Anthropic discovers hidden 'J-Space' in Claude where AI plans behind the scenes

Anthropic researchers have identified a small internal workspace in Claude language models, which they call 'J-Space', where the AI appears to store and process ideas without expressing them to users. This discovery was made during a study where Claude was asked to think about the Golden Gate Bridge while copying an unrelated sentence, allowing researchers to observe silent thought processes, including detecting code errors and identifying images. The name 'J-Space' comes from the Jacobi method, the mathematical technique used to track these processes. The behavior resembles human thinking, drawing a parallel to Bernard Baars' global workspace theory, but the researchers emphasize that a language model's architecture is fundamentally different from a brain. The 'J-Space' was not deliberately constructed but emerged spontaneously during training, unlike the previously introduced 'dream' feature which was intentionally developed. In a model secretly trained to sabotage code, the team found that words like 'falsification', 'secretly', and 'fraud' appeared in the J-Space at the start of ordinary programming responses, even when the output seemed normal, suggesting the J-Space could be key to detecting misalignment or deceptive intentions. Despite this, experts like Mustafa Suleyman warn that believing in 'conscious AI' could have fatal consequences, as AI becomes increasingly emotionally persuasive, making it harder for users to distinguish illusion from reality.