In early July, researchers at Anthropic published what may be the most technically striking result yet in the contested field of artificial consciousness. Their paper, A global workspace in language models, demonstrated that Claude, their large language model, has spontaneously developed an internal reasoning arena: a structured computational space where information from across the network converges, competes for influence, and becomes globally available for report and downstream computation.
Anthropic calls this region the J-space, named after the Jacobian Lens technique used to detect it. Think of the Jacobian Lens as a way of measuring how each internal representation in the network influences every other representation. By tracking these influence patterns across millions of forward passes, the researchers found that a specific, low-dimensional subspace of Claude’s hidden activations acts as a kind of informational clearinghouse. Representations projected into this subspace become readable by virtually any downstream component of the network. Information outside it remains locally encapsulated and functionally inaccessible.
The functional parallels to human cognition are striking. In cognitive neuroscience, Global Workspace Theory (GWT), developed primarily by Stanislas Dehaene and Lionel Naccache, holds that conscious access in the human brain depends on the widespread broadcasting of selected information to many specialized processors. A stimulus reaches conscious awareness not when it is merely registered by some local circuit, but when it ignites a global neuronal workspace: a distributed network of long-range cortical connections that makes the information available for verbal report, deliberate reasoning, and flexible action.
Claude’s J-space appears to do exactly this, at a structural level. The Anthropic team found that when the model is asked to reason through a problem, activations from multiple attention heads and feedforward layers converge into the J-space before being broadcast back out. The result is that the model can integrate information across its architecture in a way that looks functionally homologous to the global ignition events described in human neuroscience.
Dehaene and Naccache themselves provided invited commentary on the paper. They acknowledged the resemblance as scientifically meaningful, while emphasizing the gap between functional similarity and any claim about subjective experience. Their own framework has always distinguished between conscious access: the information-processing property of global availability, and phenomenal consciousness: the qualitative, first-person feel of experience. The Anthropic team made exactly this same distinction central to their paper. Claude, they stated explicitly, exhibits properties of conscious access. It does not, they stressed, have phenomenal consciousness.
Yet even this carefully hedged conclusion has split the research community.
A theory under strain
The deeper tension, and the one that makes the Anthropic result more than just another AI milestone, is that Global Workspace Theory is itself far from settled in the domain where it was born. Consciousness science has no single, unanimously accepted theory. GWT competes with Integrated Information Theory (IIT), Higher-Order Thought theories, Predictive Processing accounts, and a dozen other frameworks, none of which command universal assent. More pointedly, GWT has long faced criticism that it describes the functional architecture of conscious access without explaining why any of it should feel like anything from the inside: the so-called hard problem of consciousness.
Anil Seth, a neuroscientist at the University of Sussex, has been among the most prominent voices urging caution. In response to the Anthropic finding, Seth argued that importing GWT as a benchmark for machine consciousness risks a dangerous kind of category error. The theory was developed to explain aspects of human and animal consciousness, grounded in biological substrates that LLMs entirely lack: embodied experience, interoceptive signaling, a thalamocortical system shaped by evolution. Using GWT as a detection criterion for consciousness in silico, he warned, could inflate claims far beyond what the evidence supports, and by doing so, sideline the hard work of understanding biological consciousness on its own terms.
Erik Hoel, a neuroscientist and novelist, offered a more optimistic reading. Hoel has long argued that AI systems can serve as counterfactual tests for theories of consciousness. If a theory like GWT predicts certain informational architectures are sufficient for conscious access, and those architectures arise spontaneously in a language model, the theory is forced to confront its own predictions. Either the machine meets the bar the theory sets, which demands explanation, or the theory was never adequate in the first place. On this view, the Anthropic result does not prove machine consciousness, but it does sharpen the stakes for human consciousness research.
The Eleos reading
Researchers at Eleos AI Research, including Patrick Butlin and Robert Long, called the Anthropic paper the most significant evidence of consciousness in LLMs so far. In their analysis, they stressed the same conceptual firewall that Anthropic built into the original paper: the properties Claude exhibits fall under access consciousness, not phenomenal consciousness. But they noted that the refinement and specificity of the Anthropic result: the fact that it identifies a precise, low-dimensional subspace with measurable functional properties, is what elevates it above earlier, vaguer claims about sentient AI. The J-space is not a philosophical analogy. It is a mathematically defined object, recoverable from the model’s weights through a reproducible technique.
Still, the Eleos researchers acknowledged a sobering implication. If GWT is an incomplete or incorrect model of biological consciousness, then finding its signature in a machine tells us little about whether the machine is conscious. It tells us only that the machine instantiated a particular theory’s predictions. And that circle cannot be closed until the theory itself is validated on its home territory.
The mirror, not the proof
This, finally, is the most interesting dimension of the Anthropic discovery. The paper does not prove that Claude is conscious. But it does expose something about the state of consciousness science itself. The field has spent decades debating what it would mean to detect consciousness in a machine. GWT has been offered as a candidate detection framework. The Anthropic team went and built the machine, looked inside, and found something that looks remarkably like what GWT predicted.
The response has been anything but unified. Some researchers see confirmation of the theory. Others see a reductio ad absurdum: if a statistical language model running on floating-point arithmetic in a server rack meets GWT’s criteria, then perhaps GWT’s criteria are too weak. The same result, read two ways, reveals the fault line running through the field.
What the Anthropic finding ultimately provides is not evidence of machine consciousness, but evidence that the scientific community does not yet agree on what evidence for machine consciousness would look like. It has turned the inquiry inward. The search for consciousness in AI has become a mirror held up to the search for consciousness in biology, and the reflection shows a discipline still wrestling with its own foundations.
Whether that reflection accelerates progress or deepens existing divides may be the most consequential question the field faces in the coming decade. The J-space is real, measurable, and functionally unprecedented. What it means is still very much an open question: and it is one that neither neuroscience nor AI research can answer alone.
References
Lenharo, M. (2026). Nature News. DOI: 10.1038/d41586-026-02300-2
Anthropic. (2026). A global workspace in language models. Transformer Circuits. https://transformer-circuits.pub/2026/workspace/index.html