> Bias is when ML sees something not relevant as a pattern and uses it as a feature to determine the future.

This alone does not fit the definition of statistical bias. Statistics is not capable of proving causation. So a completely trivial fact or feature can be a great estimator nonetheless. Of course if the system changes then you have a problem (with or without a causal model).

If historic data is warped by police of course, then you're using a biased sample and will indeed get a biased estimate that way. But if you correctly use randomly sampled data and a method that has been mathematically-proven to be unbiased, then if your model says to deny loans based on race, you would likely make fewer bad loans more by following it versus ignoring this information.