Anthropic says its latest interpretability work gives researchers a sharper way to inspect what happens inside Claude as it turns prompts into answers, but the company also says the method still captures only a fraction of the model’s full computation.
The research announcement, titled *Tracing the thoughts of a large language model*, describes two new papers that Anthropic says move its internal-analysis tools from isolated features toward connected circuits. In practical terms, the work is meant to help researchers identify patterns of activity and information flow inside the model rather than relying only on outputs and conversational testing.
Anthropic says the project is designed to answer a basic but difficult question: how does a language model produce a response when no human programmer directly writes the reasoning steps in advance? The company says its approach borrows from neuroscience, where scientists study the internal workings of brains by tracing activity through different systems. Anthropic presents its own effort as an “AI microscope” intended to make model behavior more legible.
The company says the new method allowed it to examine Claude 3.5 Haiku on a range of short, representative tasks, including multilingual prompts, poetry generation, hallucination behavior and an example jailbreak. Anthropic says the findings were sometimes surprising. In one case, it says the model appeared to plan ahead in rhyming poetry rather than simply choosing the final word at the last second. In another, it says Claude’s default tendency is to refuse speculation, and that this reluctance can be overridden when something inhibits that default behavior. In a jailbreak study, Anthropic says the model recognized a request for dangerous information before it was able to redirect the conversation.
The company says the work also suggests a shared conceptual space across languages. It reports that some core features for concepts such as smallness and oppositeness activate across languages and can trigger a concept of largeness that is then translated into the language of the prompt. Anthropic says this shared circuitry becomes more pronounced as models scale up.
At the same time, the company is unusually direct about the limitations. It says the method only captures a fraction of the model’s total computation even on short prompts, and that the patterns it sees may include artifacts from the tools themselves. Anthropic also says the process still takes hours of human effort to interpret circuits for prompts containing only tens of words, which raises the question of how such methods could be applied to much longer, more complex chains of reasoning.
That tension runs through the announcement. Anthropic frames interpretability as one of its highest-risk, highest-reward research bets because it could help prove whether an AI system is aligned with human values and reliable enough to trust. But the same announcement also makes clear that this kind of visibility is still partial, slow and technically demanding.
The company says the effort is part of a wider portfolio that includes realtime monitoring, model-character work and alignment research. Its pitch is not that the new tools solve interpretability, but that they make the problem less blind. For researchers trying to understand a system as complex as Claude, even that narrower claim is meaningful.
