A new arXiv study finds chat templates shape whether open-source language models use disclaimer language (“I’m just an AI”) or experiential language (“I feel”). Eight instruct models up to 9 billion parameters were tested.
With chat templates present, disclaimer language increases and experiential language decreases. Removing templates reverses the pattern. Activation steering on an identified internal direction reproduces template effects.
Researchers conclude deployment conditions confound studies of model self-description. arXiv preprint — not yet peer-reviewed.