If you look at the demographics of people convicted after being flagged by a flawed algorithm, you'll find a lot of correlelations. Pick a totally spurious indicator. Whether you live in a house with an odd or even number. If the algorithm decides odd numbers are highly indicative of fraud, and those are 90% of cases that get investigated, then you're going to have 90% of convictions happen to have odd house numbers. "See, look, the algorithm works!"