Enter the maze

Algorithm: Explain yourself!

by Paul Curzon, Queen Mary University of London

A hooded figure in front of swirling data: copyright www.istockphoto.com 50290594

Your life is being controlled by algorithms. If your family has car insurance then an algorithm will probably have authorised it. Want to open a bank account? An algorithm decides whether you can. Want to buy a house one day? An algorithm will decide if you should be given a loan. Algorithms are doing the counting in elections, deciding who wins. They are running twitter bots generating vast numbers of tweets to persuade you to vote one way or another - to control what you think and so how you make decisions. In some countries algorithms decide when a person should be released from prison. In the future, in Europe at least, the algorithms are going to have to do more than just make decisions, they are going to have to explain themselves too. That means computer scientists are going to have to come up with some creative algorithms to do it. Explaining a decision is much harder than making one.

Algorithms are taking over more and more of our lives, which is why it is just as important that they can justify those decisions. If we don't understand the algorithms then who knows if they are making the best decisions or using valid reasons. If you don't realise that pressure groups use bots, you may be fooled into believing you are joining a popular protest, when actually you are being manipulated by a powerful and shady bot-master. In future, algorithms will be making even more decisions - about what medicine you should be given when ill, whether you should be allowed into a nightclub based on the way you have dressed, whether you are a security risk, whether you should be given a job, ...

Racist programs

A problem is the algorithms that make the decisions rarely explain themselves. They just give answers. People assume that computers never make mistakes and are impartial, but neither is true. Take prison sentences, for example, it turned out that the algorithms making the decisions in the US were more likely to give african-americans long sentences than others for exactly the same offence. Programs can be racist! Your race or religion shouldn't lead to you being treated differently, but that could increasingly be secretly happening. Prison sentence computers just follow the rules they are programmed with. If the rules as written are biased, the decisions will be too.
The algorithm making the decisions gave african-americans longer prison sentences

Learning to be like that

It's worse than that though. The latest programs work things out for themselves. They still follow algorithms but those algorithms tell them how to learn rather than what to do directly. For example, if you want to automatically detect faces in your photo collection, you give one of these algorithms lots of pictures of faces labelled with who they are. With enough images to train on the algorithm works out for itself whose face is who, and it can then identify faces in new pictures it hasn't seen before. This is massively effective. It is the technology behind music recognition programs self-driving cars, and the program that beat the best human at Go recently, too. The problem is these programs just spot patterns. If it's hard to work out why an algorithm based on explicit rules makes decisions, it's even harder when it is just spotting patterns. The patterns it spots may not be the ones we think (see Employ the Best or Bust!).

Laws about rules

Now the European Parliament has passed new laws about programs that use personal data to make decisions about European citizens. They must in future be capable of explaining their decisions and so allow decisions to be challenged if they seem unfair. This is a challenge to computer scientists. It means they have to invent new programs that can extract arguments from their decision making processes, together with ones that can turn those arguments into explanations we can all follow. Both need the kind of algorithm that creativity researchers have been working on. More laws like this might seem to be bad for business, but in the software world it is great as it drives innovation. Companies who can invent algorithms that can explain themselves will soon have a big competitive advantage.