AI-generated faces can be spotted by studying eyeballs, researchers find

Their proposed method involves analysing light reflections on eyeballs using tools that were originally developed to help astronomers study galaxies.

How does it work?

These days, AI tools are not just capable of creating a human face but are also advanced enough to capture intricate facial features such as dimples, refined nose, skin tone, etc. However, the eyes of an AI-generated face could be a giveaway since there are often inconsistencies in the light reflections cast on the eyeballs.

As a result, scrutinising eyeball reflections in images of human faces is the crux of the AI detection technique developed by the researchers. Specifically, it involves running the eyeball reflections’ morphological features through indices to see if the light reflection in the left eyeball matches with the light reflection in the right.

“If the reflections match, the image is likely to be that of a real human. If they don’t, they’re probably deepfakes,” read a blog post by the Royal Astronomical Society.

Researchers are able to automatically measure and quantify eye reflections by applying astronomy tools. These same tools have been used by astronomers in the past to determine the shapes of galaxies.

What are the obstacles?

While the new AI detection technique has a lot of potential, it is not exactly foolproof. For instance, the method works only if it is possible to zoom in and get a clear view of the eyeballs. There is also a risk that the faces of actual humans can be detected as AI-generated since real photos may have inconsistencies in eyeball reflections too.

“There are false positives and false negatives; it’s not going to get everything,” Kevin Pimbblet, professor of astrophysics at the University of Hull, was quoted as saying. Furthermore, the method could become useless if AI models evolve to generate consistent eyeball reflections.