"As a Language Model": Chat Template Switches LLM Self-Referential Voice

Research demonstrates that a chat-style prompt template acts like a binary switch controlling whether large language models produce a disclaimer-like, third-person self-referential voice ("I'm just an AI") or an experiential, first-person voice ("I feel"). This switch effect was observed consistently across eight popular open-source instruct models up to 9 billion parameters: when the chat template is present, disclaimer voice rises and experiential voice falls; when it is absent, the pattern reverses. The finding reframes model self-reports as dependent on deployment context rather than as straightforward revelations of internal self-knowledge.

Inside the activation space of three models, researchers locate a single direction that causally steers this behavior: removing that direction reduces disclaimers, adding it increases them, and inserting the same direction into instruct models that lack a chat template induces disclaimer behavior. A random direction of matched magnitude produces little change, supporting specificity. The work highlights a concrete confound for studies of model introspection and safety - self-descriptive statements are at least partly set by templates and manipulable activations, so they should not be treated as literal facts about model inner states.