For example, a couple years ago, there was a big hubbub over a Google Image labeler that labeled a black man and woman as "gorillas". A mistake for sure, but the headlines about the algorithm being "racist" were wrong. The algorithm was certainly incorrect, and it could probably have been argued that one reason it was wrong is that its training set contained fewer black people than white people, but the algorithm was certainly unaware of the historical context around this being a racist description.
Similarly, in the early days of Google driving directions I remember one commenter saying something along the lines of "You can tell that no black engineers work at Google" because it pronounced "Malcolm X Boulevard" as "Malcolm 10 Boulevard". Of course, the vast majority of time you see a lone "X" in a street address it is pronounced "ten".
It's kind of analogous to the "uncanny valley" problem in graphics. When the algorithm gets things mostly right, people think of it as "human-like", and so when it makes a mistake, people attribute human logic to it (it's quite safe to assume that a human labeling a picture of black people as gorillas is racist), as opposed to the plain statistical inferences ML models make.