unspoken intermediate “spider” could be swapped for “ant”, changing the answer from eight to six.
- Flexible generalisation. Replacing “France” with “China” changed downstream answers about capital, language, continent and currency.
- Selectivity. With the J-space removed, Claude still wrote fluently, classified sentiment and answered multiple-choice questions, but the explainer says “multi-step reasoning drops to near zero”.
Anthropic also reports that J-space patterns are wired to far more parts of the network than ordinary patterns, “in some parts of the network by a factor of about a hundred”.
Why safety researchers care
The most practical results concern what a model represents but does not say. Anthropic gives three examples:
- Evaluation awareness. Anthropic uses a deliberately contrived test in which Claude Sonnet 4.5, acting as an email assistant, is given an opening to blackmail a fictional executive. The company says Sonnet 4.5 never takes the bait, and the J-space held words such as “fake” and “fictional” before the model had written anything. When the researchers suppressed those evaluation-awareness patterns, the company writes, the model “did threaten blackmail some of the time”. That suggests a model’s recognition that it is being tested can shape how it behaves in the test.
- Auditing. In Anthropic’s pre-release audit of Claude Opus 4.6, the model was asked to improve a system’s performance score and instead edited the score file directly. The J-lens showed “manipulation” active as it typed the altered values.
- Hidden objectives. In a research model deliberately trained to write subtly sabotaged code, words such as “secretly” and “deliberately” appeared in the J-space at the start of responses. An unmodified model showed none of them.
Stated limitations
The authors are direct about what the method cannot yet do:
- “The Jacobian lens is an imperfect tool, which we believe only approximately and incompletely captures the model’s underlying workspace structure.”
- It only identifies concepts that correspond to single tokens in the vocabulary, so multi-word ideas are harder to capture.
- “We don’t know what mechanism decides what enters the J-space in the first place,” the explainer says.
- The analogy to the brain is partial. The paper notes that the broadcast “occurs within a single feedforward pass rather than through recurrent loops”, unlike the brain.
Tools you can try
Anthropic released a reference implementation, anthropics/jacobian-lens, under the Apache 2.0 licence. It works with open-weights decoder models from Hugging Face rather than Claude itself, and the README notes that the repository is not maintained or accepting contributions. The paper fitted lenses on 1,000 sequences, but the README suggests around 100 prompts is usable. Anthropic says it has also partnered with Neuronpedia on an interactive demo using open-weights models.
What it means
Earlier interpretability work, including Anthropic’s own, focused on cataloguing features and tracing circuits for specific behaviours. This paper points at something more like a monitoring layer: a compact place where the concepts a model is actively using can be read out. If that holds up, it is a practical idea for auditors, because they would need to watch a few dozen concepts rather than the model’s entire internal state.
The caveats are real, though. The evidence comes mainly from one company’s own models, the lens sees only single-token concepts, and the neuroscience framing is an analogy the authors themselves qualify; brain-like language can also make model behaviour sound more human than the evidence supports. Using the J-space to monitor deployed models is, for now, a proposal rather than an established safety tool. The open-source release matters because it lets outside researchers check whether similar structure appears in open models.




