Famous psychologists and neuro-scientists claim that a human being has two "cognition systems": S1 which is a probabilistic engine working on associations learnt from past experience, and S2 which is the true intelligence, the ability to formally think of pre-requisites, logical connections, and conclusions. It is funny that what we call "artificial intelligence" is purely S1, while when we talk about intelligence of a human being we normally mean the ability of activating and using S2.
A huge problem with different profiling and prediction techniques is that they conserve the status quo. It is a fact of the world that some people are more criminal than the other. However, we all understand that the reason for this fact is not that some people are inferior, but rather that they are put in a situation where they lack skills or possibilities to earn without crime. A logical policy would reflect this, a probabilistic automaton can just say "beer drinking + living in this area + low education = criminal". Excluding some data would not help, as the system will automatically find proxy variables (e.g. it may automatically switch from registration address, which is illegal to use in many countries, to the geography of Uber drives or Facebook check-ins, which is a piece of info one can easily buy)
As the jinn is out of the bottle, banning algorithms is not a way to go. And I don't think that anyone has a ready solution for this problem.
thanks for the comment, hard to add anything to it. I do hope that the book later on will elaborate on the range of approaches to tacking the most obvious drawbacks. However, as some data is simply physically unobtainable, we'll still have to use proxies in modelling, so it's back to square one.
Famous psychologists and neuro-scientists claim that a human being has two "cognition systems": S1 which is a probabilistic engine working on associations learnt from past experience, and S2 which is the true intelligence, the ability to formally think of pre-requisites, logical connections, and conclusions. It is funny that what we call "artificial intelligence" is purely S1, while when we talk about intelligence of a human being we normally mean the ability of activating and using S2.
A huge problem with different profiling and prediction techniques is that they conserve the status quo. It is a fact of the world that some people are more criminal than the other. However, we all understand that the reason for this fact is not that some people are inferior, but rather that they are put in a situation where they lack skills or possibilities to earn without crime. A logical policy would reflect this, a probabilistic automaton can just say "beer drinking + living in this area + low education = criminal". Excluding some data would not help, as the system will automatically find proxy variables (e.g. it may automatically switch from registration address, which is illegal to use in many countries, to the geography of Uber drives or Facebook check-ins, which is a piece of info one can easily buy)
As the jinn is out of the bottle, banning algorithms is not a way to go. And I don't think that anyone has a ready solution for this problem.
thanks for the comment, hard to add anything to it. I do hope that the book later on will elaborate on the range of approaches to tacking the most obvious drawbacks. However, as some data is simply physically unobtainable, we'll still have to use proxies in modelling, so it's back to square one.