When interacting with a chatbot like Claude or ChatGPT, it can feel like you’re talking to another conscious mind. While some prominent scientists believe this is genuinely the case, most experts remain skeptical, arguing that these systems’ impressive abilities occur without any real consciousness behind them. Anthropic recently entered this debate with a new finding: Claude appears to have a normally invisible set of internal representations that guide its reasoning and output, which the company argues echoes an influential idea known as global workspace theory.
Global workspace theory, first proposed in 1998 and later developed further, suggests consciousness arises from a central “hub” in the brain that integrates and broadcasts information for use in reasoning, behavior, and speech. Anthropic illustrated Claude’s version of this hub with an evocative video showing “ships” of conscious content sailing on a sea of unconscious processing. But whether Claude actually has a genuine global workspace is unclear, since the theory was never given a precise, formal definition.
Notably, there are real structural differences: the human brain’s workspace relies on recurrent loops, with signals cycling back through circuits over time, while Claude’s workspace emerges in a single forward pass. Human theory also invokes a process called “ignition,” where representations are non-linearly amplified before entering conscious awareness, something with no clear equivalent in Claude.
Even setting those differences aside, having a global workspace wouldn’t necessarily mean consciousness. Global workspace theory remains controversial among experts, and many believe consciousness requires more than pure computation. There’s also a deeper conceptual problem: the theory’s original architects framed it specifically as an account of “conscious access”, meaning the availability of information for recall, behavioral control, and verbal report, while explicitly leaving open whether it explains subjective experience itself. If the theory only describes conscious access rather than what it’s actually like to be something, then Claude having a workspace-like structure says little about whether there’s genuine inner experience behind its outputs.
Despite these caveats, the author argues Anthropic’s findings are still worth taking seriously, since they may nudge the artificial consciousness debate forward, even if only slightly. What’s more puzzling is Anthropic’s optimistic tone about the discovery, given that real artificial consciousness would carry enormous ethical, social, legal, and political consequences. If chatbots truly were conscious, treating them as mere tools would no longer be acceptable, and questions of AI welfare would become unavoidable.
The article closes with a pointed challenge: if Anthropic genuinely believes conscious AI might be on the horizon, and takes seriously its own suggestion that society should discuss whether such machines ought to be built at all, then perhaps research that could lead there should pause rather than continue. The author acknowledges a moratorium would be difficult to define and enforce, but warns that failing to act now risks letting the “horse bolt” before we’re prepared for what conscious AI would actually mean.