We have already seen something similar – bridging the distances – in science fiction movies. Thus, in Minority Report (2002), starring Tom Cruise and based on a short story by Philip K. Dick, there is talk of the establishment of the PreCrimen system, based on the use of three brothers with precognitive abilities capable of predicting a crime before of it being committed.
Finding a method to predict whether or not a criminal will reoffend is something that every society would want. In the United States they have been pioneers in finding it: since 1998, their courts have used a program called COMPAS. It is a sophisticated algorithm that is credited with predicting the likelihood that a criminal will re-offend in the future .
This software, marketed by the Northpointe company and which has been used to evaluate more than a million criminals, aims to predict the risk of a defendant committing a crime within two years after its evaluation taking into account 137 parameters different from the individual, including their criminal record.
However, handing over control of something as crucial as the future of a person’s life to machines can be dangerous, because, according to the results of a new analysis carried out by a team of researchers from Dartmouth College – a private university located in Hanover, New Hampshire (USA) – the algorithm fails as badly as a randomly chosen person with little or no knowledge of the crime scene.
Neither more precise nor more fair
Proponents of these systems argue that big data and advanced machine learning (or machine learning) make these analyzes more accurate and less biased than those of humans. However, we have shown that the widely used COMPAS risk assessment software is neither more accurate nor fairer than the predictions made by people with little or no experience in criminal justice ”, explain the authors of the study, published in the journal Science Advances .
“With an overall accuracy of 65% , these predictions are not as accurate as we might wish, especially when we look at it from the point of view of a defendant whose future is up in the air,” the researchers comment.
To reach their conclusions, they worked with a sample of a thousand convicts and carried out a comparison of the predictive capacity of the software in relation to that of real people – some five hundred people without any legal preparation and who underwent a online survey. They failed in a similar proportion (around that 65%) , and that humans had much less information than COMPAS, since they only evaluated seven parameters, instead of 137.
In addition, an investigation carried out in May 2016 by ProPublica – an independent and non-profit news agency based in Manhattan – indicated that, although the parameters used by COMPAS do not include the race of the subjects, other data that it collects and that are correlated with this characteristic lead to the algorithm being able to fall into a racist bias by which African-American individuals are harmed – it predicts that they will reoffend more than they will later -, while whites benefit – in this case, it predicts that they will reoffend less than they do. They came to this conclusion after analyzing the effectiveness of the software in more than 7,000 people arrested in Broward County, Florida between 2013 and 2014.
