Against the backdrop of explosive growth in large models and computing power technologies, the realism of AI-generated images has reached unprecedented levels, posing potential threats to social security. A joint study by the University of Leeds, the University of Reading, and other institutions, published in the journal *Royal Society Open Science*, focuses on the efficiency of identifying AI-forged faces. The research team invited 664 participants to take part in an experiment, using high-quality synthetic face images generated by the StyleGAN3 system as test materials. Without any prior training, the average person's recognition accuracy was only 31%, while even professional “super-recognizers” achieved a rate of just 41%. This reflects the high degree of concealment in the detail processing of current AI technology, highlighting the risks brought about by increased computing power.
Regarding the experimental design, the researchers used a visual training module to teach participants how to identify common flaws in AI-generated images, such as abnormal dental arrangements, unnatural hairlines, and asymmetrical ears or accessories. The results showed that this training took only about five minutes yet produced a qualitative leap: the accuracy of ordinary individuals rose to 51%, while super-recognizers climbed to 64%. This data not only demonstrates the plasticity of the human visual system but also underscores the importance of infrastructure in the field of AI security—efficient training tools can be rapidly deployed in practical application scenarios, such as financial authentication and social media content moderation.
Under current AI industry trends, the technical threshold for AI-generated faces is continuously decreasing, leading to their widespread misuse in creating fake social accounts, fabricating false documents, and bypassing identity verification systems. This directly threatens the stability of digital infrastructure. Dr. Eilidh Noyes from the University of Leeds noted that developing effective identification methods has become a critical issue in the security domain, especially as large models and chip technologies drive the exponential growth of AI generation capabilities. The research team plans to further explore the persistence of training effects and attempt to combine human visual advantages with AI-powered automated detection tools to build a more robust defense system.
This discovery has profound implications for AI industry trends: it not only provides new ideas for security solutions but also emphasizes the core role of computing power and infrastructure in addressing AI-related risks. In the future, as large models and chip technologies become further integrated, the short-term training model could be standardized and promoted, helping enterprises and institutions rapidly enhance their discernment capabilities and thereby maintain the healthy development of the digital ecosystem.