AIBRIEF
A new method from MIT and child safety nonprofit Thorn can flag open source models trained to produce such imagery, without ever prompting them to generate it.
Reports of AI-generated child sexual abuse material to the National Center for Missing and Exploited Children jumped from 67,000 in 2024 to more than 1.5 million in 2025 (MIT News). The surge exposes a legal catch-22 that has long shielded the open-source models behind it: until now, the only way to know whether a model could produce such material was to prompt it and inspect the output, itself a federal crime.
A new audit from MIT and child-safety nonprofit Thorn sidesteps that problem. Instead of asking a model to generate anything, the procedure reads its internal representations and infers whether it has been fine-tuned to produce child sexual abuse material. In the team's tests, the method flagged every CSAM-specialized model variation it was given, a 100% result on the paper's test set (arXiv 2604.25119). Real-world false-positive rates on broad open-source traffic are not yet established.
The technique was developed by MIT EECS Assistant Professor Ashia Wilson, graduate student Vinith Suriyakumar, and Marzyeh Ghassemi's Healthy ML Lab, with Thorn, and presented at the ICML 2026 Trustworthy AI for Good Workshop. Suriyakumar called the prior gap "a huge blind spot that some people were taking advantage of."
Hosting platforms can now screen open-weight model uploads, and law enforcement can test suspected models without generating illegal content. The method applies only to open-weight models whose internals are inspectable; closed systems remain outside its reach.