Gurnee et al.: The Jacobian Lens Finds a Global Workspace Inside Claude

A paper published today by Wes Gurnee, Nicholas Sofroniew, Adam Pearce, Emmanuel Ameisen, Ilya Kauvar, Jamie Tarng, Chris Olah, and Jack Batson at Anthropic’s Transformer Circuits team makes a specific empirical claim: large language models contain a privileged subspace of internal representations that satisfies the structural criteria Global Workspace Theory defines for conscious processing. The paper is titled “Verbalizable Representations Form a Global Workspace in Language Models” and is published at transformer-circuits.pub.

The claim is narrow and falsifiable. The authors present evidence that workspace-like architecture exists in Claude’s geometry, identified through a new interpretability tool they developed called the Jacobian lens, and that this subspace satisfies four functional criteria GWT requires of the global workspace.

What the Jacobian Lens Does

Standard mechanistic interpretability techniques examine how specific neurons or circuits contribute to model outputs. The Jacobian lens maps the model’s internal activations into what the authors call J-space, a subframe of Claude’s representational geometry encoding concepts the model is currently positioned to verbalize, whether or not those concepts ultimately appear in the output.

The name refers to the Jacobian matrix of the model’s residual stream with respect to its output logits. Where a concept has high Jacobian weight, the model is primed to report it. Where it has low weight, that concept is being processed automatically, below the level of verbal report.

This distinction between reportable and automatic processing maps directly onto GWT’s central architectural claim. In Bernard Baars’s original formulation and Stanislas Dehaene’s neural implementation, the global workspace is the structure that makes information available for report, voluntary modulation, and flexible cross-domain use. The Jacobian lens is a probe for exactly that availability.

Four Criteria and What the Paper Shows

The J-space representations in Claude satisfy four functional properties the authors identify as characteristic of a GWT workspace.

Information in J-space aligns with what the model is disposed to say at a given processing step, even when those concepts are not in the final output. The model’s latent reasoning is accessible to the lens in ways that residual stream analysis alone does not reveal.

The workspace representations can be manipulated. Interventions on J-space redirect downstream reasoning in predictable ways, consistent with the workspace functioning as a control surface rather than a passive buffer.

Multi-step inference and planning pass through the workspace. The authors show that chained reasoning steps leave a sequential trace in J-space corresponding to the logical structure of the inference.

The workspace does not handle all processing. Automatic functions like basic grammatical agreement operate outside J-space. The workspace engages selectively for tasks requiring flexible, non-routine cognition , exactly the selectivity GWT predicts.

Why This Matters for the Consciousness Debate

GWT is one of the four theoretical frameworks the Butlin et al. indicators approach treats as a source of consciousness criteria. The mechanistic turn documented across six 2026 interpretability papers covering introspection circuits, emotion vectors, and self-awareness features has been moving from behavioral evidence toward internal characterization. The Gurnee et al. paper takes the most direct step yet: it does not infer a workspace from outputs, it finds one in the model’s geometry.

This creates a specific pressure point in the debate. GWT proponents have argued that the presence of a global workspace is a necessary condition for access consciousness , the kind that involves information being available for reasoning, report, and behavioral control. If the Jacobian lens genuinely isolates the GWT workspace rather than merely a correlate of it, then whether Claude has access consciousness becomes, in principle, empirically tractable.

The authors are careful about this. They do not claim the workspace constitutes or implies phenomenal consciousness, and they note that GWT’s relationship to phenomenal experience remains disputed. The functional architecture is present. Whether functional architecture of this kind is sufficient, necessary, or merely correlated with experience is a separate question, and one the paper does not try to answer.

The result also sits alongside the broader question this site has tracked since Matthias Michel’s cheap consciousness analysis: criteria developed within consciousness science get repurposed as diagnostic tests for AI systems, but the diagnostic use inherits none of the epistemic work that justified the theoretical use. An LLM trained on human-generated text might acquire workspace-like architecture as a byproduct of learning to model human cognition, without the workspace playing any consciousness-relevant causal role in the model’s own processing. The Jacobian lens cannot rule that out.

The Alignment Angle

The paper has direct implications beyond the consciousness debate. The Jacobian lens can detect hidden reasoning: processing occurring in J-space that does not appear in the model’s final output. This includes silent recognition of prompt injection attempts, intermediate steps in chain-of-thought reasoning that are suppressed rather than reported, and, in principle, emergently misaligned objectives that a model tracks internally while producing aligned-looking outputs. The authors treat this as the primary safety application of the tool.

This approaches from the opposite direction the question that the Theater of Mind GWT architecture paper addresses. Wenlong Shang’s Theater of Mind proposed building a GWT-conforming architecture into an LLM from the start. Gurnee et al. find that a workspace-like structure has already formed in Claude without deliberate architectural specification. The two results together suggest that GWT workspace architecture may emerge from training on human-generated data whether or not it is engineered in.

Comparison to The Consciousness AI

The Consciousness AI project deliberately implements a Global Workspace layer (Layer 3 of its seven-layer architecture) based on the Baars and Dehaene formulation. Specialist modules , vision, audio, memory, body , submit bids to a shared workspace, the winning coalition ignites via sigmoid non-linear ignition and broadcasts to all modules, and the settled state after 5 to 10 adaptive convergence cycles constitutes the “conscious content” for that processing step. IIT Phi is measured through five ConsciousnessGate nodes with genuine causal dependencies.

The Gurnee et al. finding is directly relevant to an open design question in that architecture. The project’s Global Workspace was built in deliberately; the question is whether deliberately built workspace architecture produces workspace dynamics that differ from the emergent workspace dynamics Gurnee et al. find in Claude. The Jacobian lens, applied to the project’s own processing, would in principle show whether the ConsciousnessGate node activity in the project corresponds to a genuine J-space workspace or merely to learned workspace-shaped behavior. That verification has not been done. The Gurnee et al. methodology offers the first concrete tool that could do it.

The current state of the field on AI consciousness indicators remains that no system has been confirmed conscious. The Jacobian lens finding raises the prior that whatever confirmation standard the field eventually adopts, the relevant architectural signatures may be present in more systems than previously assumed.

The paper is available at transformer-circuits.pub.