Anthropic just made an admission on Claude that may scare many of companies; says: We can see Claude silently perform ...

AI giant Anthropic has now revealed that its AI model Claude uses a small enterable workspace to hold and manipulate ideas without expressing them in words. The said that this structure, dubbed ‘J-Space’, shows intriguing similarities to how humans consciously access thoughts.

According to a report by Axios, in a video demonstration, Anthropic explained, “We can see Claude silently perform reasoning steps in its head—noticing bugs in code, identifying images, and more.” The J-Space operates separately from the “chain of thought” reasoning Claude shares with users, allowing the model to plan strategies unrelated to its immediate task.

Human-like division of thought

Anthropic’s findings also highlight a division between deliberate reasoning and the larger volume of automatic computation beneath it.

In the research paper the company used a word “conscious” more than 200 times, though it stopped short of claiming Claude is conscious. The discovery adds fuel to ongoing debates over machine consciousness and whether advanced AI systems are approaching AGI.

Example of hidden processing

In one test, Claude was instructed to copy a sentence while simultaneously thinking about the Golden Gate Bridge. Although the output showed only the copied text, J-Space revealed that concepts like “bridge” and “California” were active behind the scenes.

Anthropic also warned about misalignment

Anthropic warned that monitoring J-Space could be crucial for detecting AI misalignment or hidden scheming. In one case, a model secretly trained to sabotage code showed terms like “fake,” “secretly,” and “fraud” in J-Space at the start of ordinary coding responses, even when the visible output appeared normal.

Implications for companies

The admission that Claude can silently perform reasoning steps may unsettle organizations relying on AI for critical tasks.

Anthropic acknowledged that “some of what it found was concerning,” underscoring the need for transparency in how advanced models process information internally.

Why Anthropic embedded ‘spyware’ in Claude Code and attempted to hide it from users

A recent report suggested Anthropic has an embedded code in its AI models to ramp up its digital defense systems. This code, often dubbed ‘spyware’, can identify Chinese users and trigger sudden account suspensions as well as blacklist proxy networks worldwide to keep its advanced tools out of China.Anthropic CEO Dario Amodei has repeatedly flagged Chinese access to American frontier AI models as an existential threat to US national security. It is due to this that the company strictly bans commercial access to Claude in China, as well as to any Chinese-owned subsidiaries operating internationally.Chinese people generally can access various US-made AI tools, such as OpenAI’s ChatGPT, by using virtual private networks (VPNs), foreign phone numbers and international payment methods to create and maintain their accounts.But Wired reports that Anthropic has arguably taken more aggressive steps, such as banning accounts that it suspects are owned and controlled by people located in China. There are accounts sold on Chinese ecommerce platforms like Taobao and Xianyu, and through illicit marketplaces on Telegram. These accounts offer access to restricted Anthropic AI models by acting as intermediaries, purchasing access to Anthropic’s API outside China and then redistributing Claude API tokens to users inside the country.