Since some simple and free sites have been made available to the public to ask artificial intelligence software to produce images, illustrations and fake photographs – such as DALL•E, Midjourney and Stable Diffusion – it has become a bit more complex to understand if the person a photo of which you are looking at really exists or if it was generated by a computer. However, there are details of the human physiognomy that even the most sophisticated artificial intelligences seem to have a lot of trouble recreating: feet, teeth, but above all hands.
In almost all the photos generated by these programs, in fact, people have more or less fingers than they should, or horribly disfigured, crooked, disturbing hands. A perfect example is this set of photos obtained by developer Miles Zimmerman by asking Midjourney to generate “a 2018 photo of happy 20-somethings dressed for a night out having fun at a house party”:
Midjourney is getting crazy powerful—none of these are real photos, and none of the people in them exist. pic.twitter.com/XXV6RUrrAv
—Miles (@mileszim) January 13, 2023
The fact that AIs are bad at drawing hands has become so well established that it has inspired amused comments that, in the future, people who watch AI-generated porn they will ask their partners humans to add extra plastic fingers. Recently, children’s book author Justin Boldaji also wrote a very short horror story on Twitter: “I gradually realized that in old photos from when I was a child, all my relatives had horribly disfigured fingers.”
To understand the reason for these constant errors, journalist Pranav Dixit of buzz feed spoke to a spokesperson of Stability AI, one of the various sites that allow you to generate this kind of photographs. In order to draw things like “a Simpsons-style dog” or “a platypus wearing a bow tie,” Stability AI’s AIs, like those in DALL•E or Midjourney, were trained on huge datasets that contain thousands of images of dogs, scenes from the Simpsons, platypuses and bow ties (as well as millions of other things), until “learning” that the word “platypus” corresponds to a strange animal with a duck’s bill and a marmot’s tail.
However, the spokesperson for Stability AI explained a buzz feed that, basically, the photos and drawings of human beings contained in these data sets always contain faces, but much more rarely hands, feet and smiles showing teeth, and therefore it is more difficult to train the artificial intelligence to understand how many fingers a person’s hand should have, and how long or straight these fingers should be.
According to Amelia Winger-Bearskin, artist and professor at the University of Florida who studies the aesthetics of art generated by artificial intelligence, the problem is that these technologies have not yet understood what “a hand” is after all. , and how it relates anatomically to the human body. The AI “only sees the way the hands are represented, but the hands in the images are quite articulated. They are usually bent to catch something. Other times, they’re squeezed around another person.” In short, hands do too many things in photos to be recognized as a single element by artificial intelligence.
“In the photos, the hands are very rarely open, with the fingers well apart and visible. If so, artificial intelligence would be able to reproduce them perfectly,” says Winger-Bearskin. Not surprisingly, learning to draw a hand is also one of the first things you work on when approaching the artist’s profession, and it often happens that in cartoons the characters have fewer than five fingers to simplify the work to the designers. According to the expert, however, even artificial intelligences will learn over time to reproduce believable hands, exactly as real artists do: “if we want artificial intelligence to be a useful tool for humanity, it must understand what it means to be human, and what is the anatomical reality of human beings».