MIT Develops Method to Detect CSAM-Generating AI Without Creating Illegal Content

Researchers at MIT have developed a breakthrough auditing technique that can identify AI models adapted to generate child sexual abuse material (CSAM) without ever producing illegal content. The method addresses a critical blind spot in AI safety as reports of AI-generated CSAM surged from 67,000 in 2024 to more than 1.5 million in 2025, according to the National Center for Missing and Exploited Children.

The research team, led by MIT graduate student Vinith Suriyakumar and associate professors Ashia Wilson and Marzyeh Ghassemi, collaborated with Thorn, a child safety nonprofit, to create an approach that sidesteps the legal impossibility of testing models by generating CSAM outputs.

Why it matters

The explosion of open-source generative AI has made it trivially easy for bad actors to specialize models for harmful purposes through fine-tuning techniques like low-rank adaptation (LoRA). Traditional auditing methods require generating outputs to evaluate capabilities—an approach that's both illegal for CSAM and psychologically damaging to human reviewers. This technique gives platforms and law enforcement their first practical tool to detect and remove dangerous model variants before they spread, addressing what Wilson calls "a huge blind spot that some people were taking advantage of."

How Gaussian probing works

The technique examines modifications made during the fine-tuning process rather than analyzing model outputs. Using a method called Gaussian probing, researchers feed the model random data points and analyze how it manipulates those data within its internal structure.

The approach captures these modifications at multiple points within the model's layers and averages them to create a signature of how the LoRA adaptor changed the model's computation. "We never run the model all the way to the end or prompt the model, so we never generate images," Suriyakumar explained.

Perfect accuracy in testing

When tested against variations of three model types, the auditing procedure achieved 100 percent accuracy in identifying models adapted to generate CSAM. The researchers compared results to ground-truth data from LoRA adaptors known for generating CSAM, other harmful images, and safe content.

The technique is both scalable and relatively inexpensive to implement—critical factors given that thousands of model variations are published online monthly. It's also more robust than alternative auditing methods, since evading detection would require a malicious actor to carefully alter the base model's inner workings.

Next steps

The research team plans to evaluate their technique on a larger set of model variations and explore whether Gaussian probing can detect harmful capabilities in base models before they undergo adaptation. The work was supported in part by the Bridgewater AIA Labs Research Fellowship.

The findings were presented at the "Trustworthy AI for Good" workshop at the International Conference on Machine Learning, as first reported by MIT News. The research team also includes Lena Stempfle, an MIT postdoc, along with collaborators from Boston University and Thorn.