Predicting Deepfake Enjoyment: A Machine Learning Perspective

Deepfakes, synthetic media created through advanced learning techniques, are rapidly advancing, empowering users to craft deceptive videos via voice manipulation, realistic lip-syncing, and seamless face replacements. From sophisticated methods to user-friendly 'cheapfakes', concerns persist regarding their potential misuse and ethical implications. However, despite their prevalence, our empirical understanding of deepfakes remains limited, particularly concerning the psychological processes influencing their reception and consequences. Existing experimental research primarily links deepfakes to heightened uncertainty about media content and explores their negative repercussions on social media. Additionally, most of these studies rely on linear statistics, inherently aprioristic, to comprehend the cognitive-emotional factors in deepfake processing. Addressing these gaps, our study explores individuals' engagement with non-deceptive deepfake content through a predictive non-linear approach, utilizing machine learning techniques uncommon in psychological research. Results showcase the efficiency of machine learning in predicting attitudes toward deepfakes, including enjoyment and sharing behavior. They offer highly precise information about the specific factors determining such responses. Overall, the study demonstrates that applying machine-based analysis models to responses to deepfakes can be used to automatically predict and control their virality.