New AI Safety Approaches Needed to Combat AI-Generated CSAM

Modern artificial intelligence systems introduce significant new risks to child safety, particularly through the misuse of AI to create child sexual abuse material (CSAM) and facilitate child sexual exploitation. This paper asserts that current AI safety methodologies are insufficient for addressing these specific threats. Existing techniques often rely on data accessibility, transparency, and evaluation practices that are ethically and legally unfeasible when dealing with CSAM. The authors detail how these constraints create unique technical challenges, such as limitations on dataset auditing, red teaming, and fine-tuning for prevention. They identify 15 critical open problems related to online child sexual exploitation and abuse throughout the entire AI development lifecycle, from initial dataset curation and model design to deployment and ongoing maintenance. The paper concludes with targeted recommendations for researchers, developers, and policymakers, urging a reframing of AI-facilitated child sexual abuse prevention as a central, safety-critical dimension of AI research.