Anthropic's Global Workspace Paper reveals hidden 'cognitive space' in AI models

Anthropic's much-discussed Global Workspace Paper makes four key claims: a scientific claim that models possess a 'cognitive space' storing intermediate variables, a methodological claim that Logit Lens and J-Lens can access it (with J-Lens performing better), a pragmatic claim that J-Lens is useful for interpretability (e.g., alignment audits), and a philosophical analogy to human global workspace theory. In a detailed review on LessWrong, interpretability researcher Neel Nanda (formerly of Anthropic/DeepMind) independently replicated the core findings on Qwen 3.6 27B. He finds the scientific claim 'overwhelmingly' supported, noting 'hard-to-fake evidence.' He also validates J-Lens as a cheap technique—requiring only 10 prompts of 128 tokens each—that can surface unexpected model behavior.

Despite endorsing the existence of a cognitive space, Nanda expresses caution about the fine details of the space's properties and doubts J-Lens will reliably flag all important behaviors. He sees it primarily as a hypothesis-generation tool for auditors, with false positives expected. Still, he believes J-Lens (and derivative techniques) could become a standard part of the interpretability toolkit. The paper has sparked significant community discussion, particularly around how such internal representations might generalize across models and architectures. Anthropic's work positions J-Lens as a practical step forward in model transparency, though Nanda emphasizes that 'no existing interpretability technique meets the bar of perfect reliability.'