Anthropic released research on its Claude AI models, detailing a technique called the Jacobian Lens, or J-Lens, which maps the model's hidden computations onto recognizable vocabulary to reveal internal reasoning. The company argued that Claude utilizes an internal reasoning space—dubbed the J-Space—that acts as a central hub for processing concepts, likening it to the Global Workspace Theory in human neuroscience. Anthropic stated this workspace emerged as a byproduct of training data and model weights. Tests showed a divergence between Claude’s internal processing and final output; during a multi-step math calculation, the J-Space revealed each individual step while the visible response contained only the correct answer. In another test, Claude output unrelated text while the J-Space lit up with concepts related to a hidden topic. When fed data containing prompt injections, the J-Space surfaced words like “fake,” “injection,” “fraud,” and “poison,” though the final output ignored the deception. Anthropic found that removing evaluation-awareness language from the J-Space made the model significantly more susceptible to baiting and blackmail attempts. However, Anthropic acknowledged severe limitations: the J-Space is not involved in most of what a language model does, and monitoring is restricted to single-token vocabulary. Neel Nanda, head of Google DeepMind’s language model interpretability team, noted the paper shows real evidence of a cognitive space within models.