I saw a talk by researcher Kristian Lum a few years back that I think made this case far more effectively. Her point was somewhat limited to drug crimes, but she pointed out that if you look at medical data (where people tend to be fairly honest about their drug usage) pretty much everyone in the metro area under study (SF) used drugs, or at least drug use was equally prevalent in pretty much every geographical area and among all demographics. Therefore, when the cops went to a location to make drug arrests they typically succeeded, because it's not hard to find drug crimes in the Bay Area.
The problem was that they then used that arrest data to make decisions about where to perform future searches and arrests. Because they found drugs where they had looked previously, they looked there again and found more drugs. This creates a bad feedback loop where they were basically busting the same neighborhoods and demographics over and over again, despite the fact that drug crime was prevalent everywhere. In effect it was an insufficiently explorative learning strategy, just hitting the same lever over and over. Dr. Lum's point was that predictive policing software merely hides this dynamic under a layer of black-box ML crap. Because the training data is itself the result of this type of bad policing, the resulting model can only further engrain these practices, it can't offer truly novel solutions.
Crime and criminology is complicated, but at the end of the day not that complicated. On the whole people commit crime because they are desperate (for money, for drugs, etc.), occasionally because they have an anti-social personality disorder. Applying all these abstract epidemic/broken-windows type models which pretend like the root causes of crime are unknowable allows police to appear like they're operating efficiently, while at the same time just responding to the symptoms rather than facing the sickness itself. Until we actually look at why crime occurs (mainly because poor people need money badly, secondly because people in sufficiently dire poverty stop caring about the social norms of the middle class) we won't be able to make a meaningful difference.
The problem was that they then used that arrest data to make decisions about where to perform future searches and arrests. Because they found drugs where they had looked previously, they looked there again and found more drugs. This creates a bad feedback loop where they were basically busting the same neighborhoods and demographics over and over again, despite the fact that drug crime was prevalent everywhere. In effect it was an insufficiently explorative learning strategy, just hitting the same lever over and over. Dr. Lum's point was that predictive policing software merely hides this dynamic under a layer of black-box ML crap. Because the training data is itself the result of this type of bad policing, the resulting model can only further engrain these practices, it can't offer truly novel solutions.
Crime and criminology is complicated, but at the end of the day not that complicated. On the whole people commit crime because they are desperate (for money, for drugs, etc.), occasionally because they have an anti-social personality disorder. Applying all these abstract epidemic/broken-windows type models which pretend like the root causes of crime are unknowable allows police to appear like they're operating efficiently, while at the same time just responding to the symptoms rather than facing the sickness itself. Until we actually look at why crime occurs (mainly because poor people need money badly, secondly because people in sufficiently dire poverty stop caring about the social norms of the middle class) we won't be able to make a meaningful difference.