A recent study has demonstrated that software meant to predict crimes is, well, terrible at predicting crimes. No one involved in the program apparently ever watched Minority Report. Jokes aside, this was entirely predictable. Any form of what we rather lazily call artificial intelligence exacerbates the general rule of programming: garbage in results in garbage out.
Garbage in, garbage out (or GIGO because programmers are apparently incapable of speaking in anything other than acronyms and bad puns. I have been working for my current firm for the better part of more than twenty years and there are still acronyms that no one has ever explained to me. I suspect no one actually know that they stand for.) is one of the first things you learn as a programmer, and it is the probably the most important. Computers are not magic, they are just machines. They will not solve your problem for you — they will merely do what you tell them to do in incredibly efficient fashion. Sometimes this tendency is good. Sometimes, it is hilariously or even tragically wrong. What we call artificial intelligence has this problem on steroids.
Artificial intelligence relies on training data — on being fed material that serves as the basis for the math to “figure out” what to do next. There are various different means of “figuring out” that the math can take, but in the end, they all rely to one degree or another on having a robust set of data that can provide a realistic picture of the problem domain the math is mean to figure out. Policing is not such a domain.
Relying on police data to create predictions, as these systems do, is not going to give you a reasonable picture of crime. Police departments are not good at solving crimes, for a variety of reasons. They tend to over police some areas and under police others. They tend to treat some people differently based on class and, especially, race. Crime reporting is not tracked well and tends to be underreported by certain populations. Basing decisions on police data is going to result in systems that don’t actually understand the realities of crime and thus cannot predict the most likely areas where crime will happen. Garbage in — crime and police reports — results in garbage out – a system that is hilariously bad at predicting crimes.
These systems allow data to replace expertise and as a result generate a system that is at best useless and at worst harmful. Harmful because it could direct police resources away from actual crime hotspots. And harmful because it could encourage the police to have hostile interactions with populations that are otherwise underserved or suspicious of the police, driving down trust and making policing even less effective than it already is.
Ironically, there are proven ways to use machine learning to assist in crime prevention. But that require centering expertise over data. In order to be effective, these systems require an understanding of how the limitations of existing crime data come to be and how they can be overcome by other sources of information. This is hard work, requiring people to spend time and effort to understand the nuances of the problem and to think creatively about potential solutions. It works, but is not the simple, cheap replacement of people with data that these companies promise. It costs, in time and people and treasure, in other words, and cannot be made to have a huge Silicon Valley style return on investment. Non-garbage generally costs more than garbage. But it also produces much better results.
One of the problems with the so-called artificial intelligence gold rush is the dash to replace expertise with data. The failure of police prediction systems is one example of the way these incentives create failures at the public policy level. By promising that you can substitute data for expertise, you can promise a cheap, fast solution for an intractable problem and make yourself quite a bit of money in the process.
Machine learning systems really can help augment expertise to solve hard problems. But garbage in garbage out is as close to a physical law as exists in programming. Data cannot substitute for expertise, at least not in most cases. Until we really learn that lesson, we are going to keep paying for failure instead of leveraging technology to assist expertise.
Leave a Reply