AI-generated non-consensual intimate imagery relies on configuration rather than dedicated construction: how general-purpose AI technologies facilitate sexual violence

The researchers analyzed 139 R+ guides on the CivitAI platform, known for hosting large amounts of AI-generated pornographic content, and found that these guides explain how to combine sexual and identity concepts into content-generation workflows, and sexual and identity concepts are precisely the constituent elements for producing non-consensual intimate imagery of real people (AIG-NCII); one of the guides explicitly provides an AIG-NCII workflow. The paper's conclusion is that AIG-NCII can be achieved through configuration of general-purpose tools and does not depend on a dedicated set of technical tools; multiple dual-use technical components can be combined to produce NCII outputs in an environment with almost no discussion of ethics. Because most workflows rely on open-weight models, existing technical defenses are limited. The authors recommend that model-layer and product-layer protections focus on identity content generation, and state that these findings can serve as a starting point for evaluating AI models and products based on real attacker behavior as required by new EU regulations.