COMMERCIAL CHATBOTS DECEIVING DEEPFAKE DETECTORS: THE NAÏVE EXPOSURE OF GENERATIVE AI CAPABILITIES

This article is based on research conducted by Kim et al. (2026) and analyzes how the powerful capabilities of generative AI systems, presented through user-friendly interfaces, can fundamentally undermine state-of-the-art deepfake detectors. Rather than proposing a new manipulation technique, the study demonstrates how an ordinary user, using only standard prompts that do not violate safety guidelines and commercial generative AI systems, can circumvent the most advanced deepfake detection methods. Specifically, the researchers found that generative AI systems explicitly articulate authenticity criteria, externalizing them through unconstrained reasoning processes, and transform these criteria into reusable refinement objectives. As a result, the refined images simultaneously evade detectors, preserve identity verification by commercial facial recognition APIs, and maintain significantly higher perceptual quality. Most importantly, widely accessible commercial chatbot services pose a much greater security risk than open-source models, as their high realism, semantic controllability, and low-barrier interfaces enable even inexperienced users to achieve effective evasion [1].