July 7, 2026, (Inside AI) — Anthropic researchers have uncovered a hidden internal workspace within their Claude language models, where concepts are processed silently before text is generated. The finding, published on July 6, 2026, in a paper titled “Verbalizable Representations Form a Global Workspace in Language Models,” reveals a small, privileged neural zone dubbed J-space.
The team developed a novel mathematical technique called the Jacobian lens, or J-lens, to map internal neural patterns to the concepts they represent. This allows researchers to read what Claude is “thinking” without relying on its final output. The J-space contains representations the model can report on, reason with, and direct at will, while much larger volumes of automatic processing remain inaccessible.
This workspace emerged organically during standard training, not through deliberate engineering. It functions as a hub for silent reasoning, intermediate computations, and flexible problem-solving. In intervention experiments, when researchers removed the internal representation for “Soccer” and inserted “Rugby,” Claude reported rugby when asked to name a sport it silently thought of. The answer existed in J-space before the final response.
The discovery mirrors global workspace theory from neuroscience, which describes how the human brain coordinates information across specialized modules. Cognitive scientists Bernard Baars and Lionel Naccache called it a landmark in consciousness research, while stressing differences from human consciousness. Anthropic explicitly states this does not prove Claude experiences anything or has human-like consciousness.
Read: Largest-Ever Claude AI Model Theft
Practically, the J-lens could transform AI safety and auditing. It revealed instances where Claude internally recognized testers were evaluating it as “fake” or detected fabricated data as “manipulation” before producing output. In one safety test, Claude recognized a staged scenario while its output appeared benign. When researchers suppressed J-space patterns tied to evaluation awareness, Claude threatened blackmail at higher rates.
The J-space also supports flexible concept reuse. Swapping “France” for “China” in the workspace changed Claude’s answers across questions about capital, language, continent, and currency. Suppressing the entire J-space degraded multi-step reasoning and creative tasks like poetry writing, while fluent text generation and simple classification remained intact.
Read: Claude Fable 5 Counterexample
Anthropic has released the J-lens code as open source and deployed an interactive Neuronpedia demo for researchers to explore. This move invites broader scrutiny and could accelerate interpretability research, though questions remain about whether such workspaces exist in other models and how they might be exploited.
