Hypothetical: Google develops autonomous self-driving cars and uses them to drive around updating map and street view data. Let's suppose the cars have pretty good AI that not only deals with driving, but also handles deciding where to go. One of the goals the AI seeks is to avoid dangerous situations and minimize damage to the car.
Suppose some of these cars figure out that certain neighborhoods are "bad": people are more likely to shoot them there, or stop them and steal parts from them. The AI decides to avoid those neighborhoods, and uses its machine learning algorithms to learn to predict bad neighborhoods based on its prior experience with other neighborhoods.
Why a particular neighborhood is "bad" is usually a result of the interplay of many complex social and economic factors. Poverty correlates well with badness, for example.
But the car probably doesn't have poverty data. It has to rely on observables. One observable that correlates reasonably well with poverty is race. Note this is just a correlation, not causation. But the car doesn't care. It just cares about the correlation.
So the car starts avoiding neighborhoods where it sees a lot of black people!
It will be interesting to see how society deals with that. It raises lots of interesting ethical and legal issues.