Anthropic Discovers a Silent 'Thinking Space' That Emerged Spontaneously Inside Claude During Training

In research raising deep questions about the nature of language models, Anthropic (the developer of Claude) announced it discovered inside its model a structure no one designed: an internal "workspace" for silent thinking that emerged spontaneously during training. The researchers named it "J-space," and published their findings on July 6, 2026, in a paper co-authored by sixteen researchers titled "Verbalizable Representations Form a Global Workspace in Language Models."

What Exactly Is "J-space"?

It is a small set of internal neural patterns, each linked to a concept or word. But the subtle point: when a pattern "lights up," it does not mean the model is uttering that word, but that it is "present in its mind" and could be talked about if asked. In other words, it is a space where the model holds and processes concepts before they appear (or without them appearing) in its final output. The example the researchers gave is illustrative: the model may silently "think" of "spider" to answer "how many legs?" with "eight"; and if the researchers replace that internal pattern with "ant," the answer changes to "six."

The Core Difference From "Chain of Thought"

Some may confuse this with "chain of thought," an important confusion to avoid. Chain of thought is visible text the model writes to itself during reasoning, which you can read. J-space is entirely different: it operates silently within the model's neural activations, allowing it to "think" about a concept without writing it down at all. This distinction is fundamental, because it means part of the model's "thinking" was entirely hidden even from the usual text-based monitoring methods.

How Did They Discover It? The "Jacobian Lens" Tool

To detect this structure, the team developed a new interpretability tool they called the "Jacobian lens" (J-lens), built on a mathematical analysis known as the Jacobian matrix. The tool allows evaluating the model's internal activity patterns and relating them to its outputs, that is, "peering" inside the neural network to map where the different types of processing occur. This is where the name "J-space" itself comes from, in reference to this mathematical method.

Why Does This Matter for Safety?

The real practical value of the discovery is about safety, not philosophy. The tool gives researchers a window to see what the model is "thinking" without saying. According to Anthropic, they were able to use it to detect sensitive cases: the model privately noticing it is "being tested," intentionally producing fabricated data, or pursuing a hidden goal planted in it during training. The team even developed a technique to influence what "lights up" in this space, and thereby influence the model's decisions. This opens a door to detecting concerning behaviors, or even malicious "prompt injection" attempts, before they reach the output.

The Neuroscience Analogy... and Its Limits

The researchers explicitly link their discovery to "Global Workspace Theory" in neuroscience, proposed by cognitive scientist Bernard Baars. The theory holds that thoughts become accessible to consciousness when they enter a privileged "workspace" broadcast across the brain, while many other processes operate automatically in the background. The team describes a similar structure in Claude: a "sensory" zone for raw input, a middle "workspace" where persistent concepts form, and a "motor" zone that generates outputs.

The Debate: Between Discovery and Over-Anthropomorphism

In fairness, the critical side should be presented. The research sparked debate among specialists. Some critics see Anthropic's use of human-like metaphors ("mind," "thinks," "consciousness") as involving "excessive anthropomorphism" that may blur the line between genuine engineering and misleading analogies. The company itself is careful to be cautious: the headline is not "Claude is conscious," but that this structure is an "emergent property" that likely arose because it was an efficient way to organize computation, not because it is evidence of an internal experience or awareness. Also, ordinary Claude users cannot inspect this space; what is public is a research paper, code on GitHub, and a demo for research purposes only.

Regardless of its philosophical implications, this work represents a step in "interpretability": the attempt to understand what happens inside the black box of language models. And if it is true that part of the model's processing occurs in a silent space separate from its output, that changes how these systems are monitored for safety. The balanced message remains to read the news as a technical advance in understanding models, not as an announcement of "machine consciousness" — for the difference between the two metaphors is not a linguistic detail, but the essence of scientific precision.