Deepfake Detection Needs Social Theories, Not Just Artifact Analysis

A new arXiv position paper challenges the prevailing approach to deepfake detection. Authors Jessee Ho, Shweta Khushu, and Shaina Raza argue that traditional artifact-based methods—which classify media as real or synthetic—are failing against rapidly improving generative models. They identify five core assumptions driving current detection, all of which are eroding, leading to what they call the "Generalization Illusion." The paper contends that for interactive deepfakes (e.g., impersonation in live video/voice calls), the real harm is the act of deception, not the manipulated signal itself.

To address this, the authors propose a complementary analytical layer grounded in three established frameworks from philosophy of language and social psychology: Speech Act Theory (examining intent behind utterances), Grice's Cooperative Principle (analyzing conversational norms), and Cialdini's principles of influence (evaluating listener response). This multi-level approach examines deception at the utterance, conversation, and listener-response levels, aiming to catch manipulation that current detectors miss. The paper concludes with open research problems for developing more robust forensic methods that treat deception as a social phenomenon.