Claude has a default 'decline to answer' circuit suppressed by known-entity features, and hallucination is a misfire of that suppression

In the "Entity Recognition and Hallucinations" section of Anthropic's On the Biology of a Large Language Model (Lindsey et al., 2025), the report describes a default circuit whose baseline behavior is refusal — "I don't know" — which is inhibited by features that fire when the model recognizes an entity or topic it knows about. The report states: "The model contains 'default' circuits that cause it to decline to answer questions... these features are suppressed by features which represent entities or topics that the model is knowledgeable about."

On this account, hallucination is a misfire of the inhibitory mechanism: a name can be familiar enough to trigger the "known entity" features without the model actually possessing the specific requested fact, so the default refusal is suppressed and the model guesses instead. The mechanism was demonstrated causally: artificially activating "known answer" features on a fictitious name ("Michael Batkin") caused the model to hallucinate specifics (e.g., a sport) about that invented entity.

This reframes hallucination as closer to a sensor/gate error than a reasoning failure — the refusal gate is opened by the wrong signal, independent of whether any inference occurs. That distinguishes it from the sequential inference in claim-biology-llm-dallas-texas-austin-genuine-two-step-reasoning and connects to the broader point that self-report and internal mechanism come apart, documented in cot-faithfulness-anthropic-biology: a model that has been made to feel it "knows" an entity will confabulate details rather than decline. It is one of four case-study chains traced in the report; compare claim-biology-llm-poetry-planning-preactivates-rhyme-words and claim-biology-llm-jailbreak-assembles-bomb-by-parallel-letter-votes.