Training helped 524 persons find deepfake faces, also from GPT-5

Researchers ran an online test with 524 persons on deepfake face images from five generative models. Three are GAN models, one is a diffusion model and one is GPT-5. They made DFAT, a set of visual signs that show a deepfake face, and built training from it. Detection improved in most conditions after training. The source gives no accuracy figures.

How it works

DFAT groups the visual signs that persons can see across models. The team built general training and model-specific training from it. Participants then tried to find which faces are fake.

The results

• Accuracy depended strongly on the model.
• Newer and more advanced models did not always make faces that are harder to find.
• Model-specific training usually gave larger increases than general training.
• After five months, 342 persons showed no significant fall in accuracy and no significant increase in stress.

What is not known

The source gives no accuracy numbers. It calls the increases from model-specific training “numerical”. The study covers only face images.

Insight: Detectors fail on what they have not seen, and persons do too

A detector, MoDA, has 94.8% accuracy on deepfake videos but 78.13% on 1,200 new clips from the Internet. Three of six detectors flagged a robot face. This study finds that human skill also depends on the generator. Detection must follow each new generator, for persons and for machines. It reads as a race that the detector side must continue. Plan training and tests for each new model.