How to Recognize a Photo Created by Artificial Intelligence
When I first explained to my son as a preschooler that cartoon characters are not real, but drawn and animated on a computer, it amazed him. How could something that looked so realistic be fake? I think now I know how he felt; lately, the more I learn about AI photographs, the more my brain wants to rebel against this whole idea.
You may have heard of the This Person Does Not Exist website, which features realistic ( but completely fake ) photographs of people who certainly exist in this world, but who do not. AI (artificial intelligence) photographs created with computer algorithms have become so advanced that it is becoming increasingly difficult to distinguish real from fake.
Certainly some stellar special effects in movies and video games can do it. And we are normalizing the practice of creating fake people for other entertainment purposes with sites like Have They Faked Me, which entice us to look for our fake doppelgangers.
But in the wrong hands, these algorithms can also be useful in producing propaganda and other fake media. In our time full of fake news, we have had to learn not to always believe everything we read; now we too must not believe everything we see.
Created by Jevin West and Karl Bergstrom , What Face Is Real website aims to educate users to less trust in photographic images and be more critical of potentially synthetic “photographs”:
Computers are good, but your imaging systems are even better. If you know what to look for, you can spot these fakes at a glance – at least for a while. The hardware and software used to create them will continue to improve, and it may only be a few years before humans fall behind in the arms race between counterfeiting and detection.
There are a few more clear signs that the photo is fake . Here’s what West and Bergstrom say we should look before we assume the photograph is real:
- Water stains . Shiny droplets that look like water spots are the hallmark of NVIDIA’s current “StyleGAN Algorithm”.
- Strange backgrounds . If the background looks like a torn photograph, has strange shapes, or other distortion, it could be fake.
- Asymmetry . Algorithms have all sorts of symmetrical problems, including facial hair, earrings, fabric, and glasses in particular.
- Hair . The fake person’s hair may look too straight, streaked, or have a halo or glow around it.
- Fluorescent bleeding . Look for vibrant colors that flow from the background to the hair or face of the person in the photo.
- Teeth . This is one of the first things I look at. It’s not just that the teeth look imperfect, but that they are odd (for example, three incisors), asymmetrical and in some photographs have an almost pixelated appearance.
At this point, West and Bergstrom say in What a Real Face is that there is still a silver bullet that you can use to determine if a face on the Internet belongs to a real person:
We are currently not aware of any software that can generate multiple images of the same fake person from different perspectives. Therefore, if you want to be sure that your tinder is a real person, insist on viewing two or more photos. If he has a headshot and a photo of himself with a tiger, he is a real person. Of course, the person texting you might be someone other than the person in the picture, but the person in the picture is definitely real.
Ready to see how well you can tell a real person from a fake person? You can take their test here .