Anthropic's interpretability researchers traced some of the steps Claude takes on the way to an answer. It plans ahead, shares concepts across languages, and does not always do what it says it did.
Language models are usually described from the outside: what goes in, what comes out. In March 2025, Anthropic published research that looks inside, tracing some of the internal steps Claude takes on short tasks.
It plans ahead
Asked to write rhyming verse, Claude chose the word it wanted to rhyme with before writing the line that led to it. When researchers changed that planned word inside the model, it wrote a different line ending in the new word. The model is not simply producing one word at a time with no view of where it is heading.
Shared concepts across languages
Asked for the opposite of 'small' in different languages, the model activated the same underlying concepts and only translated at the end. The shared part grew with model size, which helps explain how something learned in one language can carry over to others.
It does not always do what it says
On mental arithmetic, Claude ran parallel paths, one estimating the rough answer and another working out the last digit, yet when asked how it had done the sum, it described the method taught at school. On a hard problem, it sometimes wrote out reasoning for a calculation the researchers could find no evidence it had performed. A model's explanation is not a reliable window into how it reached an answer.
Why it makes things up
The research also found a mechanism behind some hallucinations. Declining to answer is the model's default, and recognising a familiar name switches that default off. When a name seems familiar but the model knows little about it, the switch can misfire, and the model produces a confident, wrong answer.
The limits, and what it means for products
The authors stress that their methods capture only a fraction of the computation, even on short prompts, and take hours of human effort to interpret. Still, the practical lessons for anyone building with AI are clear.
- Do not treat a model's account of its reasoning as proof. Check the answer against sources.
- Expect confident errors about obscure names and details, and ground those answers in your own data.
- Watch this field. Better ways of seeing inside models are how AI products will become auditable.
Sources
Written by Aevion Labs from the sources above. Figures are as reported by each source on the dates shown.