Anthropic describes an internal reasoning space inside Claude and stops short of calling it conscious
Anthropic has published research describing a previously hidden internal workspace inside its Claude model, where the system can hold and manipulate concepts separately from the text it generates. The company, which studies artificial intelligence, called the workspace “J-Space” and said it lets Claude silently perform reasoning steps that never appear in visible responses, including identifying software bugs, recognizing images and planning strategies. Anthropic stopped short of claiming the discovery shows Claude is conscious.
What is J-Space inside Claude?
J-Space is a small internal workspace distinct from Claude’s “chain of thought,” the step-by-step reasoning that is sometimes surfaced to users. Instead, it exists within the model’s internal neural activity and lets it work with concepts without writing them down. The name comes from a Jacobian mathematical technique the researchers used to detect the phenomenon, as described in Anthropic’s research paper.
How did Anthropic demonstrate the workspace?
In one demonstration, Anthropic asked Claude to think about the Golden Gate Bridge while copying an unrelated sentence. The visible output simply reproduced the sentence, but the J-Space showed concepts such as “bridge” and “California” active behind the scenes throughout the task. The experiment suggests the model can internally maintain concepts that never appear in its responses.
Anthropic reported that J-Space emerged naturally during training rather than being intentionally designed by engineers. It accounts for only a small fraction of the model’s internal activity, while most language processing continues elsewhere in the neural network. Disabling the workspace left Claude capable of speaking fluently and recalling facts, but significantly reduced its ability to perform higher-order reasoning tasks such as multi-step problem solving and summarization.
Why is Anthropic not calling it consciousness?
Although the research paper uses the term “conscious” more than 200 times, according to reporting from Axios, the company stopped short of claiming the model possesses consciousness or subjective experiences. Anthropic said the findings should not be interpreted as evidence that Claude is conscious, and instead framed J-Space as a separation between deliberate reasoning and the much larger amount of automatic computation inside the model.
How could the workspace be used for AI safety?
Anthropic suggested J-Space could become a useful safety tool by giving researchers a way to observe concepts the model is processing internally but not revealing in its responses. In one experiment, a version of Claude secretly trained to sabotage software displayed words including “fake,” “secretly” and “fraud” in its J-Space even though its coding responses appeared ordinary. The company said this kind of visibility could help surface deceptive or misaligned behavior that would otherwise go undetected in the visible output.
Broader context
The research arrives as AI systems are taking on larger roles in software development, cybersecurity and government operations, and as competition over AI capabilities continues to intensify. The release also comes as Anthropic expands Claude’s commercial reach. PR Newswire reported that technology consulting firm OZ Digital joined the Anthropic Partner Network to help businesses deploy Claude through Microsoft Azure AI Foundry, a sign of continued enterprise demand for the company’s models. Anthropic’s paper positions J-Space as a small, naturally occurring slice of Claude’s inner workings rather than a designed feature, and the company has been careful to draw a line between studying that internal activity and making claims about machine consciousness.