This is how you will recognize an image generated by AI. Just pay attention to the face

Counting extra fingers or looking for distorted jewelry is no longer an effective way to detect graphics created by artificial intelligence. As described by Spider’s Web, Australian scientists have developed a method based on the analysis of general facial features, not individual errors.

Researchers argue that modern AI models are increasingly better at eliminating typical artifacts, which is why it was necessary to develop a new approach.

Six elements worth paying attention to

A team from the Australian National University identified six features that help distinguish real photographs from AI-generated graphics. These are:

• symmetry,

• proportionality,

• attractiveness,

• expressiveness,

• distinctiveness,

• memorability.

As Spider’s Web reports, faces created by artificial intelligence are usually more symmetrical, proportionate and attractive. In turn, photos of real people often seem more characteristic, easier to remember and richer in natural expression.

Where do these differences come from?

The authors of the study explain that image generators learn from huge sets of photographs and create new faces based on statistical relationships between millions of examples.

The result is faces that resemble a mathematical average – harmonious and visually correct, but often devoid of individual features characteristic of real people.

Short training significantly improved effectiveness

In the study, participants first independently assessed the authenticity of the presented faces, and then underwent a short training involving the analysis of six indicated features.

According to Spider’s Web, the average effectiveness of recognizing images generated by AI increased from 41.4%. to 81.1%, and the best participants achieved almost 100% accuracy. The results were then confirmed by an independent team of researchers from the University of Victoria in Canada.

However, the authors state that the study concerned images generated by the StyleGAN model. Further work will check whether the method will be equally effective in the case of newer generators and audio and video materials using deepfake technology.