What is it about?
AI-generated faces are now so convincing that they fool most people, with average accuracy little better than chance. Despite the lack of obvious clues, we showed that people can be trained to spot them simply by viewing and rating AI and human faces on six global qualities that distinguish them, such as symmetry and attractiveness. With less than an hour of training, participants' accuracy almost doubled, and they also became faster at spotting AI faces.
Featured Image
Photo by Danielle-Claude Bélanger on Unsplash
Why is it important?
AI-related fraud is projected to cost 40 billion USD by 2027, driven by image-generation technology that is improving faster than we can keep up with. Algorithms offer one line of defence, but their decision-making remains opaque and recent benchmarking has revealed serious weaknesses. Our training approach offers a transparent, human-centred alternative. Because it relies on implicit learning through exposure to current AI images, and targets global qualities rather than specific defects, our training is well placed to keep pace as AI models continue to evolve.
Perspectives
We have shown our training is effective for some of the most convincing fakes available, StyleGAN3 faces. Now, we need to find out whether that training generalises to other AI-face models, AI-voices and deepfake videos. We are also working on how to optimise the training – making it shorter and ensuring the benefits last over time. Interested in participating in our AI face detection training? Register here: https://tinyurl.com/ai-face-study-register
Amy Dawel
Australian National University
Read the Original
This page is a summary of: Training humans to detect AI-generated faces, Proceedings of the National Academy of Sciences, June 2026, Proceedings of the National Academy of Sciences,
DOI: 10.1073/pnas.2602122123.
You can read the full text:
Contributors
The following have contributed to this page
