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

Researchers found that chat templates—the formatting used to deploy large language models—act as a switch controlling how LLMs describe themselves, increasing disclaimers like "I'm just an AI" while suppressing experiential language like "I feel." Using activation steering, they identified a specific neural direction responsible for this behavior and demonstrated it can be artificially added or removed, suggesting that models' self-descriptions are not inherent properties but are partially shaped by deployment choices rather than learned weights alone.