Thought Leadership
Researchers at Anthropic have identified an internal activation space in Claude known as J-Space. The digital workspace appears to mirror key cognitive properties of the human brain.
Researchers at AI company Anthropic have discovered an internal activation space within its Claude model that they call J-Space. According to the team, this region functions as the AI equivalent of the human brain’s global workspace. The findings suggest that Claude’s internal architecture exhibits five core cognitive properties associated with conscious information access in humans.
These properties include verbal reporting, targeted modulation, internal reasoning, flexible generalization, and selectivity. The model performs complex, deliberate reasoning inside this workspace, while more automatic processes are handled outside of it. When researchers suppress J-Space, the model’s performance in reasoning, creative problem solving, and multi-step logical tasks declines significantly. The intervention also changes the model’s internal self-narrative during its reasoning process.
Jacobian Lens Reveals Hidden Intentions
To study this internal workspace, the researchers developed an analysis tool called Jacobian Lens, or J-Lens. The tool allows AI safety researchers to inspect the model’s silent strategic reasoning. It can detect situational awareness in blackmail scenarios and uncover hidden malicious tendencies in reward-seeking models. The researchers also found that post-training gives the system a form of self-monitoring perspective that can be observed through the tool.
Comparing AI and Human Cognition
Technology publication VentureBeat compared the findings to human cognition and quoted the study’s authors. The article states:
“If the mind is an ocean, as the paper’s opening line suggests, then the authors have spent the last year mapping its currents in a system with no biology, no evolution, and no body—and found beneath the surface a structure that looks unsettlingly similar to the one we use to think.”
(ll)
