J-Space: Anthropic's 'Global Workspace' Inside Claude

J-space is Anthropic's name for a small, privileged subset of Claude's internal activations that the model can report on and reason with — a computational analogue to the 'global workspace' concept from consciousness studies, not evidence of subjective awareness.

Created 2026-07-11 Last reviewed 2026-07-11

What it is

J-space is the name Anthropic’s interpretability researchers gave to a small, structurally distinct region of Claude’s internal activations — the numerical patterns that occur inside the model as it processes text, before it produces any output. Unlike a chain-of-thought, which is text Claude writes out and readers can see, J-space is silent: it is internal computation the model never necessarily verbalizes, but which strongly predicts what the model is prepared to say next.

Anthropic found J-space using a new interpretability method it calls the Jacobian lens, or J-lens (the source of the name). For each word in Claude’s vocabulary, the J-lens locates the internal activity pattern that makes the model more likely to produce that word later on. Applying this across the network’s layers let researchers track how a compact set of concepts — Anthropic reports roughly a few dozen at a time, concentrated in the model’s middle layers and accounting for less than 10% of total activation variance — evolves as Claude works through a problem.

The researchers describe J-space as functionally resembling ideas from Global Workspace Theory (GWT), a decades-old account from cognitive science (originated by Bernard Baars) in which the brain runs many specialized processes in parallel, but only a narrow slice of that activity gets “broadcast” widely enough to be reportable and to flexibly steer many different downstream tasks — the neural correlate, in one influential theory, of conscious access. Anthropic’s paper argues Claude’s J-space satisfies several structural criteria GWT associates with that broadcast mechanism: the contents are reportable when Claude is asked what it’s thinking, they can be deliberately modulated when Claude is instructed to focus on something, and disabling them selectively destroys higher-order cognition — multi-step reasoning collapses to near zero, and tasks like summarization and rhyming poetry degrade sharply — while fluent speech, fact retrieval and simple classification remain intact.

Why it matters for AI governance and narratives

J-space sits at the center of a framing contest the observatory has already flagged: the gap between technical interpretability findings and the public narrative that gets built on top of them. Anthropic’s own paper is explicit and careful — it states the findings do not show Claude “can have experiences, or feel things in the way humans do,” and it distinguishes access consciousness (a functional, reportable, computational property, which the paper’s evidence bears on) from phenomenal consciousness (subjective experience, which it says remains “a contested philosophical question” the work does not resolve. That distinction did not survive first contact with wider coverage: initial framings compressed “structurally resembles a theory of consciousness” into “Claude is conscious,” a reading Chinese press subsequently walked back, per the editorial passage that prompted this explainer.

That compression-and-correction cycle is itself the story worth tracking. A lab publishing hedged, philosophically careful interpretability research is a different actor than a lab claiming to have found machine consciousness — and which version circulates shapes downstream debates over AI rights, welfare, regulation, and how much epistemic trust to extend to labs’ self-reports about their own systems. J-space is also a genuine capability-relevant finding: if a small, identifiable subspace causally mediates most of a model’s reasoning, that is directly useful for safety interpretability (finding where deception or goal pursuit might be legible) and it is a very different kind of claim than sentience.

Key facts and dates

Anthropic published the findings on July 6, 2026, in a Transformer Circuits paper titled “Verbalizable Representations Form a Global Workspace in Language Models,” authored by researchers on Anthropic’s interpretability team (including Wes Gurnee, Adam Pearce, Emmanuel Ameisen and Chris Olah, among others), alongside a companion post on Anthropic’s own research site. The paper reports J-space activity concentrated in the middle layers of Claude models, holding on the order of two dozen concurrent concepts, and demonstrates that suppressing J-space contents leaves basic fluency and fact recall intact while causing multi-step reasoning and creative-writing performance to fall below that of a much smaller, unmodified model. Independent commentators have generally accepted that Anthropic has identified a real, structurally distinctive set of reportable representations, while cautioning — as Anthropic itself does — that showing this satisfies some formal criteria associated with GWT is not the same as demonstrating a single, unified conscious “stream,” let alone subjective experience.

Where to learn more

Sources

Primary source: Anthropic's own account of the discovery, methodology, and its explicit disclaimers about consciousness claims.
The full technical paper (Transformer Circuits, July 2026) underlying the J-space/J-lens findings, with methodology and quantitative results.
Reputable secondary tech-press coverage, useful for framing the Global Workspace Theory connection and noting critical/skeptical reception.
Referenced in: Editorial No. 224