Recognize an image created by artificial intelligence: here are six tips you can really use

Okay, dad in the big coat wasn’t real.

Many people initially assumed that it was, mainly because it was so harmless that few thought anyone would fake it. But someone did it, and many fell for it. As AI grows, more and more of these photos will appear.

So, here is your guide to know if an image is generated by artificial intelligence or not.

text. Image-generating AI just don’t understand text. Sometimes they guess the letter with random luck, but mostly they create incoherent scribbles. If the image contains fully legible text, it is probably not an AI image. Take a look at this photo of the Duke of Sussex walking into McDonald’s. The AI ​​nails the logo, but when trying to make a different text, it gives out gibberish.

However, it is possible to add text later using Photoshop, so a good fake can find a way around this.

Hands. This used to be a big problem because the AI ​​seemed to have no idea what hands looked like, and specifically didn’t know how many fingers a hand had. This has pretty much been resolved and they now usually have five fingers. But AI is still struggling with its hands, as many images don’t quite understand how they interact with their surroundings.

You can usually see it if you look closely. Looking again at the photo of Prince Harry, his hand seems right at first, but you can actually see the crushed little finger while his index finger is supposed to be holding the bag, but it doesn’t look like it, and the bag hangs incorrectly as a result.

Background. As good as foreground images are, if there is background content, it often gets confused. Check out this statue of the Pope on a bike. You can tell that the railings don’t always line up, while many in the crowd have faces that don’t exist or are flawed.

Cars. In fact, it refers to everything that is made by human hands. AI does a good job with “organic” material – faces or clothes. He is less good at reproducing the tougher, man-made things. Look at the bike above. At first glance, it looks like a bicycle. But if you look closely, it doesn’t make sense anymore. Brake cables go in and out. The rods from the axles to the fenders look bent and bent, and most egregiously, the Pope’s legs appear to be straight, as if the pedals are going through the bike. Ultimately, AI cannot understand how a machine or other human-made object works. He just knows what he is trained on and what the next generation pixel should look like.

You see similar effects with glasses and other artificial objects. In the famous depiction of the Pope in a robe, his glasses are actually curved, with only a single string holding them up at the crotch. Basically, look for things that defy physics – a belt without a buckle, or a loose belt that still follows the contour of the item underneath.

This error can also be seen in the picture below, where Boris Johnson is wearing a Balenciaga down jacket. The AI ​​seems to have tried to create drawstrings for the hood, but didn’t tie them properly to the jacket itself, meaning they don’t really make any sense.

Eyes. If everything else looks wrong, it’s always worth getting your eyes checked. Like the pedals on a bicycle, AI doesn’t know that it’s normal for the eyes to look in the same direction. So if you look closely at Boris Johnson’s eyes in this picture, you’ll see that they don’t really match.

Connections, edges and contact. AIs can struggle with where things touch, blur, get confused or redefined. There is a very subtle example of this in Boris Johnson’s image, where the transition between his fur and jaw is too sharp and doesn’t convey the shadows that should exist there. In the image of the Pope, the upper left corner of his head has created an unusual white artifact.

However, the problem with all of the above is that AI companies are trying to fix these bugs. Over time, they will probably all disappear. Even now, a skilled user of image-generating artificial intelligence can prevent or resolve these issues after the image has been generated, although this is often time consuming.

But this guide is meant to help you catch more.