Researchers say deepfake studies ignore non-consensual intimate imagery harms

A new position paper published on arXiv argues that mainstream AI/ML deepfake research is fundamentally misaligned with the reality of AI-generated non-consensual intimate imagery (AIG-NCII). The authors contend that highly cited technical interventions predominantly address viewer-centric epistemic harms, such as fraud or misinformation, while neglecting subject-centric dignity harms associated with sexualized abuse. Through a landscape analysis of existing literature, the researchers demonstrate that current detection tools rarely account for AIG-NCII specificities and may even exacerbate victim harm by validating synthetic content as authentic. The paper recommends updating threat models to prioritize subject safety and urges researchers to implement strict guardrails when studying this high-risk domain. This critique suggests that prevailing industry safety benchmarks are insufficient for addressing the most prevalent form of generative AI abuse currently affecting individuals.