Just a thought: avoiding political bias could it's own form of bias. I would think the goal is mostly to get "correct" answers based on facts, logic, and so on. If a "correct" answer can be computed (what does that mean with an LLM?), then it might be biased, but still desirable.
This sounds like category error to me. LLMs output string completions. Humans should judge the set of outputs to be representative of what humans would write. All humans, not biased to the subset of liberal humans.
The bias is thinking the liberal ideas are correct ideas.