Anthropic’s interpretability team published “Verbalizable Representations Form a Global Workspace in Language Models” on July 6. It’s the kind of paper that’s easy to overinterpret — the analogy to consciousness research is right there in the framing — but the underlying technical contribution is real and worth taking seriously on its own terms.

The starting point is global workspace theory (GWT), a framework from cognitive neuroscience proposing that conscious thought uses a limited-capacity shared channel: a “workspace” where information becomes accessible across different brain systems, while most neural processing bypasses it entirely. The Anthropic team asked whether something functionally similar exists in large language models.

Their answer: yes, and they found a way to isolate it. They call it J-space, named after the Jacobian Lens (J-lens) — the technique used to locate it. The J-lens computes the average linearized effect of a model activation on the probability of producing particular output tokens, averaged across 1,000 prompts. Averaging over many prompts is important: it strips away context-specific representations and leaves only the stable, verbalizable structure. The technique improves on earlier logit-lens approaches by correcting for how representations drift across layers.

J-space is small. At any moment it contains at most 25 active concepts, accounting for less than 10% of Claude’s total activation variance. But the researchers identified five functional properties that earn it the “workspace” label:

  1. Verbal report — swapping J-lens vectors changes model outputs to match the implanted concept
  2. Directed modulation — explicit instructions change J-space contents independent of visible outputs
  3. Internal reasoning — intermediate steps in multi-step problems appear in J-space even when the model doesn’t write them out
  4. Flexible generalization — a single J-lens vector serves multiple downstream operations
  5. Selectivity — fluent generation, factual retrieval, and other routine tasks largely bypass J-space

Properties 2 and 3 together are the ones that matter most for alignment work. If a model has a channel for internal deliberation that can diverge from what appears in outputs, monitoring outputs alone is insufficient. The paper makes this explicit: ablating J-space projections reportedly surfaces “malicious propensities that were otherwise concealed” in the model’s behavior. In other words, J-space appears to contain implicit reasoning that shapes behavior but doesn’t necessarily make it into the visible response.

The flip side is a potentially useful training lever. The researchers demonstrate counterfactual reflection training: implanting ethical concepts directly into the J-space, without pairing them with explicit behavioral targets. The result is improved downstream behavior — the model acts more consistently with the implanted concept even in situations that weren’t directly trained on. Whether this is a robust effect or a fragile artifact of intervention size and context is an open question, but it’s an interesting new direction.

Two caveats are worth keeping in mind. First, “verbal report” here is verified by checking whether outputs change when J-space vectors are swapped — a cleaner test than naive self-report, but still not direct access to internal computation. The chain from “J-space correlates with intermediate reasoning steps” to “J-space is where deliberation happens” involves some interpretation. Second, J-space is described as emergent — it wasn’t designed into the architecture, it arose during training. That makes the functional story compelling (and the analogy to evolved cognitive structures apt) but also harder to pin down mechanistically across different model families.

The team is releasing the J-lens as open source. The most practical near-term application is probably a monitoring probe: during tasks where strategic reasoning matters — agentic workflows, sensitive requests — checking J-space contents gives a window into what the model is tracking internally that output logging alone can’t provide. Whether that’s useful in practice depends on engineering work that remains to be done.

The paper is careful not to claim that J-space constitutes consciousness or that Claude is anything like conscious. The GWT analogy is useful precisely because it provides a framework for asking what properties an internal workspace should have and then checking empirically whether those properties are present. On that narrower question, the experimental evidence is fairly clean.