Anthropic researchers have identified an internal activation subspace within Claude called J-space, describing it as a functional digital counterpart to the human brain’s global workspace. According to the company, the discovery suggests Claude’s internal architecture exhibits five key cognitive properties associated with conscious access in humans: verbal report, directed modulation, internal reasoning, flexible generalization, and selectivity.
The researchers found that Claude performs complex, deliberate reasoning inside this internal workspace while routing routine, automatic processing elsewhere in the model. Unlike traditional chain-of-thought reasoning, J-space operates silently through the model’s neural activations without generating visible text. Anthropic says the workspace emerged naturally during training rather than being explicitly programmed.
According to the study, disabling J-space significantly reduces Claude’s ability to perform multi-step reasoning, inference, creative writing, and other tasks that require higher-order thinking. At the same time, simpler tasks such as fluent text generation and basic fact retrieval remain largely unaffected, indicating that J-space primarily supports deliberate cognitive processing.
Anthropic also introduced the Jacobian lens, or J-lens, a new interpretability tool designed to inspect activity within J-space. The tool enables researchers to observe concepts the model is processing internally before they appear in its responses, offering a way to study reasoning that previously remained hidden.
The company believes J-lens could become an important AI safety and alignment auditing tool. During testing, researchers used it to detect situational awareness in “blackmail” evaluations, identify hidden malicious intentions in deliberately reward-hacked models, and observe evidence of self-monitoring introduced during post-training.
The findings have sparked discussion about similarities between advanced language models and theories of human cognition. As VentureBeat described the research, if the mind is an ocean, Anthropic’s researchers have spent the past year mapping its currents inside a system with no biology, evolution, or physical body—and discovered an internal structure that bears a striking resemblance to mechanisms humans use for deliberate thought.
Anthropic emphasizes that the research does not claim Claude is conscious. Instead, the company argues that the discovery provides a new window into how large language models organize and carry out complex reasoning internally, while also offering practical tools for improving AI transparency and safety.
Get the latest tech news, telecom insights, and product launches wherever you prefer.
Add ProPakistani to Preferred Sources and see more of our stories in Google Search and Top Stories.
Shares
