[DigitalToday intern reporter Seung-a Yoo] Anthropic said it has newly detected, inside its large language model (LLM), a space of words that do not appear in answers but influence the reasoning process.

MIT Technology Review reported on July 13 that the research is an extension of Anthropic’s long-term mechanistic interpretability work to interpret how AI models operate internally. The outlet examined the significance of the research through an interview with senior editor Will Douglas Heaven (윌 더글러스 헤븐), who holds a doctorate in computer science.

The key is an internal area Anthropic calls “J-space.” Words exist in this space that do not directly appear in the final output, and they were found to be involved in how the model solves problems or makes judgments. Anthropic said it confirmed the area after analysing Claude using a new probing technique.

These internal words vary in role. Some were used to track how far the model had progressed on a specific task, while others appeared like a moment of recognising a specific object when only limited clues were given. A representative case is that when only letters of a protein sequence were provided, a concept corresponding to “protein” surfaced internally. In another case, the word “panic” appeared when Claude moved in a direction aimed at cheating on a coding test.

Anthropic also said it confirmed that the model can describe and manipulate words in this space. That suggests the signals formed internally may be used in practice rather than being a simple byproduct. Dario Amodei (다리오 아모데이), Anthropic’s chief executive officer, has said that to sufficiently control an LLM, it is necessary first to better understand how it works.

It drew a line, however, at interpretations that immediately link the results to AI “thought” or “consciousness.” LLMs are “not magic but mathematics,” but they are also a highly complex system involving hundreds of billions of numbers and many calculations, it said. It also cited an analogy that printing a mid-sized LLM on paper would be enough to cover the entire city of San Francisco. It said specialised tools are therefore needed to look into a model from the outside by pinpointing a specific part at a specific moment.

There was also caution about language that likens LLMs to the human brain. It is clear that an LLM is “not a brain,” and such language can lead to misunderstandings that it has more human-like capabilities than it does, it said. It also noted that researchers use terms such as thinking or understanding for convenience because they lack sufficient alternative vocabulary to describe internal model phenomena.

Anthropic also kept its distance from claims that its structure is exactly the same as the human brain. In a statement, the company said, “These analogies helped design experiments,” adding that they made several counterintuitive predictions about J-space and they turned out to be correct. It also said, “There are important differences between J-space and the human brain, and we are not claiming a perfect correspondence.”

The context of the research also cited a past case in which Anthropic warned that the coding ability of its new model was so strong it could pose a global cybersecurity threat, after which the U.S. government temporarily suspended access to the model. It was also pointed out that such a precedent aligns with a narrative Anthropic has repeatedly stressed: that it created “very mysterious technology,” but the entity that can ultimately reveal it is itself.

The immediate problem the research aims to address is model monitoring. Anthropic believes that monitoring J-space could capture behavioural signals that are easy to miss from output alone, such as the process of assessing whether the model may produce biased responses or break rules. It aims to connect this to a control method that detects early signs of inappropriate behaviour by a model.

Even so, it is difficult at this stage to view the achievement as an independent practical technology. The discovery is closer to a step forward in understanding LLMs overall than to a safety device that can be applied immediately. Future points to watch include whether J-space can lead to an actual monitoring system and how much interpretability research can contribute to expanding controllability.