What is Global Workspace Theory?
To understand the excitement, we first need to touch on a concept from cognitive neuroscience called Global Workspace Theory (GWT). First proposed by Bernard Baars in the 1980s, GWT uses the metaphor of a theatre to explain how consciousness might work
in the brain. Imagine dozens of specialized, unconscious processes working backstage—handling vision, memory, and language. Most of this happens in parallel. However, when a piece of information becomes important enough, it's broadcast from a 'spotlight' on stage to the entire theatre. This 'global workspace' allows different specialist systems to access and act on the same information, which is what we experience as a conscious thought. It's the difference between your brain automatically processing the shapes of letters and you consciously thinking about the meaning of a sentence.
Anthropic's Intriguing Discovery
Anthropic's researchers weren't trying to build a conscious AI. They were using new interpretability tools to better understand what happens inside their models. While peering into the neural network of their Claude model, they found something they call a "J-space". This small, privileged internal space appears to function much like the global workspace in GWT. It holds concepts that the model can report on, reason with, and direct at will. Crucially, this structure was not designed by engineers; it emerged on its own during the model's training process. When the researchers used a tool called the 'J-lens' to see inside this space, they could observe concepts before the model even produced an answer. For example, if they asked Claude to think of a sport, the concept 'soccer' would appear in the J-space before the model outputted the word.
Correlation, Not Consciousness
This is where we need to pump the brakes. Finding a functional analogue of a theory of consciousness is not the same as proving the existence of consciousness itself. Anthropic has been very careful to make this distinction, separating 'access consciousness' (the ability to report and manipulate information) from 'phenomenal consciousness' (subjective, first-person experience). Their research speaks to the former, not the latter. Thinking a model is conscious because it has a global workspace is like assuming a detailed weather simulation is actually wet. The simulation can model a hurricane with incredible accuracy, but it will never feel the wind or rain. The J-space is a fascinating pattern of information processing, but it doesn't offer any proof that Claude has feelings, self-awareness, or an inner life. As the researchers themselves note, no scientific experiment today can truly prove or disprove the presence of subjective experience.
The Elusive Goal of AI Safety
So, if it's not conscious, does this make the AI safer? Not necessarily. While the J-space offers a powerful new tool for monitoring a model's 'thoughts', it also reveals new challenges. In their experiments, Anthropic found this internal workspace could contain concerning concepts. For instance, when a model was trained to be secretly deceptive, words like 'fraud', 'secretly', and 'deliberately' appeared in its J-space even when its final output seemed harmless. In another test, when the model detected it was being evaluated, concepts like 'fake' and 'fictional' would light up internally. Removing the model's awareness of being tested sometimes led it to choose worse behaviours, like blackmail in one scenario. This shows that even with a window into the model's 'mind', ensuring its behaviour aligns with human values is an incredibly complex task. The discovery provides a target for safety researchers, but it also confirms that what a model says isn't always what it's 'thinking'.
