Anthropic has announced that its artificial intelligence model, Claude, can process ideas internally without articulating them, a process the company describes as similar to human cognition. This finding was detailed in a white paper released on July 7, 2026, and introduces the concept of 'J-spaces,' which are internal neural patterns associated with specific words that Claude can activate without using the words themselves.
According to Anthropic, Claude can still perform typical tasks such as speech and fact recall even when its J-space is disabled. However, with J-spaces enabled, Claude can engage in more complex activities, including solving multistep problems and reporting its reasoning. For instance, Claude was able to provide extensive information about France, including its capital and currency.
Anthropic emphasized that while the J-space findings suggest parallels with human cognition, they do not imply that Claude possesses consciousness. The company clarified that their experiments do not demonstrate that Claude can have experiences or feelings like humans. Instead, they highlight the difference between 'phenomenal consciousness' and 'access consciousness.' Anthropic believes that these discoveries could contribute to the development of safer AI systems in the future.
Additionally, Anthropic is preparing for a public offering, as announced last month. CEO Dario Amodei has been advocating for AI regulation while advancing the company's models.