It is very easy to design experiments that find "bias" when the null hypothesis is that a language model gives consistently dead-even answers to an arbitrary political compass quiz regardless of phrasing and sampling

Comments