MIT researchers developed Gaussian probing, a method to detect AI models fine-tuned for child sexual abuse material (CSAM) without generating images. Led by graduate student Vinith Suriyakumar and associate professors Ashia Wilson and Marzyeh Ghassemi, the team collaborated with child safety nonprofit Thorn to inspect models' internal adaptations rather than outputs.

The technique achieved 100% accuracy identifying CSAM-specialized models, addressing a critical gap as 1.5 million AI-generated CSAM reports flooded the National Center for Missing and Exploited Children in 2025, up from 67,000 in 2024. Traditional output-based audits are illegal and unscalable; Gaussian probing feeds random data through models to analyze shifts from LoRA adaptors, creating a fingerprint of harmful specialization without generating content.

The method integrates into platforms like Hugging Face and Civitai for automatic screening, resisting evasion better than output filters. Supported by the Bridgewater AIA Labs Research Fellowship, the research also involved MIT postdoc Lena Stempfle and Boston University collaborators. The team plans broader testing and preemptive detection in base models.

While experts note it addresses only LoRA-based fine-tuning—not models trained from scratch on abusive data—Gaussian probing offers a legally sound, scalable solution for the open-source ecosystem where LoRA dominates customization. "Hopefully, we can have a transformative impact," Ghassemi said.