All models must be biased, there must be value judgments in curating things into it.

These models can be steered so easily with a shockingly small number of training examples, that it's unthinkable the models wouldn't all have some kind of (perhaps unintentional) bias.