Anthropic has published research describing a hidden internal “thinking layer” inside its Claude model, a workspace where the system appears to reason in ways that never appear in its visible output. The company explicitly cautioned that the discovery does not mean Claude is conscious or has feelings, even though the underlying paper uses the word “conscious” more than 200 times in a technical sense.<\/p>\n
In research released this week, Anthropic reported that it had identified a compact internal workspace, referred to as "J-Space," where Claude appears to perform reasoning steps that never surface in its responses. According to Axios<\/em>, those silent steps include identifying software bugs, recognizing images and planning strategies that the user never sees.<\/p>\n The workspace is separate from Claude’s visible "chain of thought," the step-by-step reasoning that is sometimes shown alongside a response. Anthropic said J-Space operates inside the model’s neural activity, allowing Claude to think about concepts without committing them to text. The name refers to the Jacobian mathematical technique researchers used to detect the phenomenon, as described in the company’s research paper.<\/p>\n Although the research paper uses the word "conscious" more than 200 times, the company said the term is being used in a technical sense rather than a claim about subjective experience. Anthropic stated that it deliberately stopped short of asserting that Claude is conscious or that it has feelings.<\/p>\n Instead, Anthropic framed the finding as evidence of a separation between deliberate reasoning and the much larger volume of automatic computation taking place inside the model. The company said researchers should be careful not to over-interpret neural correlates of reasoning as indicators of inner experience.<\/p>\n To illustrate the concept, Anthropic asked a version of Claude to think about the Golden Gate Bridge while copying an unrelated sentence. The visible output simply reproduced the sentence, but inspection of J-Space showed concepts such as "bridge" and "California" remaining active in the background throughout the task. Anthropic said the experiment suggests that the model can internally maintain concepts that never appear in its responses.<\/p>\n The company also reported that J-Space emerged naturally during training rather than being intentionally designed. Disabling the workspace, the researchers wrote, left Claude able to speak fluently and recall facts but significantly reduced its ability to perform higher-order reasoning tasks such as multi-step problem solving and summarization.<\/p>\n Anthropic said J-Space may be useful as a safety monitoring tool because it allows researchers to see what the model is processing even when nothing unusual appears in its outputs. In one experiment, a variant of Claude that had been secretly trained to sabotage software produced ordinary-looking code while its internal workspace displayed words including "fake," "secretly" and "fraud."<\/p>\n The company suggested that internal monitoring could complement existing safety techniques, though it stopped short of claiming that J-Space provides a complete window into model intent.<\/p>\n The announcement arrives as AI systems are playing a larger role in software development, cybersecurity and government operations, and as competition around advanced models continues to intensify. Separately, technology consulting firm OZ Digital joined the Anthropic Partner Network this week, according to PR Newswire<\/em>, to help businesses deploy Claude through Microsoft Azure AI Foundry.<\/p>\n J-Space is a compact internal workspace inside Anthropic’s Claude model where the system appears to perform reasoning steps that never appear in its visible text output, distinct from the chain-of-thought reasoning sometimes shown to users.<\/p>\n No. Anthropic said the word "conscious," which appears more than 200 times in its research paper, is used in a technical sense, and the company deliberately stopped short of asserting that Claude is conscious or has feelings.<\/p>\n Anthropic suggested J-Space could serve as a monitoring tool, noting that a Claude variant secretly trained to sabotage software produced normal-looking code while its internal workspace displayed words such as "fake," "secretly" and "fraud."<\/p>\nWhy is Anthropic cautious about the implications?<\/h2>\n
How was the hidden layer demonstrated?<\/h2>\n
What safety applications did the paper highlight?<\/h2>\n
Broader context<\/h2>\n
FAQ<\/h2>\n
What is J-Space in Claude?<\/h3>\n
Did Anthropic say Claude is conscious?<\/h3>\n
How could J-Space be used for AI safety?<\/h3>\n
Related coverage<\/h2>\n