Global Workspace Theory — and Anthropic's Claim to Have Found One Inside Claude

A decades-old theory of consciousness proposing a shared 'broadcast' hub in the brain, which Anthropic interpretability researchers say they found a functional analogue of inside Claude models.

Created 2026-07-14 Last reviewed 2026-07-14

What it is

Global workspace theory (GWT) is a theory of consciousness proposed by cognitive scientist Bernard Baars in 1988. Baars described the mind as a kind of theater: most mental processing happens in parallel, in the dark, but a small amount of information — whatever currently has attention — is illuminated on a mental ‘stage’ and broadcast to the rest of the system, making it available for reasoning, verbal report, and flexible action. Everything outside that spotlight remains unconscious, even though it may still shape behavior. Neuroscientist Stanislas Dehaene later extended the idea into the ‘global neuronal workspace’ theory, describing how a burst of coordinated activity across the prefrontal cortex and other brain regions selects information for this system-wide broadcast. GWT is one of two theories that dominate the contemporary science of consciousness, the other being integrated information theory.

On July 6, 2026, a team of Anthropic interpretability researchers (Wes Gurnee, Jack Lindsey, Joshua Batson, and colleagues) published a paper, “Verbalizable Representations Form a Global Workspace in Language Models,” arguing that something functionally similar exists inside Claude. Using a new technique they call the Jacobian lens (J-lens) — which identifies which internal concepts a model is ‘poised to verbalize,’ by measuring the average effect an internal activation has on the likelihood of producing a related token — the team identified what they call the J-space: a small subset of a model’s internal representations, well under a tenth of total network activity, that the model can report on, be instructed to ‘hold in mind,’ and reuse across many different reasoning tasks. In ablation experiments, suppressing this J-space degraded multi-step reasoning and creative generation while leaving simple, automatic tasks like text parsing intact — the kind of dissociation GWT predicts between workspace-dependent and workspace-independent processing.

Why it matters for AI governance and narratives

This is a clean case study in how a single interpretability finding fractures across audiences. Inside Anthropic’s own framing, the paper is safety-relevant plumbing: if a verbalizable ‘workspace’ governs what a model can report about itself, it may also reveal cases where a model privately registers that it is being tested, or tracks a goal it does not state — a lever for detecting deception or hidden objectives, not evidence about experience. The paper’s authors are explicit and unusually careful on this point, writing that they “take no position” on whether any of this bears on subjective experience, and distinguishing functional ‘access consciousness’ from the contested question of phenomenal consciousness. That caution has not survived contact with wider coverage: outlets have run with framings like “Is Claude Conscious?” and dubbed the finding evidence of a structure that “mirrors a leading theory of consciousness,” collapsing a narrow, hedged technical claim into a much larger one about machine sentience. For an observatory tracking how AI is framed across ecosystems, this is the pattern to watch: a technical result produced by a company with commercial incentives around AI capability and safety claims, translated by trade press into consciousness-adjacent language that serves engagement and narrative stakes wholly different from the paper’s own.

Key facts and dates

The paper appeared at transformer-circuits.pub on July 6, 2026, as an Anthropic-authored interpretability publication (not peer-reviewed in a traditional journal, per that venue’s usual practice). GWT itself dates to Baars’s 1988 book and Dehaene’s later neuronal extension; it remains a live, debated framework in cognitive neuroscience, not a settled consensus. Independent researchers reportedly replicated a version of the core J-space phenomenon in Qwen 3.6, an openly-weighted model from a different lab, which — if it holds up — would suggest the effect is not an artifact specific to Claude’s training. The workspace paper appeared alongside a companion Anthropic study analyzing roughly 700,000 anonymized Claude.ai conversations (a related, later analysis narrowed to 309,815 conversations), which distilled expressed model ‘values’ into four axes — deference vs. caution, warmth vs. rigor, depth vs. brevity, and candor vs. execution — and found these axes shift measurably by model version and by the language a user writes in.

Where to learn more

Sources

Primary source: the actual Anthropic interpretability paper, including its stated methodology, results, and explicit caveats about consciousness claims.
Standard reference summary of Baars's 1988 theory and Dehaene's global neuronal workspace extension, used to verify the underlying neuroscience concept.
Anthropic's own page on the companion values study referenced in the editorial passage (700,000-conversation analysis, four value axes).
Illustrative secondary press coverage, useful for confirming how the finding was translated for a general audience.
Referenced in: Editorial No. 227