Deepfake Research Misaligned with Addressing Non-Consensual Intimate Imagery Harms

Current research in AI/ML concerning deepfakes is critically misaligned with the most prevalent and damaging forms of generative AI abuse, specifically AI-generated non-consensual intimate imagery (AIG-NCII). The paper highlights that while deepfake research predominantly addresses "epistemic harms" related to truth and authenticity, it largely overlooks the severe "dignity harms" experienced by individuals targeted by AIG-NCII. A landscape analysis of highly-cited works demonstrates that technical interventions for deepfakes almost exclusively focus on authenticity detection, failing to address the core issue of sexualized imagery abuse. The authors argue that merely knowing an image is synthetic does not mitigate harm to the subject and can even worsen it. They recommend a significant shift in the field, urging researchers to update their threat models to include subject-centric harms, integrate AIG-NCII into AI safety research, and establish robust safety guardrails and partnerships with sexual violence prevention experts before engaging in this high-risk domain.