Earlier this month, a research paper from Nvidia, maker of graphics processing units, showed that the company is able to generate photo-realistic images of people who never existed. It marked another step toward a world in which any media can be easily and believably generated from scratch, a looming crisis for truth online.
Earlier this month, a research paper from Nvidia $NVDA, maker of graphics processing units, showed that the company is able to generate photo-realistic images of people who never existed. It marked another step toward a world in which any media can be easily and believably generated from scratch, a looming crisis for truth online.
Fortunately, AI-generated faces bear some telltale signs. This week, computational artist Kyle McDonald published a guide on how to identify a fake. These tips probably won’t be reliable forever, and they’re certainly not applicable to every picture—some generated images are extraordinarily convincing. But every little bit of information helps.
Join 500,000+ readers who start their day with Quartz.
By subscribing, you agree to our Terms of Service and Privacy Policy.
The following pictures are screenshots grabbed from Nvidia’s latest paper, which Quartz previously covered by imagining how these AI people might interact on Tinder.
The picture above has a bunch of clues. One of the easiest is the big weird blob floating off the top right of the person’s head. That kind of artifact or aberration is common in AI-generated imagery, consistent with Google $GOOGL’s DeepDream experiments a few years ago.
But the image is also slightly asymmetrical, when you look around the person’s ears. Not only is the hair fuzzy and strange on the right side, but the ear is missing an earring that’s present on the left side. (Or you could say that the left ear has an earring not present on the right side, as an exercise in perspective.)
Algorithms have no common sense, and no rule that if an earring is present on one side it has to be present on the other, except maybe when generating faces of boy bands from the late 90s or people who wanted to be in boy bands in the late 90s. Because of this, the algorithms sometimes don’t generate facial features or accessories as they should.
Algorithms also don’t know the number or orientation of teeth, instead drawing on examples of teeth from all different angles. Sometimes the algorithm messes that up, leaving its creations with odd and mangled teeth.
This might be a little more difficult to spot than a missing or disfigured earring, but if you look closely at the above image, the pictured woman has a small, third middle tooth. There’s also some weirdness around the ears, so it’s definitely a fake.
Patterns are sometimes a mystery for people-generating algorithms, which can lead to strange structures in the background of images, or strange clothing on the generated subject. McDonald says to especially watch out for text in the background of generated images, since it’s almost always malformed.
While the subject of the image above looks stunningly like a human, it appears to be in some sort of geometric jade prison, or perhaps Superman’s Fortress of Solitude.
Below, you can see this manifest in clothing. It just doesn’t make any sense. There’s also a strange artifact of an earring in the photo, and some of the subject’s hair isn’t actually attached to anything, which is also common, McDonald says.
Stay vigilant.
