Far from Italy, a serious and important debate is underway on facial recognition software, on the limits of this technology and on the risk that the algorithm amplifies the racism that circulates. A striking example comes from a very beautiful project that takes us back two thousand years. It was carried out “during the quarantine”, a Canadian researcher, Daniel Voshart, who studies the applications of the so-called opposing generative networks (GAN) for machine learning in an original way. In short, he reconstructed photos of 54 Roman emperors.
Technically they are “portraits made with machine learning”. In fact, he took the last statues, when they were there, added the descriptions of historians about their morphological and even character details, inserted all these data into a software, Artbreeder, which allows you to create portraits using artificial intelligence, and the end result is actually intriguing. They seem alive.
They suddenly become so close that they seem like our neighbors. The faces of the 54 emperors photographed by artificial intelligence are the faces of Romans you meet at the bar, or on a bus, or at the market. They are the Romans – even when they came from far away, they were Romans – described by Pasolini or that we have seen in the films of Deaf and Verdone. There is nothing heroic or epic about these images, this was not the purpose, explained Voshart in his post, “I rather wanted them to look like they really were”.
So far little to add, apart from the fact that the poster of the 54 emperors can be purchased online in precious paper and in three different formats for a maximum of 72 dollars. The beauty comes when a mysterious Italian researcher, Davide Cocci, comes across the work of his Dutch colleague and discovers that something is wrong: some emperors in the photos of Voshart are blond. Seven. But were the emperors blond?
It is an old question, which brings us back to a debate that began during Nazism when a certain racist rhetoric, in search of the origins of the Third Reich, tried to pass off the alleged Aryan origins of some emperors. This thing obviously didn’t die there but survive in some online forums. In one in particular, there is a table that reconstructs the Roman emperors based on hair color, citing some authoritative historical sources, such as Pliny and Suetonius, and others less so. In particular, he quotes a work by a German historian, Wilhelm Sieglin, published in 1935 but written in 1905 and therefore long before the Nazi era, entitled “Die blonden Haare der indogermanischen Völker des Altertums”, or “Blond hair in the ancient Indo peoples” -European “.
In turn, the German historian refers in part to the work of many centuries before a minor Greek historian, considered unreliable on the point. In short, of the 46 emperors for whom there are sure descriptions of physics, writes the Italian researcher, only 4 were with light hair, but one dyed them (Commodus), while Nero and Augusto were brown.
This is not the place to deepen the historical question but rather to highlight the functioning and the limits of artificial intelligence that works, to make it simple, applying a series of algorithms to a database: the error, and often racism, can be among the former, in the way the algorithms are constructed, because they are rules that reflect our vision of the world; but also among the data that is fed to the software.
After the post by the Italian researcher, Voshart has removed the data of the Nazi forum and those of the Greek historian, and the hair of three emperors have returned dark.