The bias has to be small enough that if you ask an LLM to repeat some passage of text like the national anthem, either from the training data or from the prompt it doesn't change random words.

Gotta be hard to tune that.