Pictures show the reality. Even in times of social media filters and Photoshop, we tend to see photos basically as proof of reality. Especially when they show people. Yes, the details may not all be exactly right, but the people in the photo are guaranteed.
But what if not?
Two algorithms play ping-pong
Self-learning algorithms are getting better and better at creating photos with artificial faces. These are pictures of people who do not exist in real life. The AI creates it based on a database of other (real) photos.
With the information from other faces, intelligent algorithms learn to create even images that look like real people.
The KI method that underlies this is called GAN (Generative Adversarial Networks). Simply explained, two algorithms communicate with each other, with the goal of training each other.
On the one hand, an algorithm generates the artificial photos. On the other hand, the other algorithm is supposed to detect fake photos. The longer this ping-pong game runs, the better both algorithms get.
In the early days of technology, it was still relatively easy for humans to recognize artificial AI images.
But now the KIs are so good that it is getting harder and harder.
Because this technology opens the door to lies, fraud, and manipulation in the wrong hands, experts are now giving tips on how to differentiate between real and artificial faces.
Notice “bullshit” faster
One of them is Kyle McDonald, an artist who works with code for his works . He has, for example, recently a very detailed overview compiled by which you can distinguish authentic from counterfeit KI images.
Scientists at the University of Washington in Seattle even go one step further. You’ve dedicated an entire site titled “Which Face Is Real” to the topic.
They want to achieve one thing above all else: users should be able to recognize “bullshit” faster. This gives visitors information about GAN technology and what weaknesses it still has. There is even a game where you can test your own knowledge.
Real or fake? How to recognize AI pictures!
So how exactly can you tell the real ones from the wrong photos?
So far, there are some key bugs that seem to undermine the algorithms when creating an artificial photo.
1. Funny hair
Human hair, in its combination of much information and a lot of detail, seems to be sometimes difficult for the algorithms. Often smooth hair looks like an oil painting or the hairstyle has incoherent swirls. Not infrequently, hair is also placed in the wrong place.
Although the AI can produce eyes, ears, eyebrows or earrings, but it often happens with the symmetry. Sometimes the eyes have different colors. Other times earrings are not the same length. Paying attention to strange symmetry errors helps to expose fake pictures. Especially glasses are apparently still very difficult to represent for the algorithms.
3. Strange teeth
Teeth, with their near-symmetry, are obviously also a challenge. Often, the algorithms overlap the teeth or they pick a combination of teeth that do not match.
4. Blurred background
Since the algorithms focus all their skills on their faces, other aspects are often displayed very poorly in many AI images.
Anything that can be seen in the background is usually too much information for the algorithms that they can not coherently merge. It looks similar to lettering, which therefore often appear only as blurred letters.
Another indication of an artificial photo is weird color palettes. Again, the algorithms do not combine the information very well, often resulting in color patterns that are more reminiscent of watercolors and less of true colors.
The ultimate trick
In addition, there is an ultimate trick to expose AI images. If you are not sure if the photo is real, ask for a second photo of the same person.
So far, no algorithm has been able to create two images that show the same person in different perspectives or situations.
So if someone sends you a portrait photo and a photo of you at the shooting festival, at least the person in the photo is real.
Of course, the AI will eventually get that done too. But until then we can see the differences by looking closely.