Overview
- The peer-reviewed study, published November 12 in Royal Society Open Science, evaluated 664 participants from the Universities of Reading, Greenwich, Leeds and Lincoln.
- Without instruction, typical observers identified AI-generated faces 31% of the time, while super-recognizers reached 41%.
- A five-minute lesson highlighting common artifacts such as irregular hair patterns and teeth anomalies raised accuracy to 51% for typical participants and 64% for super-recognizers.
- Researchers caution that performance after training still nears chance levels for many people and that synthetic faces are often judged more trustworthy because they appear more average and familiar.
- The team tested images from StyleGAN3 and recommends pairing brief human training with super-recognizers and automated detectors, with further work planned on how long the training benefits last.