DYNA JOURNAL ENGINEERING

This research work, led by Michal Kosinsky, "analysed" different profiles and boasted of correcting the sexual orientation of the candidates with a great rate of success.

Years later, the same author returned with another article entitled Facial recognition technology can expose political orientation from naturalistic facial images; and nothing more and nothing less, it was published on Nature.
The article quickly raised controversy and provoked an avalanche of critics from other researchers who claimed that Kosinsky had developed an algorithm worthy of a 13-year-old teenager, and that he claimed the return of an idea classified as a pseudoscience long ago: phrenology.

Phrenology consists in the personality evaluation through the measures between face parts (eyes, mouth, nose, ears…). This work became a clear example of monkey do, monkey see. Phrenology was ambitioned since ancient times, but its golden age was the Victorian Era, in which the ugliness or beauty of an individual was the main criteria for local authorities to accuse someone of any unsolved violent crime. Perhaps phrenology was promoted by other sciences that revolutionized the police investigation of the time, among them the beginning of the use of ether as a general anesthetic (1844), the beginning of the use of fingerprints in the identification of people (1879), the discovery of X-rays (1895)...

Going back to our time, what did Mr. Kosinsky's exactly do? He scraped photos of about 75,000 profiles of men and women from online dating websites. Next, a deep neural network for pattern identification was trained with those images. As Evgeny Morozov warns in his work The Madness of Technological Solutionism, with the right algorithm and precise learning data, we can all be classified into any group. And that's what happened. The researchers concluded that their tool was valid to distinguish heterosexual people from those who are not.

In experiments in which volunteers were asked to distinguish between two photos which of them is more likely to be homosexual, they got 54% correct for women and 61% for men. While 'the tool' of the researchers got 71% and 81% respectively. However, in subsequent tests, with photos taken from Facebook, the algorithm's success rate dropped sharply. Moreover, any serious statistician would alert from the highly biased dataset of Kosinsky, which only obtained young white people. In this text, gaydar tool is just presented, but Kosinsky idea to distinguish political ideologies out of photos works similarly.

One of the biggest risks of artificial intelligence is that it perpetuates the mistakes and prejudices of the past, camouflaging them under a sense of objectivity. And it is this type of study that many companies wield to continue installing and developing facial recognition systems for the (wrong) detection of criminals, immigrants, violating a multitude of human rights under the aura of bad science. Reviews of over a thousand research articles in Psychology reveal that you cannot guess anyone's mood or thinking from their movements and facial features. Perhaps it draws more attention to talk about the technological unemployment or the advent of Artificial General Intelligence, but silently, more risky artificial intelligence tools are being developed. History explains that civil rights are hard to reconquer.