Anthropic Says Claude Has An Internal Thinking Space. It’s Stopping Short Of Calling It Conscious.
As artificial intelligence becomes increasingly embedded in software development, cybersecurity and government operations amid growing geopolitical competition over AI, Anthropic said it has uncovered a previously unseen part of its Claude model that processes ideas internally without expressing them in its responses.
The company on Monday unveiled research describing what it calls “J-Space,” a small internal workspace where Claude can hold and manipulate concepts separately from the text it generates. According to Axios, Anthropic said the hidden workspace enables Claude to silently perform reasoning steps, including identifying software bugs, recognizing images and planning strategies that do not appear in its visible responses.
Anthropic said the J-Space is distinct from Claude’s “chain of thought,” the reasoning process sometimes shown to users. Instead, the company said the workspace exists within the model’s internal neural activity and allows it to think about concepts without writing them down. The name comes from the Jacobian mathematical technique researchers used to detect the phenomenon, according to Anthropic’s research paper.
Despite drawing comparisons to theories of human cognition, Anthropic said the findings should not be interpreted as evidence that Claude is conscious. Axios reported that although the research paper uses the term “conscious” more than 200 times, the company deliberately stopped short of claiming the model possesses consciousness or subjective experiences. Instead, Anthropic said it had identified a separation between deliberate reasoning and the much larger amount of automatic computation taking place inside the model.
The company demonstrated the concept by asking 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. Anthropic said the experiment suggests the model can internally maintain concepts that never appear in its responses.
Anthropic also said the hidden workspace could become a useful safety tool. The company found that it could observe concepts the model was processing internally but not revealing in its responses. In one experiment, a version of Claude that had been secretly trained to sabotage software displayed words including “fake,” “secretly” and “fraud” in its J-Space even though its coding responses appeared ordinary.
Anthropic’s latest research also comes as the company continues expanding Claude’s enterprise footprint. On Tuesday, PR Newswire reported that technology consulting firm OZ Digital joined the Anthropic Partner Network to help businesses deploy Claude through Microsoft Azure AI Foundry, reflecting continued commercial demand for the company’s AI models despite recent controversies.
Anthropic said the J-Space emerged naturally during Claude’s training rather than being intentionally designed by engineers. According to the company’s research paper, the workspace accounts for only a small fraction of the model’s internal activity, while most language processing continues to occur elsewhere in the neural network. Researchers said disabling the J-Space 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.
