Position: AI/ML Deepfake Research is Misaligned with AI-Generated Non-Consensual Intimate Imagery (AIG-NCII)

Research focus is misaligned with real-world harms: The paper's landscape analysis of highly-cited deepfake research shows that technical interventions almost entirely ignore AI-generated non-consensual intimate imagery (AIG-NCII), despite this being the dominant form of generative AI abuse. Instead, research concentrates on "authenticity detection tools" designed to identify synthetic media, which addresses fraud and scams rather than sexual abuse.

Knowing an image is fake doesn't protect victims: A critical finding is that detecting whether intimate imagery is synthetic does not mitigate harms to the people depicted and may actually worsen them. This means current technical solutions fail to address the core problem—the violation of dignity and consent experienced by subjects, regardless of whether viewers can identify the content as artificial.

The field prioritizes "viewer-centric" over "subject-centric" harms: Existing deepfake research focuses on epistemic harms (threats to truth and authenticity) that affect viewers, while ignoring dignity harms that directly affect the people depicted in non-consensual imagery. This represents a fundamental misalignment between where research efforts are directed and where the most severe harms occur.

AI safety research must expand its scope: The paper calls for threat models in AI safety to explicitly consider subject-centric harms like AIG-NCII, not just viewer-centric concerns. This requires updating how researchers conceptualize and prioritize risks when developing and evaluating generative AI systems.

Responsible research requires safety guardrails and expert partnerships: The authors caution that researchers working on AIG-NCII should implement safety protections for both research subjects and themselves, and must partner with domain experts in sexual violence prevention—establishing a higher standard for responsible research in this high-risk area.