This neuron detects formatting and structural markup in the text (headings, emphasis/bold markers, section bullets and similar layout tokens).

discussions centered on remote/hybrid work and return-to-office topics, including policies, practices, and collaboration for distributed teams.

Tokens that are part of the model/assistant's detailed explanatory responses (i.e., content words in the assistant's reply).

statements where the assistant refuses a request by citing safety rules, limits, or that it is "programmed" to be safe (i.e., refusal/safety-policy language).

The neuron is essentially flagging the assistant’s own “long‐form” explanation turns (the multi‐paragraph, bullet‐list responses) as opposed to user utterances. In other words, it turns on for tokens in the model’s detailed breakdowns.

assistant safety-refusal boilerplate: declarations that the AI cannot comply and references to its safety guidelines, ethical principles, and programming by its creators.

discourse connectors and prepositional function words that signal relationships and structure within explanations or requests.

sentences or passages where the assistant introduces itself or describes its identity, training, capabilities, and availability.

Tokens that occur in the model's long explanatory responses (assistant-generated, contentful reply text).

sentences where the assistant refers to itself and issues safety/refusal disclaimers (e.g., "I am programmed..." / "As such, I cannot...").

This neuron detects the model’s self-description phrase “As a large language model” (and similar self-referential disclaimers).

The neuron strongly activates on the pattern where the model refers to itself as “a large language model,” i.e. self-identification phrases stating “As a large language model…”

the start of an assistant/model reply (the token marking the beginning of the model's response).

The neuron is identifying when the model is discussing its own nature, limitations, or role as an AI language model.

mentions of artificial intelligence, especially references to AI assistants, models, or AI-powered technologies.

The neuron strongly activates on the acronym “AI.”

This neuron spots the assistant’s self-descriptive policy-and-safety statements, especially “I cannot/absolutely cannot” refusals based on its programming constraints.

The neuron fires on self-referential AI identity phrases (e.g. “As a large language model,” “AI,” “model,” etc.).

statements where the speaker issues a disclaimer identifying themselves as an AI and noting limitations or non-advisory status.

The neuron activates on the self‐referential “As a large language model” style disclaimer phrase.