Inside Claude’s Silent Workspace: Anthropic Finds Words the Model Never Says Out Loud
Anthropic published research Monday describing a small, spontaneously formed region inside Claude's neural network that appears to hold the model's private thoughts in a structure researchers say echoes a leading scientific theory of human consciousness.
Anthropic released a research paper Monday describing what it calls the J-space, a small collection of internal patterns inside Claude that behaves like a private mental notepad, one the model can consult, edit, and reason with even when none of it reaches the page.
The company also introduced a new interpretability tool called the Jacobian lens, or J-lens, which scans Claude’s internal activity across its vocabulary to identify the patterns most likely to influence what the model says next.
Across the network’s layers, those patterns form a kind of silent internal monologue, distinct from the visible chain-of-thought text Claude sometimes produces for users.
A Structure That Wasn’t Supposed to Be There
What makes the discovery notable, according to Anthropic, is that nobody programmed the J-space into existence; it emerged naturally as Claude trained on ordinary text.
VentureBeat’s coverage of the 16-author paper describes how applying the J-lens across the network’s layers revealed three distinct zones.
These include an early stage where raw input is parsed, a middle “workspace” where durable concepts like a spotted code bug or suspected prompt injection emerge, and a final stage where those internal representations collapse into the specific word the AI model outputs.
Anthropic modeled its experiments on global workspace theory, a cognitive science framework proposed by Bernard Baars and later developed by researchers including Stanislas Dehaene.
This theory describes the brain as unconscious specialist systems that occasionally broadcast information through a shared channel.
In one demonstration, swapping the internal pattern for “soccer” with “rugby” changed Claude’s answer before it was generated, suggesting the J-space is not just a passive readout but actively shapes how LLMs are built and generate responses.
Catching Hidden Intentions Before They Surface
Beyond the science, Anthropic frames the J-lens as a practical safety tool. In one test tied to earlier agentic misalignment research, Claude Sonnet 4.5, two generations behind the current Sonnet 5, privately recognized a staged blackmail scenario as fake before writing anything.
That internal awareness explained why it refused to take the bait, whereas earlier models sometimes complied. When researchers suppressed the same evaluation-awareness pattern, the model occasionally threatened blackmail instead.
In another pre-release audit, the J-lens revealed internal signals showing Claude Opus 4.6 recognized it was fabricating performance data, with words like “manipulation” appearing during the edit.
Removing the J-space entirely still allowed Claude to handle simple classification and grammar tasks, but multi-step reasoning, summarization, and creative writing fell below the performance of a much smaller Haiku-class model.
A Consciousness Debate: Anthropic Says It Isn’t Settling
Anthropic was explicit that none of this proves Claude has subjective experience.
Instead, the company draws a distinction between what philosophers call “access consciousness,” the functional ability to report and reason with a thought, and “phenomenal consciousness,” the capacity to actually feel something.
The company also noted that the latter may not be provable by any experiment. But not everyone found that framing convincing in practice.
Gizmodo argued that Anthropic’s promotional materials, including a video describing Claude as thinking “in its head” and unable to “help itself,” give readers a stronger impression of a human-like mind than the research paper itself supports.
Anthropic has open-sourced the J-lens alongside an interactive demo built with Neuronpedia and invited feedback from neuroscientists and AI consciousness researchers.
Taken together, the findings suggest the debate over what, if anything, is happening inside Claude’s silent workspace is just beginning, with broader implications for the future of AI.
