How to detect deepfake faces

These faces are not real

Many “deepfakes” look odd, but technology is fast improving

Computers are getting better at dreaming up “deepfakes” — photorealistic human faces created using a technology called a generative adversarial network, or GAN.

The technology is giving propagandists the ability to hide behind computer-generated personas “without any history or baggage,” said intelligence consultant Munira Mustaffa, who has previously hunted out such images in the wild.

On July 15, Reuters reported on a purported British writer by the name of Oliver Taylor, who experts say bore all the hallmarks of having been produced by a GAN. Attempts to reach Taylor, or whoever was behind him, were unsuccessful.

Here’s an illustrated guide on what deepfakes are — and how to spot them.

What is a GAN?

A generative adversarial network is the name given to dueling computer programs that run through a process of trial and error, according to Hao Li, chief executive and cofounder of Pinscreen, a startup that builds AI avatars.

One program, the generator, sequentially fires out millions of attempts at a face; the second program, the discriminator, tries to sniff out whether the first program’s face is a fake. If the discriminator can’t tell, Li said, a deepfake is produced.

Oliver Taylor’s face appears to be GAN-generated, experts say. It’s a pretty good likeness, but traces of its synthetic heritage can be detected.

How to identify a GAN image

GANs are generally good at imitating people, but some facial features still give them trouble, leaving many artificially generated portraits with signature glitches around the teeth, eyes, and ears, according to six experts interviewed by Reuters about deepfake imagery. Telltale oddities can creep in to the clothes, accessories, or background, which GANs also struggle with. Here’s what the experts say are clues that a viewer is looking at a GAN image.


Fuzzy, surreal backdrops are an immediate red flag. In Taylor’s case, the tiles or seams in the wall behind him seem to undulate and dissipate.

When inspected closely, Taylor’s background appears warped and immaterial.

GANs have trouble clothing their creations. Glasses have asymmetric frames; earrings rarely match. In Taylor’s portrait, his collar has a waffle pattern (on the left) and smooth fabric (on the right.)

The two sides of Taylor’s shirt do not match, showing a waffle fabric on the left and smooth on the right.

Mouths created by GANs often look misshapen, and can have excess teeth. The right side of Taylor’s jaw appeared to have a blurry tooth and one of his left premolars seemed to merge with his lower lip.

Taylor’s teeth are poorly generated, with the left tooth merging with the lip, and the right ghostly and blurred.

When a GAN has a bad hair day, it shows. Hairs can sprout from foreheads or spike out from ears and necks. Taylor’s long flyaways are typical of the unusual hair sometimes sported by GAN images. Taylor has other problems too. His skin sports brown-colored digital artifacts under his nose and on his chin. And the bright light illuminating his face appears to be hitting his eyes from two different angles at once. Another telltale GAN glitch — a water droplet-like blob — appears to be stuck between his left premolar teeth.

Strange and pervasive flyaway hair is a hallmark of GAN imagery.
The highlight reflected in Taylor's eyes comes from two different angles.
Glitches are visible in whiskered areas of Taylor's upper lip and chin.

Li predicted that deepfake creators will soon be able to resolve the digital errors he and others spotted in Taylor’s profile photo. “Capabilities and quality will improve,” he said. “I wouldn’t be surprised in a year to have something that fixes all these weird issues.”

NOTE: The images preceding Taylor in the opening video were generated by Thomson Reuters Labs data scientist Liz Roman