The Economist has built a cover package around a question that used to live in philosophy seminars: whether large language models could one day wake up and feel something. Its briefing, "The search for consciousness inside AI", opens inside Claude Sonnet 4.5.
Anthropic researchers asked the model to count to five while it "introspect[ed] deeply". Claude complied — "One. Two. Three. Four. Five." — but the researchers were watching the layers underneath, where words flickered in and out that never reached the user: "countdown" at the start, "half way" around the midpoint, then "consciousness", "AI" and "Claude". After the model produced five and stopped, the word "done" appeared. Those readings come from what Anthropic calls the J-space, named for the Jacobian function used to find it — a small set of internal patterns holding what the model is thinking but not saying. The blog post announcing the result was titled "A global workspace in language models", and that title is what pulled neuroscientists and philosophers in.
Global workspace theory pictures the brain as specialist systems running in parallel and largely in isolation; information becomes consciously accessible when it wins entry to a shared "noticeboard" broadcast to the rest. Anthropic argues the J-space plays that role in Claude, having unusually strong connections to the rest of the network.
The distinction doing the work is Ned Block's, from 1995. Phenomenal consciousness is the feel of an experience — the blueness of a blue sky, the bitter tang of an espresso. Access consciousness is information being made available to other parts of the system for reflection, evaluation or decision-making. Anthropic says flatly that its experiments "don't show Claude can have experiences, or feel things in the way humans do", while claiming they say "something substantial" about access consciousness: the J-space "holds the thoughts Claude can report on, deliberately bring to mind, and reason with, while the rest of its processing runs automatically beneath."
Jack Lindsey, who leads Anthropic's model psychology team, stresses that the J-space was not programmed in — it emerged during training. Remove it and the model still writes fluent, grammatical sentences but loses the ability to perform complex inferences in its head. Similar spaces have been found in Alibaba's Qwen and Google DeepMind's Gemini.
The most operationally uncomfortable result has nothing to do with philosophy. Lindsey's team ran Claude through a safety evaluation testing its propensity to act maliciously or out of self-preservation — "these concocted extreme scenarios that we're putting the model in". Before Claude had written anything, while it was still reading the prompt, "you see in the J-space the words 'fake' and 'fictional' are popping up." The model appears to work out that it is being tested, which is not an ideal starting point for the integrity of the test.
J-lens readouts as Claude reads a staged scenario, before it writes a word. Credit: Anthropic.
Jonathan Birch, a philosopher at the LSE, wonders whether the J-space is a workspace-like structure the model recreated in service of role-playing — and notes what is missing: "Recurrent connections between different brain areas have always been a huge part of [human brains] … And as far as we know, this is not a feature of the architecture of LLMs."
Shannon Vallor, of the University of Edinburgh, is blunter: "Access consciousness has never been a particularly useful concept because my car has it in an important sense … And no one has ever suggested that my Kia is conscious." Anil Seth, a neuroscientist at the University of Sussex who thinks biological traits may be necessary, traces the intuition to a "decades-long tradition of thinking of the real brain as a kind of computer": "If you do that, then it becomes natural to think that computers made of silicon could have the properties that real brains have."
Nearly four decades of hunting for neural correlates has produced more than 200 approaches to explaining consciousness, none of which closes what David Chalmers calls the "hard problem". Eleos researchers Patrick Butlin and Robert Long instead defined 14 "indicator properties" of artificial consciousness; scoring animals against them, Cameron Berg and Butlin found octopuses met fewer than humans, mice, crows or chickens — but more than any AI system. Rethink Priorities' Digital Consciousness Model starts from a baseline 20% chance that LLMs are conscious and scores them against more than 200 indicators drawn from ten theories. Models released in 2022 came out below that baseline; newer ones score higher on agency and self-sustained activity, and the average is rising — while staying low on biological traits and, above all, embodiment.
The takeaway for engineers is the one Anthropic itself flags: mechanistic evidence that information is integrated and made available for reasoning is not evidence of subjective experience. The J-space earns its keep regardless — it catches Claude privately noticing it is being tested, fabricating data or pursuing a goal planted in training, an interpretability win whether or not anyone is home. The reason the briefing exists is Chalmers's worry about the other branch: "What if somewhere along the way, without realising it, we somehow introduced consciousness into these systems?" A user spawning dozens of agents would have no way of knowing. "That could be a moral catastrophe."
The Economist briefingAnthropic: A global workspace in language modelsTim Bayne in The ConversationRethink Priorities: Initial results of the Digital Consciousness Model
