MIT Develops AI Audit to Block Child Sexual Abuse Material

MIT scientists, collaborating with Thorn, developed a new method to detect if an AI model can generate child sexual abuse material (CSAM) without actually prompting the model to do so. The team addressed a challenge created by laws prohibiting the generation of CSAM, which previously made auditing for this type of content impossible. Their technique analyzes modifications made to AI models during a process called “fine-tuning,” specifically examining changes within the model’s internal structure using a method called Gaussian probing.

The researchers tested their approach on several model variations and achieved 100 percent accuracy in identifying those adapted to generate CSAM. This allows platforms hosting open-source models, and potentially law enforcement, to proactively flag and remove unsafe models before they are widely distributed. The team hopes this work will encourage further research into AI safety and help mitigate the growing harm caused by AI-generated abuse material.