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

This paper shows that a model’s “self-referential” tone—such as disclaimers like “I’m just an AI”—can be significantly changed by the chat template used in deployment, not just by the model weights themselves. For CIOs and technology leaders, that means LLM behavior is partly an application-layer and governance issue: IT teams need to treat prompt templates, system wrappers, and deployment conventions as controllable levers that can affect user trust, compliance posture, and evaluation results across different environments.