Anthropic finds Claude reasons with abstract concepts that never become words

Anthropic has found that its Claude language model developed, on its own, an internal "system" space for storing and processing concepts that aren't expressed in words. The mechanism looks surprisingly similar to the way the human brain handles thoughts before putting them into words. The company announced the finding itself and noted that it surprised even its own researchers.

What the researchers found

Anthropic's team studied the model's internal representations: how Claude stores and links concepts inside its layers, well before it produces any finished text. They found that the model processes some concepts in abstract form. Instead of translating them into specific words, it unpacks them into text only later, at the output stage.

• Anthropic describes the mechanism as a "system space" for concepts that haven't been put into words
• The company stresses the similarity to how human thinking handles ideas before they are spoken
• The finding came out of interpretability research, which tries to understand what happens inside a model, not just what it outputs

Why companies are looking inside the model

For a long time, large language models worked as a "black box": developers could see the input and the output, but not what happened in between. Anthropic is one of the leaders in interpretability, a field that tries to locate individual concepts inside neural networks and trace how they are processed. The company has been publishing research for several years on how models build internal representations of concepts, and it periodically reports new behavior patterns that catch even its own researchers off guard.

Findings like this matter beyond science. Understanding how a model processes information before it generates a response helps explain its decisions more precisely, track down the causes of errors and hallucinations, and build trust in systems that are increasingly used in business and everyday tasks.

What it means

If models really do build their own structures for storing meaning, structures that can't always be translated into words, that changes how we think about the depth of "thinking" in modern language models. It also makes the job of fully interpreting and controlling them harder.