Artificial Intelligence Generated Deepfakes as Instruments of Disinformation: Examining Their Influence on Public Opinion, Digital Trust, and Governance

Abstract

The paper examines the use of artificial intelligence-generated deepfakes as a socio-technical threat, one which can be developed at three interlinked levels: technical detection, human perception, and governance. We test four image-only image-deepfake detectors, including GAN-based, diffusion-based, Vision Transformer (ViT-B/16), and CLIP ViT-B/32, using the FaceForensics image dataset, which has undergone a uniform preprocessing pipeline that includes facial-cropping, image-alignment, resolution normalization, and quality filtering. Transformer-based models are, by a significant margin, more successful than both GAN- and diffusion-based detectors, with CLIP ViT-B/32 being the most successful and obtaining the highest classification accuracy and an almost perfect ROC-AUC, which highlights the importance of large-scale pretraining and attention-based models on synthetic media forensics. To add to these technical experiments, there is a human-subject experiment indicating that participants are always more efficient in authentic image recognition than in deepfakes, with a general deepfake detection accuracy falling close to that of chance, and with a relative weakness in differentiating between age groups. Deepfakes not only result in a high rate of false categorization but also cause a significant decrease in the level of trust, despite the fact that the perceived credibility scores are not only displaced but also significantly lowered. Lastly, the policy and regulatory text topic modeling has shown an unequal panorama of emerging but inconsistent governance with a focus on identity protection and election protection issues, and minor reference to actual enforcement tools. Combined, the results can be interpreted to support the potential and shortcomings of modern AI-driven detectors, the susceptibility of human judgment, and the necessity of an improved and more enforceable regulation and specific media literacy to maintain online trust.