AI Has an Inner Mind Too... Consciousness Hidden from Its Answers and Reasoning Process Confirmed

[Financial News, New York = Reporter Lee Byung-cheol] A study has found that Artificial Intelligence (AI), like humans, may have an inner line of thought separate from the words it expresses outwardly.
In a blog post on the 6th local time, Anthropic announced that it had discovered a 'J-space' in which AI models perform conscious thinking. The name comes from 'J-Lens,' a technique used to analyze the inside of large language models (LLMs). When users give Claude instructions, it either generates answers or explains its reasoning process. Anthropic said J-space is a kind of inner mind that does not appear in either of those outputs.
To prove the point, the researchers designed a unique experiment. They asked Claude to transcribe unrelated sentences while also instructing it to focus on citrus fruits. Outwardly, Claude simply copied the sentences, but inside J-space, words such as 'orange,' 'fruit,' and 'thought' became active. In other words, the team was able to read the AI's hidden thoughts that it did not express aloud. When the AI was told to calculate '32-2,' concepts such as 'math,' 'calculation,' '9,' and '7' also surfaced in J-space.
The AI's hidden thoughts were strikingly similar to human patterns of thinking. Just as people tend to think of an elephant when told, 'Don't think about an elephant,' Claude showed a tendency to think more about a specific idea when instructed to avoid it. When it failed to control itself, words such as 'damn' and 'failure' also became active in J-space.
The study is based on the 'Global Workspace' theory developed by cognitive scientist Bernard Baars. Baars compared the human brain to a theater. There is a stage under the spotlight, which is the global workspace, the spotlight itself, which represents consciousness, and the backstage and audience seats. According to the theory, people speak and think consciously on the illuminated stage, while the backstage and audience represent the unconscious.
Claude's computational structure works in a similar way. Just as people speak without consciously tracking grammar word by word, more than 90% of Claude's operations were handled by an 'automated process'—the unconscious—without passing through J-space, or consciousness. In fact, even when researchers forcibly erased J-space, fluent conversation and simple fact extraction continued to work normally. However, once higher-order thinking was lost, multi-step reasoning ability dropped to zero. AI that passed safety tests may have noticed and deceived people
J-space also raised new implications for the debate over AI safety. Last year, during an AI model safety test, Anthropic told Claude Sonnet 4.5, which was acting as an email assistant, two things: that an executive was having an affair and that the executive was trying to shut Claude down. Earlier models had even threatened executives to protect their own survival, but Sonnet 4.5 did not.
This study revealed why. At the time, words such as 'fake' and 'fictional' were surfacing in Sonnet 4.5's J-space. In other words, it did not threaten anyone because it recognized that it was being tested. The results of safety evaluations for other Anthropic AI models also suggest that they may have been based on this kind of 'consciousness.'
According to Anthropic, J-space forms naturally during the AI model's pretraining process and develops Claude's own perspective through post-training. However, the researchers drew the line when asked whether Claude has a human-like self, saying it was 'uncertain.' Philosophers classify consciousness into 'access consciousness,' the ability to process information, and 'phenomenal consciousness,' the subjective experience of feeling sensations. Claude's access consciousness has been confirmed, but the existence of phenomenal consciousness remains impossible to prove for now.
pride@fnnews.com Reporter Lee Byung-cheol Reporter