AI supposedly learns to identify criminals by their faces, takes us back to the 19th century

The system wouldn't have worked on Charlie here because he's got facial hair

Measuring the worth or character of a person by their physical traits is generally thought to be a sad staple of the quasi-scientific 19th and 20th centuries. But some of these ideas seem to be making a comeback, apparently dressed in full scientific garb, and accessorizing in trendy buzzphrases like machine-learning and artificial intelligence. That’s the case with a new “scientific” paper, currently waiting to be published, which reports that researchers in China managed to teach a machine to identify criminals just by looking at their faces.

The idea that criminals inherently have some sort of noticeable physical trait that sets them apart from “regular” people, was debunked more than a hundred years ago, by using statistical data. However, it recently came back when a psychological study from Cornell University supposedly found that humans were good at actually recognizing criminals when they see them. Here’s a fun, related fact: a meta study from 2015 suggested more than half of major psychological studies can’t be reproduced. A secondary fun fact is that researchers then found that the debunking study had its own replication errors. Make of that what you will.

Now, in an effort to take things out of human hands, or brains in this case, Chinese researchers trained a machine-learning algorithm to analyze the faces of almost 2000 humans without facial hair, half of which were criminals. They used 90% of the samples to train the system, while letting the AI itself analyze the remaining 10% and say whether it was looking at a law-abiding citizen or a criminal.

The researchers claim that the machine got it right 89.5 percent of the time, identifying criminals with a remarkably high degree accuracy. The system reportedly identified that criminals had a larger curvature of the lip by 23% than regular people, a 6% smaller distance between two inner corners of the eyes, and a more acute angle by 20% between two lines drawn from the tip of the nose to the corners of the mouth.

The end result is that criminals reportedly have bigger differences between them and the regular population than law-abiding citizens do. In other words, the more people look like each other, the less likely they are to be criminals.

The “study”, which took place on a population of almost identical ethnical background, raises serious questions and concerns for the future. While criminality may have a genetic component, most researchers believe it has a lot more to do with environmental socio-economic factors, rather than anything genetic.

If tentative, highly dubious studies such as these take root in the public psyche though, it may lead to a dangerous future where “justice” is done a priori, based not on reality, but on the semi-random bits that a machine puts out.

Source: arXiv Via: MIT Tech Review

Report a problem with article
1479817994_blackfriday2016
Next Article

Save 90% off The Immersive Angular 2 Bundle; further 15% off site-wide at Neowin Deals

lumia-950-xl
Previous Article

Microsoft cuts Lumia 950 XL price to £289 in the UK, including a free Display Dock

23 Comments - Add comment

Advertisement