OK, the Pope in the big coat wasn’t real.

A lot of people initially assumed it was – largely because it was so harmless that few would have thought anyone would bother to fake it. But someone did, and a lot of people fell for it. As AI ramps up, more and more photos like this will appear.

So here’s your guide to working out if an image is AI-generated or not.

Text. The image-generating AI’s simply do not understand text. Occasionally, they’ll work out a letter through random luck, but more often than not they’ll produce incoherent scribbles. If there’s any fully readable text in an image, there’s a good chance it’s not an AI image. Take a look at this image of the Duke of Sussex picking up some McDonald’s. The AI nails the logo, but comes up with gibberish when it tries to do any other text.

Prince Harry isn’t picking up McDonald’s – this image is entirely AI-generated.

However – it’s possible to add the text in with Photoshop afterwards, so a good fake might find a way around that.

Hands. This used to be the big one because the AI’s seemingly had no idea what hands looked like, and specifically, didn’t know how many fingers hands have. That’s mostly been fixed, and they usually have five fingers now. But the AI still struggles with hands, with a lot of images not quite nailing how they interact with whatever is around them.

This is usually evident when you look closely. Again looking at the Prince Harry image, his hand initially appears correct, but you can actually see a squished little finger, while his index finger should be holding the bag but doesn’t actually appear to be, and the bag isn’t hanging correctly as a result.

Background. As good as the foreground images often are, if there is background content, it’s often confused. Take a look at this image of the Pope on a bike. You can see the railings don’t always match up, while a lot of the crowd have faces that are non-existent or glitchy.

The Pope wasn’t actually riding a bike, just as he wasn’t wearing a big coat.

Machinery. This refers to essentially anything made by humans. AI is pretty good with “organic” material – faces or clothing. It’s less good at replicating more rigid, human-designed things. Take a look at the bike above. At first glance, it seems like a bike. But as soon as you look closer, it stops making sense. The brake wires fade in and out. The bars from the axles to the mudguards look bent and curvy, and most egregiously, the Pope’s feet seem to be at the same level, as though the pedals continue straight through the bike. AI, ultimately, can’t understand how a machine, or any another human-created object works. It only knows what it’s been trained on, and therefore what the next pixel in a generation should look like.

You see similar effects around glasses and other human-created items. On the famous image of the Pope wearing a coat, his glasses are actually bent, and his cross only has a single string holding it up. Essentially, look for things that defy physics – a belt without a buckle, or a loose strap that nevertheless traces the contours of the item below it.

This flaw is also present in the below image of Boris Johnson wearing a Balenciaga puffer jacket. The AI appears to have tried to create some drawstrings for the hood, but hasn’t correctly connected them to the jacket itself, meaning they don’t actually make sense.

Boris Johnson, as far as we know, has not yet modelled for Balenciaga

Eyes. If there’s nothing else looking wrong, eyes are always worth checking. Much like the pedals of a bike, the AI doesn’t know that eyes should typically look in the same direction. So if you look closely into the eyes of Boris Johnson in this image, you can see they don’t actually line up.

Joins, edges and contact. The AI’s can struggle with where things touch each other, blurring, confusing, or overly defining them. There’s a very subtle instance of this on the Boris Johnson image, where the transition between his coat and jaw is over-defined, and doesn’t capture the shadows that should exist there. On the image of the Pope, the very top left of his head has created an unusual white artefact.

The problem with all of the above, however, is that AI companies are trying to fix these errors. Given time, all of them are likely to disappear. Even now, a skilled user of image-generating AI can prevent these issues occurring, or fix them after the image has been generated, though it’s often time-consuming.

But this guide should help you catch more.

By admin