Anthropic has released new interpretability research that offers insight into the internal workings of Claude, how it processes language, reasons through tasks, and generates responses. Rather than relying solely on outputs, this work involves tracing the model’s internal computations, revealing patterns that help explain its behavior.

Drawing inspiration from neuroscience, the team has developed tools to examine “features” and “circuits” inside the model. These tools helped uncover how Claude handles core tasks like translation, reasoning, poetry, and math. Key findings include:

1. Multilingual Concepts: Claude shows evidence of shared internal features across languages, suggesting it operates in a common conceptual space before translating into the requested language.
2. Poetry and Planning: In composing rhymed verse, Claude plans ahead, thinking of rhyming words before generating a full sentence, rather than picking a rhyme at the end.
3. Mental Arithmetic: Instead of mimicking human math strategies, Claude appears to use multiple computational paths, some for approximating, others for computing specific digits.
4. Faithfulness of Explanations: When reasoning through simpler math, Claude’s internal steps align with its written explanation. But with more difficult tasks, it may construct plausible steps after arriving at an answer.
5. Defaults and Hallucinations: The model tends to refuse to answer questions by default. Only when it detects sufficient internal signals of knowledge does it proceed. Errors can happen when these signals are misapplied, leading to hallucinations.
6. Jailbreak Dynamics: In some adversarial cases, internal features for coherence and grammar may momentarily override safety mechanisms. The model often recovers once the sentence concludes, offering a refusal afterward.

These insights, while limited to specific tasks and model instances, demonstrate how interpretability tools can reveal otherwise hidden processes. The work remains time-intensive and incomplete, each example studied represents only a small fraction of the model’s total behavior, but it marks meaningful progress toward transparency in AI systems.

Explore the full research and case studies on Anthropic's platform:

[https://www.anthropic.com/research/tracing-thoughts-language-model?utm_source=alphasignal]