Beijing Time, July 7, 2026, Anthropic Research Finds Claude Has Developed a “Human‑Like Conscious Workspace” Internally.
Anthropic today released research showing that its large language model, Claude, has spontaneously formed a neural activity region called “J‑space” (Jacobian space) during training, which functionally resembles the “global workspace” in the human brain that can be consciously accessed.
The research team explicitly stresses that this does not prove Claude possesses subjective feelings, but the region does indeed carry the model’s “unspoken thoughts” and silent reasoning steps. However, this discovery opens a new window into the internal cognitive mechanisms of large models, and the head of Google DeepMind’s interpretability team publicly praised the paper as “extremely valuable.”
A Human-Brain-Like “Broadcast Hub”
The discovery of J‑space was inspired by the “Global Workspace Theory” in neuroscience, the idea that only information entering the brain’s “broadcast space” can be reported and flexibly used. Anthropic located this region within Claude’s billions of parameters using a self‑developed interpretability tool called the “Jacobian Lens.”
J‑space is extremely small, holding only a few dozen concepts at a time and accounting for less than one‑tenth of the model’s total activity, yet it carries the core information that the model “thinks but has not yet spoken.” Unlike “chain‑of‑thought,” it operates entirely silently at the neural activation level, holding and transforming concepts without outputting any text.
Five Key Properties
The research shows that J‑space has five distinctive properties. Specifically, it has reportability. Claude can accurately tell what is currently activated in this region; controllability – when asked to think quietly or solve a problem silently, the region activates accordingly.
Internal reasoning ability – intermediate steps of multi‑step reasoning appear sequentially here; flexible generalisation – the same concept can be reused for different tasks; and selective participation – it does not engage in basic conversation, but is indispensable for higher‑order reasoning.
(Schematic diagram of ablation experiment)
Experimental Verification
The team confirmed the functional reality of J‑space through multiple intervention experiments. In a silent‑thinking experiment, Claude was asked to think of a sport without saying it; before it spoke, reading J‑space showed “Soccer” already ranked first – and when that was replaced with “Rugby,” the answer changed accordingly.
In a reasoning test, when asked “How many legs does an animal that weaves webs have?”, J‑space activated “spider” mid‑process; replacing it with “ant” changed the answer from 8 to 6.
In a cross‑task reuse test, replacing “France” with “China” in J‑space caused the model to give China‑related answers to four different questions.
These experiments prove that the final output is indeed read from J‑space, not merely recorded passively.
Does Deleting It Make Claude “Stupid”?
When Claude was prevented from using J‑space – i.e., after an ablation experiment – the model could still hold normal conversations and perform sentiment classification, but its multi‑step reasoning accuracy dropped to nearly zero, and its ability to write poetry and summarise also declined sharply.
The ablation experiment proves that this region is indispensable for higher‑order cognitive functions.
Is This Consciousness? Anthropic’s Cautious Response
Anthropic makes it clear that J‑space exhibits “access consciousness” – being reportable, reason‑able, and able to guide action – rather than “phenomenal consciousness” (subjective feelings).
The paper stresses that “our experiments do not prove that Claude can experience or feel things like a human,” but acknowledges that its neural activity is indeed functionally accessible. This finding in itself is enough to provoke profound scientific and philosophical discussion.
Implications and Outlook
Anthropic says the discovery is already being used to monitor AI safety risks: by reading J‑space in real time, it can detect early signs of potentially risky output tendencies.
The related code has been open‑sourced, and an interactive demo has been launched in collaboration with Neuronpedia. This research provides an experimentally verifiable new path for understanding the “inner mind” of large models, moving the question of AI mentality from philosophical speculation to the scientific frontier.