Anthropic discovers that Claude thinks in abstract concepts without words

Anthropic discovered that its language model Claude independently formed an internal "system" space for storing and processing concepts that cannot be expressed in words — a mechanism unexpectedly reminiscent of how the human brain operates with thoughts before verbalizing them. The company itself noted that the finding surprised even the researchers.

What exactly the researchers found

Anthropic specialists studied the model's internal representations — how Claude stores and connects concepts within its layers, long before it outputs finished text. It turned out that the model processes some concepts in abstract form, without translating them into specific words of the language, only later unfolding them into text on output.

• Anthropic describes the discovered mechanism as a "system space" for concepts not clothed in words
• The company emphasizes the similarity to how human thinking operates with ideas before voicing them
• The discovery was made as part of interpretability research — attempts to understand what happens inside the model, not just what it outputs

Why companies look inside the model

Large language models have long worked as a "black box": developers saw input and output, but not what happened between them. Anthropic is one of the leaders in the interpretability field, which tries to find individual concepts and processing chains inside neural networks. The company has been publishing interpretability research for several years and periodically reports new patterns of behavior that surprise even the researchers themselves.

Such discoveries are important not only from a scientific standpoint. Understanding how the model processes information before generating a response helps explain its decisions more accurately, find causes of errors and hallucinations, and increase trust in systems that are increasingly being applied in business and everyday tasks.

What it means

If models really do form their own structures for storing meaning that are not always translatable into words, this changes our understanding of how deeply "thinking" is organized in modern language models — and simultaneously complicates the task of their full interpretability and control.