The first is one relates to this application of facial recognition targeting poor people. This woman had no documentation on her. I'd wager this is really common in Brazil. How could she possibly prove she is _NOT_ the person the software with 99.999..% accuracy says she is?
If the actual criminal wasn't already in jail, would she be set free in the same day she was arrested?
The second issue is perhaps more universal to other automations, but certainly potentializes the first one, dealing with _scale_. Human cops can't check every person they see against the entire database of wanted people.
25% of false-positives from every cop is a hell of a lot less than 0.00001% of an automated system when applied to every crowd, everywhere, all the time.