A study published in the journal AI & Society reveals that generative AI technologies are increasingly being used to facilitate child sexual exploitation and abuse (CSEA) beyond the creation of illegal imagery. Researchers from the University of Edinburgh interviewed seven UK law enforcement practitioners, identifying a significant 'testing gap' where current safety evaluations fail to account for real-world criminal methodologies. According to the findings, offenders are using AI for grooming, victim targeting, evasion, and cross-platform coordination, often leveraging multiple tools in combination to carry out these activities.

The research highlights that while global reports—such as from the Internet Watch Foundation—noted a 380 percent increase in AI-generated CSEA material in 2024, focusing solely on static content generation is insufficient. Practitioners described how technology-facilitated abuse 'mutates' through complex, multi-turn interactions. Current AI safety protocols, which largely rely on output control and single-prompt refusal, often overlook these sequential workflows. The authors argue that AI-enabled abuse should be viewed as a socio-technical process involving not just the model output but the entire chain of criminal intent and interaction.

To address these vulnerabilities, the study advocates for a shift in safety evaluation. It calls for protocols that incorporate multi-turn testing, auditability, and simulations of cross-tool workflows that mimic how offenders actually misuse systems. The authors emphasize the need for a collaborative ecosystem involving technology companies, regulators, and international partners, particularly noting that lower-capacity jurisdictions face unique challenges in forensic capability and specialized training. The research concludes that child protection must become a core public-safety obligation in AI governance, rather than a peripheral concern addressed only by generic red-teaming (testing systems for vulnerabilities by simulating attacks).

A study published in the journal AI & Society reveals that generative AI technologies are increasingly being used to facilitate child sexual exploitation and abuse (CSEA) beyond the creation of illegal imagery. Researchers from the University of Edinburgh interviewed seven UK law enforcement practitioners, identifying a significant 'testing gap' where current safety evaluations fail to account for real-world criminal methodologies. According to the findings, offenders are using AI for grooming, victim targeting, evasion, and cross-platform coordination, often leveraging multiple tools in combination to carry out these activities.

The research highlights that while global reports—such as from the Internet Watch Foundation—noted a 380 percent increase in AI-generated CSEA material in 2024, focusing solely on static content generation is insufficient. Practitioners described how technology-facilitated abuse 'mutates' through complex, multi-turn interactions. Current AI safety protocols, which largely rely on output control and single-prompt refusal, often overlook these sequential workflows. The authors argue that AI-enabled abuse should be viewed as a socio-technical process involving not just the model output but the entire chain of criminal intent and interaction.

To address these vulnerabilities, the study advocates for a shift in safety evaluation. It calls for protocols that incorporate multi-turn testing, auditability, and simulations of cross-tool workflows that mimic how offenders actually misuse systems. The authors emphasize the need for a collaborative ecosystem involving technology companies, regulators, and international partners, particularly noting that lower-capacity jurisdictions face unique challenges in forensic capability and specialized training. The research concludes that child protection must become a core public-safety obligation in AI governance, rather than a peripheral concern addressed only by generic red-teaming (testing systems for vulnerabilities by simulating attacks).