AI output is probabilistically the average opinion of everyone on the internet so it shares the common biases of the general public. Even with a bit of RLHF to “balance out” the models. Also it probably doesn’t help to anthropomorphise them. They don’t have opinions, they just autocomplete based on prior input