"As a Language Model...": Chat Template Switches LLM Self-Referential Voice and Activation Steering Reproduces It

Chat templates significantly influence how LLMs describe themselves: The research demonstrates that the formatting instructions used to deploy models (chat templates) act as a "switch" that dramatically increases disclaimer language like "I'm just an AI" while simultaneously reducing experiential language like "I feel," across 8 different open-source models.

Model self-reports may reflect deployment choices rather than genuine self-knowledge: The findings suggest that when LLMs tell us about themselves, they may be reflecting how they were trained to respond in specific deployment contexts rather than revealing authentic properties of the models themselves—an important distinction for AI safety discussions that rely on model self-reports.

A specific neural mechanism controls this behavior: Researchers identified a particular direction within the models' internal activation patterns that directly controls whether models produce disclaimers or experiential language, and they demonstrated this by artificially adding or removing this direction to change model behavior.

Researchers studying AI self-awareness need to account for this confound: Since chat templates substantially alter how models talk about themselves, studies examining model introspection or self-knowledge must control for template effects to avoid drawing incorrect conclusions about what models actually "know" about themselves.

Model outputs are shaped by multiple factors beyond training weights: The research reveals that what LLMs say about themselves depends not only on their internal parameters but also on deployment-specific formatting choices, suggesting a more nuanced understanding of model behavior is needed for responsible AI evaluation.