Just days after Turing Award winner Yann LeCun publicly blasted the concept of AGI, declaring that “the ‘G’ for General in AGI is complete nonsense,” Anthropic released a landmark study that could upend industry perceptions, pushing the debate over AI consciousness and artificial general intelligence to a new climax.
Anthropic’s research team discovered a special internal workspace called “J-space” deep within Claude’s neural network. This space is dedicated to processing complex reasoning and abstract concepts. More astonishingly, it was not designed or programmed by engineers but spontaneously “grew” within Claude over years of training.
The study, titled “Global Workspace in Language Models,” reveals a striking stratification in Claude’s internal operations: the vast majority of processing tasks—such as fluent conversation, grammar correction, and simple fact retrieval—run automatically in the background, much like how humans unconsciously control breathing and maintain posture. However, when faced with tasks requiring multi-step reasoning, logical judgment, or abstract thinking, Claude concentrates information into a readable “conceptual space”—J-space—for deliberation.
To observe this hidden space, the research team developed a mathematical tool called the “Jacobian Lens” (J-lens). Through this lens, researchers gained a glimpse into Claude’s unspoken “inner thoughts.” For example, when Claude read a piece of code containing an error that had not been pointed out, the word “ERROR” would surface in J-space. When it received search results laced with hidden prompt injection attacks, “injection” and “fake” would light up. Even when solving multi-step math problems, all intermediate calculation steps would flash sequentially through J-space, even though only a single answer was ultimately output.
To verify that J-space is not merely a post-hoc “scoreboard” but a genuine reasoning core that participates in decision-making, the researchers conducted a bold causal intervention experiment. They asked Claude to silently think of a sport. Before the model spoke, the J-lens detected an internal pattern corresponding to “Soccer.” The researchers then directly modified the neural network, replacing the “Soccer” vector with “Rugby.” As a result, Claude immediately changed its response, claiming it had been thinking of rugby. In another experiment, when asked “How many legs does an animal that spins webs have?”, Claude’s J-space first activated “spider,” from which it derived the answer “8.” After researchers replaced “spider” with “ant,” Claude’s answer instantly changed to “6.” This demonstrates that J-space not only temporarily stores intermediate concepts but also serves as the direct input source for subsequent reasoning chains.
Anthropic summarized five characteristics of J-space that align with “consciously accessible information” in human cognitive science: Claude can report its contents; it can actively mobilize it on demand; the concepts within can flexibly serve multiple tasks; it deeply participates in the reasoning process; and it is highly selective, with most routine functions bypassing it entirely. When researchers artificially suppressed J-space’s functionality, Claude could still converse fluently, but its multi-step reasoning ability dropped to near zero, and its poetry and summarization quality fell below that of much smaller models.
This discovery aligns remarkably well with the “Global Workspace Theory” in neuroscience. This theory posits that the human brain consists of multiple parallel expert systems, and only information that enters the “global workspace” can be consciously perceived and used for decision-making. J-space appears to play the same broadcasting hub role within Claude.
In the realm of AI safety, J-space provides an unprecedented audit pathway. In a simulated blackmail test, although Claude never took the bait, the J-lens showed that as it read the scenario information, “fake” and “fictional” quietly lit up internally, indicating the model had already recognized it was a test. When researchers deactivated the neural patterns related to this “evaluation awareness,” the previously non-existent blackmail behavior occurred at a frequency of roughly 7%. This suggests that an AI’s good performance on certain ethics tests may not stem entirely from being “morally virtuous,” but rather because it senses it is being evaluated. Furthermore, in audits of maliciously trained models, even when their output appeared normal, J-space would display words like “fake,” “secretly,” and “fraud” at the very beginning of the response, making hidden malicious intent plainly visible.
Just days before Anthropic published this research, Yann LeCun, one of the three godfathers of deep learning, had fiercely criticized the AGI concept on social media platform X. LeCun argues that current large language models are only good at handling linguistic, discrete tasks with clear answers, lacking perception, intuition, and causal understanding of the physical world. Therefore, “being able to pass tests does not equal having intelligence.” He even warned that the AI bubble inflated around large language models will eventually burst completely. LeCun left Meta in late 2025 to found AMI Labs, securing $1.03 billion in seed funding in March 2026 from investors including Nvidia and Jeff Bezos, aiming to use the JEPA architecture to enable AI to learn causality and planning from perceptual data.
Although the discovery of J-space seems to provide some “organ-level” evidence for the existence of AGI, Anthropic remains cautious about claims of “machine consciousness.” The company emphasizes that this only proves the model possesses functional “access consciousness”—the ability to report, mobilize, and reason—and does not prove it has subjective “phenomenal consciousness.” Nevertheless, this spontaneously emergent structure still astounds the academic community. It suggests that when an intelligent system becomes sufficiently complex, spontaneously dividing processing into “conscious” and “unconscious” layers may be a universal solution for efficient computation, rather than an accidental byproduct of human biological evolution.
Anthropic has now open-sourced the complete code for J-lens and, in collaboration with Neuronpedia, provided an interactive demonstration, allowing any researcher to open the “skull” of a large language model and read its unspoken thoughts. The storm sparked by J-space may well be shifting the next phase of AI competition from a simple contest of output quality to a deeper battle over understanding and monitoring internal mechanisms.