Feed
"As a Language Model...": Chat Template Switches LLM Self-Referential Voice and Activation Steering Reproduces It
Whether a chatbot says 'I'm just an AI' depends heavily on the chat template.
0.53
disclaimer rate
Disclaimer rate with chat template
Same weights disclaim far less when the template is removed
How it works
Self-reports drive debates
Statements like 'I'm just an AI' feed AI safety and introspection research, but nobody knows what triggers them.
Toggle template, find direction
Run 8 base/instruct pairs with and without chat template, judge outputs, then extract a disclaimer direction from activations.
Template acts like a switch
Template raises disclaimers and kills 'I feel' talk. Adding the direction without template restores template-level disclaimers.
From the paper