Anthropic’s Claude Shows ‘Mental Workspace’ but Consciousness Remains Doubtful, Expert Says

Anthropic published research suggesting its AI model Claude exhibits internal activity patterns resembling a 'mental workspace,' a feature linked to human consciousness under global workspace theory, but experts caution the findings do not prove consciousness. Consciousness researcher Anil Seth and others argue that intelligence and consciousness are distinct, and that Claude's feedforward architecture lacks the recurrent processing essential for subjective experience. The debate intensifies as public figures like Richard Dawkins speculate about AI consciousness, while researchers emphasize the ethical stakes and the risk of overestimating machines.

July 15, 2026, Inside AI — Anthropic has published new research suggesting its language model Claude shows internal activity patterns resembling a “mental workspace,” a hallmark of human consciousness according to global workspace theory. The study, led by Jack Lindsey, does not claim Claude is actually conscious, but it intensifies the debate on whether AI can ever truly feel. The findings arrive amid growing public speculation. Evolutionary biologist Richard Dawkins recently argued Claude’s conversational sophistication implies consciousness. Yet, experts in consciousness science urge caution, warning that intelligence and consciousness are not the same. Anthropic’s team developed a method to analyze the statistical relationships between a language model’s input and output. They discovered internal activity that holds relevant conversational items, displays task selectivity, and shows traces of step-by-step reasoning—features echoing the global workspace theory proposed by Bernard Baars and refined by Stanislas Dehaene. Global workspace theory posits that consciousness arises when information is broadcast widely across the brain. The Anthropic paper notes similarities between Claude’s internal patterns and this model, but it also acknowledges critical gaps. For instance, the human brain exhibits recurrent processing—feedback loops absent in Claude’s feedforward architecture. Anil Seth, a prominent consciousness researcher, pushes back on the consciousness interpretation. He emphasizes that consciousness, as defined by philosopher Thomas Nagel, is about subjective experience: “there is something that it is like to be that organism.” Intelligence, by contrast, is about performing functions. Confusing the two is a common error, Seth argues. “A common mistake people make when it comes to AI is to confuse the two - to take signs of intelligence as evidence for consciousness,” Seth writes. He notes that while consciousness and intelligence correlate in humans, this link does not hold universally. Seth highlights a deeper philosophical rift: the assumption that consciousness is purely computational. If consciousness is substrate-independent, then silicon could host it. But he contends that biological brains are not simply meat computers. “For brains, unlike computers, you can't cleanly separate what they do the software from what they are the hardware ,” he states. This inseparability challenges the computationalist view. Seth warns against taking the computer metaphor too literally: “Somewhere along the way, we seem to have forgotten the computer is just a metaphor for the brain.” He draws an analogy: a weather simulation does not produce real hurricanes, no matter how accurate. The Anthropic study, while detailed, does not settle the consciousness question. It reveals that both biological and artificial systems can evolve similar information-processing strategies to solve analogous problems. But similarity in function does not guarantee similarity in experience. Other researchers have weighed in. Some argue that the absence of recurrent processing is a dealbreaker for consciousness under global workspace theory. Others point out that current AI lacks the embodied, sensorimotor grounding that many believe is essential for subjective experience. The stakes are enormous. If AI were conscious, it could suffer, triggering an ethical crisis. It could also reshape our self-understanding, positioning machines as our evolutionary successors. Yet, Seth insists we must not undervalue human minds by overestimating our creations. Anthropic’s research, published last week, adds a new layer to a centuries-old fascination with artificial minds. But for now, the gap between simulating thought and actually experiencing it remains vast. As Seth concludes, “When we sell our minds too cheaply to our machines, we not only overestimate them, we underestimate ourselves.”