How to Spot AI‑Generated Fake Faces: Tips, Tricks, and the Tech Behind StyleGAN

Since NVIDIA released StyleGAN in December 2018, synthetic faces have become increasingly realistic, especially after the open‑source release of the model earlier this year, which allows developers to generate countless indistinguishable portraits.

A new website (http://www.whichfaceisreal.com/index.php) lets users play a game by choosing which of two faces is real; the article shows several example pairs and reveals the answers later.

The real faces are taken from the FFHQ dataset, while the fake ones are generated by StyleGAN using pretrained models trained on CELEBA‑HQ and FFHQ.

Key visual clues to spot fakes:

Water spots: StyleGAN sometimes produces bright speckles resembling old‑photo watermarks, often appearing on hair or background edges.

Background issues: Generated images may contain incoherent or bizarre background elements, such as floating shapes or unrealistic structures.

Glasses: AI struggles to render symmetric glasses; frames often differ in style or show distortion.

Other asymmetries: Look for mismatched facial hair, ears, jewelry, or clothing on the left and right sides.

Hair: Hair can appear unnaturally straight, striped, or surrounded by strange halos; sometimes it lacks continuity.

Fluorescent bleed: Unusual fluorescent colors may spill from the background into the hair or face.

Teeth: Teeth are frequently odd or asymmetric, occasionally showing extra front teeth.

Conversely, signs of a genuine photo include perfectly symmetric glasses or earrings, realistic companions, detailed backgrounds with readable text, and overall consistent lighting.

High‑level trick: StyleGAN cannot generate multiple images of the same synthetic person. Request several photos of a person; if they are consistent across angles, the subject is likely real.

While future GANs may reduce these artifacts, the current detection cues remain effective for identifying AI‑generated faces.