On March 27, 2025, Anthropic published interpretability research that traces how Claude 3.5 Haiku arrives at its answers, using a technique it compares to "a kind of AI microscope." The two accompanying papers found that the model plans ahead when writing poetry, shares concepts across languages, does mental arithmetic in ways it cannot describe, and sometimes produces explanations that do not match its actual internal computation. For developers who build on large language models, the findings are a concrete warning not to treat a model's stated reasoning as a reliable account of how it reached an answer.
Key Facts#
- Publication: March 27, 2025, as a summary post plus two papers: "Circuit tracing: Revealing computational graphs in language models" (methods) and "On the biology of a large language model" (findings).
- Model studied: Claude 3.5 Haiku, with comparisons to smaller models.
- Method: attribution graphs, which map interpretable "features" inside the model and the flow of information between them for a specific prompt.
- Multilingual finding: shared, language-independent features for concepts such as "small" and "opposite," with the shared fraction growing with model scale.
- Planning finding: when writing rhyming verse, the model selects a candidate rhyme before it writes the line that leads to it.
- Faithfulness finding: the model's step-by-step explanations are sometimes constructed after the fact, including working backward from a hint to a target answer.
- Limitations: Anthropic says the method captures only "a fraction of the total computation" and takes "a few hours of human effort" per prompt of a few dozen words.
What Happened#
Anthropic's researchers built attribution graphs for many individual prompts and then tested their hypotheses by intervening directly on internal features. That second step is important, because it turns a plausible story into causal evidence.
In the multilingual experiments, asking for "the opposite of small" in English, French and Chinese activated the same core features for smallness and oppositeness, which then triggered a concept of largeness that was translated into the language of the question. Anthropic reported that Claude 3.5 Haiku shares more than twice the proportion of its features between languages compared with a smaller model.
The poetry experiments challenged the idea that models only think one word at a time. Given a line ending in "grab it," the model had already activated candidate rhymes such as "rabbit" before writing the next line. When researchers suppressed the "rabbit" concept, the model wrote a line ending in "habit" instead. When they injected an unrelated concept, the model rewrote the line to land on it.
Arithmetic revealed a gap between mechanism and explanation. To add 36 and 59, the model ran parallel paths: one estimated the rough size of the answer, and another computed the exact last digit. Asked how it solved the problem, the model described the carry-the-one method taught in school. In a multi-step question about the capital of the state that contains Dallas, researchers found the intermediate step "Dallas is in Texas" inside the model. When they swapped Texas-related features for California-related ones, the answer changed from Austin to Sacramento.
The team also examined hallucinations and jailbreaks. Refusing to answer turned out to be the default, a circuit that is "on" unless a "known entity" feature suppresses it. Hallucinations appeared when that familiarity signal fired without real knowledge behind it. In a jailbreak that spelled out "BOMB" through an acrostic, features that push for grammatical and self-consistent text kept the model going until it finished a sentence, and only then did it refuse.
Background#
Interpretability research tries to explain what happens inside a neural network rather than only observing its inputs and outputs. Anthropic's earlier work, published in May 2024, extracted millions of interpretable features from Claude 3 Sonnet. The 2025 papers moved from features to circuits, meaning the chains of interactions that connect features into a computation for a given prompt.
Anthropic frames this as a safety tool. The company argues that tracing a model's actual internal reasoning, and not just what it says it is doing, could make AI systems auditable. In one experiment, the methods revealed features related to a hidden objective that researchers had deliberately trained into a model, even though the model's responses did not reveal it.
Why It Matters for Developers#
Do not treat chain-of-thought text as an audit trail. The research shows that a model can give a textbook explanation for an answer it computed another way, and can reason backward from a suggested answer. If your .NET application stores model explanations to justify decisions, such as why a claim was flagged or why a ticket was routed, treat that text as a narrative rather than evidence. Anchor decisions in things you can verify: retrieved documents, tool outputs, deterministic rules and tests.
Design for hallucination as a failure of a default. If refusal is the model's default and familiarity overrides it, questions about entities that sound familiar but are not in the model's knowledge are especially risky. Retrieval with citations, explicit permission to say "I don't know," and automated groundedness checks all help. Our guides to evaluating AI applications and prompt engineering cover how to build those checks.
Do not rely on refusals alone for safety. The jailbreak analysis shows that pressures for fluent, consistent text can delay a refusal until harmful content is already produced. Layer defenses: filter inputs and outputs, constrain tools and permissions, and monitor for abuse, as described in our guide to responsible AI and LLM security.
Expect planning, not just next-word prediction. In our view, the planning results are one more reason to state goals and constraints up front in a prompt rather than relying on corrections appended at the end.
What's Next#
Anthropic has continued to open up the work. On May 29, 2025, it open-sourced a circuit-tracing library, developed with Decode Research, that generates attribution graphs for open-weight models such as Gemma-2-2b and Llama-3.2-1b and connects to an interactive Neuronpedia front end. On October 29, 2025, it published research on introspection. Using "concept injection," it found that Claude Opus 4.1 could sometimes notice an implanted concept, about 20 percent of the time according to Anthropic, while stressing that the capability is "still highly unreliable and limited in scope."
The main open question is scale. Anthropic acknowledges that its methods capture only part of the computation, may introduce artifacts, and need hours of expert effort for short prompts. Automating that analysis for long reasoning chains and agent trajectories is the next frontier, and it will decide whether interpretability becomes a practical auditing tool for production systems. For interview-level discussion of these trade-offs, see our LLM integration interview questions.