It's becoming a common occurrence to fall victim to false positives from algorithmic filters. False positives are just always going to be part of any kind of machine-learning algo. In the end it's statistical probability, and probability doesn't equal certainty. Sure there are techniques to reduce false positives, but I sincerely think that any user-affecting policy based on ML / algos should always have a real-live human based support fallback to get real resolution to the real issue of false positives.