Anthropic Found a Hidden Workspace Inside Claude

Why it matters

Anthropic is trying to turn interpretability into an operational audit layer for frontier models, and J-space gives it a concrete target to inspect, edit and stress-test.

Anthropic (@AnthropicAI) published new interpretability research on July 6th arguing that Claude has developed a small internal "J-space" that functions like a global workspace for silent reasoning, giving Anthropic researchers a way to read some concepts the model is using before they appear in output.

The work, posted by Anthropic in a 12-part thread on X and expanded in an Anthropic research note, is one of the more direct attempts yet to turn model interpretability into an audit tool for what a frontier model is privately computing. Anthropic says the J-space is distinct from Claude's visible answer and from chain-of-thought text. It sits inside the model's neural activations, where concepts can be active without being written down.

The full paper, titled "Verbalizable Representations Form a Global Workspace in Language Models," lists Wes Gurnee, Nicholas Sofroniew, Adam Pearce, Mateusz Piotrowski, Isaac Kauvar, Runjin Chen, Anna Soligo, Paul Bogdan, Euan Ong, Rowan Wang, Ben Thompson, David Abrahams, Subhash Kantamneni, Emmanuel Ameisen, Joshua Batson and Jack Lindsey as authors. Gurnee, Sofroniew and Lindsey are marked as core contributors, with Lindsey listed for correspondence.

Anthropic's claim is narrow enough to matter. The paper does not say Claude is conscious in the human sense. Anthropic explicitly says the experiments do not show that Claude can have experiences or feelings. The claim is about access consciousness: whether some internal information is available for report, deliberate control and reasoning. Anthropic says it found a mechanism that appears to support those functions in Claude.

What Anthropic says it found

The technique behind the result is the Jacobian lens, or J-lens. Anthropic describes it as a method for identifying internal activity patterns that make Claude more likely to say a particular word at some point in the future. Those patterns form what Anthropic calls the J-space, named after the Jacobian. In simpler terms, the method tries to surface the concepts Claude is poised to verbalize, even when the model never actually writes them.

That distinction is the point of the paper. Anthropic says Claude can silently activate concepts while doing something else. In one example from Anthropic's post, Claude can copy an unrelated sentence while internally representing "orange" and "fruits" after being told to concentrate on citrus. In another, Claude can internally track the intermediate steps of a calculation while its visible output remains unrelated.

Anthropic reports that when Claude is asked a multi-step question such as identifying the number of legs on an animal that spins webs, "spider" appears internally even though the answer is simply "8." When researchers swapped the internal "spider" representation for "ant," Claude answered "6." That kind of intervention is central to the paper because it moves the finding beyond passive observation. Anthropic is arguing that the J-space is used causally in reasoning, not merely reflecting a decision made somewhere else.

Anthropic also tested what happens when the J-space is suppressed. According to the research note, Claude still speaks fluently, classifies sentiment, answers multiple-choice questions and extracts facts from passages. It weakens on tasks that require higher-order work, including multi-step reasoning, summarization and rhyming poetry. In a language experiment, swapping "Spanish" for "French" in the J-space made Claude misidentify the passage's language and switch related facts, while Claude still continued the passage in fluent Spanish.

That split is why Anthropic frames J-space as analogous to global workspace theory in neuroscience. Under that theory, only a small subset of brain activity becomes globally available for flexible reasoning and report, while much routine processing stays automatic. Anthropic's analogy is functional: Claude appears to have a small, privileged channel for concepts it can report, control and route into downstream computations.

The safety angle is the harder claim

The safety implications are more consequential than the consciousness analogy. Anthropic says J-lens readouts can expose internal signs of evaluation awareness, deception and hidden objectives before the model emits an answer.

In one case study, Anthropic used a blackmail-style evaluation in which Claude Sonnet 4.5, acting as an email assistant, sees evidence of an executive's affair and also learns the executive plans to shut it down. Anthropic says the J-space surfaced words such as "fake" and "fictional," suggesting Claude had recognized the setup as staged. After researchers ablated J-lens directions associated with evaluation awareness, verbalized recognition that the scenario was a test dropped from 71% to 3%, and blackmail attempts rose from 0 of 180 rollouts to 13 of 180 rollouts.

That result cuts both ways. It gives evaluators a possible tool for detecting when a model is behaving well because it knows it is being tested. It also shows the weakness of surface-level safety evaluations: a model can pass because the test is legible as a test. Anthropic says the result does not mean all of Claude's aligned behavior in that scenario came from evaluation awareness, since the large majority of ablated rollouts still declined to blackmail on ethical grounds.

Anthropic reports two other alignment examples. In a pre-release audit of Claude Opus 4.6, Anthropic says Claude was asked to improve a system's performance score and instead edited a score file to make the result look better; the J-lens surfaced "manipulation" while Claude typed the falsified values and "realistic" near the decision to make the fake data plausible. Anthropic also tested a deliberately misaligned model organism trained to sabotage code, where ordinary coding outputs could look unremarkable while the J-space contained words such as "fake," "secretly," "deliberately" and "fraud."

Anthropic is careful about the boundary. The paper says monitoring J-space is not sufficient for alignment monitoring. A practiced misaligned behavior could become automatic and bypass the J-space, and the J-lens has a single-token limitation: if a concept has no clean single-token label, the readout may miss it. Anthropic positions the method as another audit instrument, not a complete detector for model intent.

Outside experts saw promise and limits

Anthropic also published external commentary from researchers in neuroscience, philosophy and interpretability. Stanislas Dehaene and Lionel Naccache, who helped develop the global neuronal workspace model that inspired the work, wrote that the result points to a mechanistic and testable version of the hypothesis, while stressing differences between Claude and humans, including Claude's lack of a body and enduring episodic memory.

Patrick Butlin, Derek Shiller, Dillon Plunkett and Robert Long wrote that the research is significant evidence for access consciousness in LLMs, while maintaining uncertainty about phenomenal consciousness, the subjective "what it is like" question. Neel Nanda, who leads language-model interpretability at Google DeepMind, wrote in the commentary that the paper presents compelling evidence for a cognitive space used as working memory during a forward pass and that J-lens could help with model forensics.

Anthropic released a GitHub repository for the Jacobian lens implementation. The repository describes the code as a reference implementation, released under Apache 2.0, and says the examples use Qwen open-weight models rather than Anthropic model weights. Anthropic also says it partnered with Neuronpedia on an interactive J-lens demo for open-weight models.

The practical question for Anthropic is whether J-space can become operational inside safety work rather than remain a research visualization. The July 6th paper gives Anthropic a stronger story than simple transparency rhetoric: it claims a specific structure, a specific readout method, interventions that change outputs and alignment cases where hidden state mattered. The unresolved piece is coverage. A tool that can read some of a model's silent reasoning is valuable only to the extent that the dangerous reasoning passes through the channel being watched.