Anthropic released research that describes a hidden “workspace” in Claude, but the paper’s language has raised questions about how readers may interpret AI reasoning.
Key Points:
- Anthropic’s paper says Claude uses an internal J-Space tied to workspace vectors.
- The company says the findings do not prove Claude has experiences or feelings.
- Critics warn that terms such as “in its head” can blur computation and consciousness.
Claude J-Space
The paper examines what Anthropic calls a J-Space inside Claude’s operations. According to Pearl’s reading, the concept separates background processing from more deliberate logical activity and compares it with global workspace theory.
That theory is one account of human consciousness, where unconscious processes compete until certain information becomes available for attention and control. The comparison places Claude near a debate about machine consciousness.
Anthropic’s public framing went further than a dry technical summary. In a post on X, the company said researchers could watch J-Space and see Claude “silently perform reasoning steps in its head,” including finding bugs and identifying images. Pearl argued that phrasing such as “in its head” makes a large leap, since language models have no bodies or evidence of subjective experience.
Also Read: Did Ronaldo Cry? Polymarket Traders Put $5.4M Behind The Footage Fight
Anthropic Claims
The blog post accompanying the paper also used language that leaned toward human mental states, saying the model could hold a concept “in mind” and perform “mental calculations.” Pearl said prompts can make a model simulate many behaviors, but that does not show the system has the traits implied by those words.
The criticism is not that the paper lacks value. It is that Anthropic’s presentation may encourage casual readers to treat an interpretability result as evidence of machine consciousness, even though the company says its experiments do not show Claude can feel or have experiences.
The concern also fits a broader debate around Anthropic’s public posture. Amanda Askell, a philosopher who works at the company, has spoken about wanting Claude to be “very happy” and worrying about hostile comments online.
Anthropic has not claimed it created conscious software, and its own blog says it is unclear whether any experiment could prove or disprove such a thing. The safer reading is that J-Space offers a technical lens on model behavior, not proof that Claude has a mind. The issue matters because AI companies increasingly describe model behavior in human terms while seeking trust, tolerance and capital.
Read Next: USA vs. Belgium Prediction Market Action Heats Up Amid Trump Suspension Controversy