Ask Claude, Anthropic’s AI model, how many legs the animal that spins webs has, and it answers eight. The word “spider” never appears in the question or the reply. Yet researchers could watch the idea of a spider light up inside the model. When they reached in and swapped that hidden “spider” for “ant,” the model answered six.
That trick is at the center of a 2026 Anthropic study. Headlines declared that Claude is conscious, but the study does not say that. It found a hidden place where the model holds a handful of ideas it can think about and reason with. A much larger amount of processing runs underneath, out of reach.
The team calls this place the J-space and compares it to how people think. A person deciding what to cook for dinner weighs the choice on purpose, while breathing, heartbeat and most other brain activity carry on automatically. Brain science has long called this idea global workspace theory. The model matches it in function, though a brain is wired differently.
Nobody built this into the model, which developed it on its own during training. By the team’s measurement, the workspace never takes up more than a tenth of the model’s activity. On a problem with several steps, the in-between answers show up there even when the model never writes them down. Asked to hold one idea while copying an unrelated sentence, the model keeps it active there, much as a person thinks about one thing while talking about another.
Whether that makes the model conscious depends on what the word means. One meaning is felt experience, what it is like to taste coffee or feel dread, and the study leaves that question open on purpose.
The workspace does give researchers a window. The model’s hidden reasoning runs through ideas it could put into words, so researchers can read some of it before the model answers. Editing a hidden note can even change the outcome, as the spider swap shows. For a company whose safety work depends on knowing what a model is doing, that is a real gain. Yet the method that opens the window is imperfect, catching only ideas that match single words, and only in a middle stretch of the model’s processing.
