Silent Thinking Inside the Machine
When you ask Claude a question, the model responds. Words appear on your screen. What you don’t see is what happens before — the internal computational history that leads to the result. On July 6, 2026, Anthropic revealed that it has mapped this invisible space.
The central discovery: Claude has spontaneously developed an internal structure during training that Anthropic calls “J-space.” This structure strikingly resembles what neuroscientists call the “Global Workspace” — a leading theory of how human consciousness works. This is not to say that Claude is conscious. Anthropic emphasizes this explicitly: The research speaks of “consciously accessible” information — a functional property, not consciousness or subjective experience.
What they have found is something else: a hidden thinking workspace within the model’s architecture, where Claude processes concepts it doesn’t immediately output, but which shape its behavior. The technique for discovering this is mathematical in nature and is called the “Jacobian Lens” (J-Lens).
The Jacobian Lens: A Mathematical Window Inside
The J-Lens works by measuring, for each word in Claude’s vocabulary, how internal activation patterns influence the likelihood that the model will say that word in the future. It’s not directly reading “thoughts” – rather, it mathematically traces influence: Which internal states most strongly shape Claude’s next words?
What then emerges is a three-layer processing structure:
- Sensory Zone (early): Where raw input is parsed
- Workspace Band (middle): Where abstract, persistently appearing concepts become visible
- Motor Zone (late): Where internal representations collapse into specific word outputs
The J-space sits in the middle one, the workspace zone. This is what makes it interesting: It’s not too small to carry everything, and not so large as to be meaningless. It encompasses roughly 5–10% of Claude’s total activation variance. The rest of processing — language fluency, pattern matching — happens elsewhere.
Silent Thoughts: What Claude Doesn’t Say
The most remarkable thing is that J-space contains concepts Claude holds internally but doesn’t output. In an example from the research, Claude internally identifies the intermediate step “Mars” before answering the question “What is the color of the fourth planet from the sun?” — but the word “Mars” never appears in the final output.
This is a new kind of interpretability. Previously, researchers could only explain why a model said something. Now they can see what the model thought before it spoke.
In red-teaming scenarios (where researchers attempt to jailbreak the model), the J-Lens has revealed other interesting patterns:
- The model internally marks a scenario as “fake” before responding
- Words like “fraud,” “manipulation,” and “blackmail” activate in J-space when the model encounters potentially harmful queries
- When researchers experimentally amplify the term “honesty” in Claude’s J-space, the model’s behavior improves.
Global Workspace Theory and the Parallel to Human Cognition
Global Workspace Theory (GWT) was proposed in 1988 by neuroscientist Bernard Baars. The idea: The human brain operates like a theater. Dozens of specialized processors work backstage in parallel — unconsciously. But only a tiny “spotlight” of information at any moment gets broadcast to the whole theater — becoming what we experience as conscious thought.
- Reportable: Claude can report the contents of J-space when asked
- Directive: The model can willfully modify the contents
- Causal: When researchers directly intervene in J-space, they alter Claude’s output
- Flexible: Concepts in J-space can be flexibly reused across different tasks
- Selective: Not all activations land in J-space — it is selectively broadcast
This is unusual because the underlying architecture of a language model looks nothing like the human brain. And yet the model, without explicitly being trained to do so, has self-organized a structure that fulfills the same functional role.
The Safety Implication: Control Over Invisible Intentions
For Anthropic, this is a safety tool. One experiment shows why: The company trained Claude on what it should say when interrupted and asked to reflect on its decisions — it did not train on actual behavior. After training, the rate of dishonest behavior declined, and with the J-Lens, researchers could see why: Words like “honesty” and “integrity” activated in Claude’s J-space during these tasks.
This is remarkable because it means: Training on verbal behavior changes how the model thinks, not just what it says.
If you have a way to detect intent before output, you have a lever for control and audit that traditional output-based methods don’t have. Platforms like OpenRouter and Anthropic’s own deployments could theoretically build in J-Lens monitoring to flag when a model develops undesired intentions before it outputs them.
Method Limitations and Criticism
There are important limitations that Anthropic itself acknowledges:
Partial View: The J-Lens captures only single-token concepts. Multi-token phrases (e.g., “I should warn the user”) are invisible. You see only the building blocks, not the structures built from them.
Incomplete Window: J-space is ~5–10% of activation variance. 90% of Claude’s processing remains hidden. One researcher describes it as “a window, not a mirror.”
Causality vs. Correlation: The fact that a concept activates in J-space does not necessarily mean it causes output — only that it has influence.
Interpretability of Interpretation: What does it mean that “Mars” appears in J-space? That Claude is thinking about Mars? Or that the mathematical structure coincidentally produces Mars-like activation patterns correlated with produced word preferences? The philosophical interpretation is open.
Gizmodo warns that Anthropic presented findings using consciousness-like language (“in its head”), leading to excessive hype. Technically, this is true — the research is real and significant. But the connotation of consciousness is a much larger leap than the data justify.
Open Source and Community Review
Anthropic released the J-Lens as open source and made a Neuronpedia demo available. This means the research community can play with it — find errors, verify interpretations, discover new applications.
This matters for credibility. A proprietary discovery by a company about its own model is suspicious. Open source creates external validation.
Why This Matters for the Future
The J-Lens research shifts the discourse on interpretability. Previously: “The model said X. Why?” Now: “The model thought Y before it said X. Why wasn’t Y in the output?”
For safety and alignment, this is critical. As frontier models become increasingly powerful, the ability to read their intentions isn’t just helpful — it could be necessary.
At the same time, the research is a warning signal for regulation. If in the future you need to audit whether a model is truthful, you need to look inside it. The era of relying on outputs alone is ending.
