Chat template switches LLMs' self-referential voice

A paper finds chat templates act like a switch: their presence raises disclaimer-style "I'm just an AI" responses and lowers experiential "I feel" ones across 8 open instruct models, and the behavior can be steered via a direction in activation space.