Overview
- The peer‑reviewed study published in PNAS on Monday, June 29, 2026, found that brief training raised average detection accuracy from about 40 percent to roughly 80 percent.
- The method teaches people to judge six global facial qualities—symmetry, proportionality, attractiveness, expressiveness, distinctiveness and memorability—rather than look for fleeting artifacts.
- An independent replication by University of Victoria researchers reproduced the effect and reported similar accuracy gains plus faster response times, with top performers approaching near‑perfect scores.
- Researchers tested the approach using StyleGAN images and warn the training has not yet been shown to generalize to other generators, to audio or video deepfakes, or to vulnerable age groups and must be tested for durability as models evolve.
- Because the training is short, cheap and explainable, authors say it could be paired with opaque algorithmic detectors to give people a practical tool for reducing fraud and helping humans stay in the loop.