Models don't have preferences, they have context

A language model's output reflects training data and context, not an opinion. Asking an empty chat yields React because it dominates the corpus. Adding project files like Svelte changes the answer. The model's "preference" disappears when context shifts, as Gao and Kreiss documented.