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

The study examined eight open-source instruct models up to 9 B parameters. It found that the chat template acts as a switch, amplifying disclaimer statements (“I’m just an AI”) and suppressing experiential language (“I feel”). Activation analysis identified a specific hidden-state direction that controls this behavior. Removing or adding this direction toggles the disclaimer voice. The results show that a model’s self-reports are partially governed by the chat template, not solely by its weights, and researchers must account for this confound when evaluating self-knowledge.