MIT, Thorn Build 100% Accurate Audit to Detect CSAM AI Models Without Generating Images
Updated
Updated · MIT News · Jul 13
MIT, Thorn Build 100% Accurate Audit to Detect CSAM AI Models Without Generating Images
1 articles · Updated · MIT News · Jul 13
Summary
MIT and Thorn said their new "Gaussian probing" method identified AI model variants tuned to generate CSAM with 100% accuracy, without ever prompting the model or producing an image.
The approach targets LoRA fine-tuning adaptors and analyzes hidden internal representations from random inputs, sidestepping a legal barrier because generating CSAM for testing is illegal in the U.S. and many other jurisdictions.
Tests across three model families compared against known CSAM, other harmful, and safe adaptors, suggesting hosting platforms could flag or block unsafe uploads before they spread.
The need is growing fast: the National Center for Missing and Exploited Children received more than 1.5 million reports of AI-generated CSAM in 2025, up from 67,000 in 2024.
Researchers said the technique is relatively cheap, scalable for thousands of monthly model uploads, and harder to evade than some alternatives, though they still plan broader testing.
A new AI tool has 100% accuracy in finding illicit models. But are developers already creating AI that can outsmart it?
New laws now criminalize AI models, not just their illegal outputs. Is this the future for regulating artificial intelligence?
With AI making abuse content easier to create, can technology alone ever solve a deeply human problem?
Non-Generative AI Auditing: Gaussian Probing as a Breakthrough Against Harmful Specialization and CSAM
Overview
The rapid rise of generative AI brings both powerful new abilities and serious risks, especially the threat of models being fine-tuned to create harmful content like child sexual abuse material (CSAM). Traditional methods for detecting such misuse are not only ineffective but also illegal and unethical, as they require generating the very content they aim to prevent. Addressing this urgent challenge, MIT and Thorn have introduced a groundbreaking solution called Gaussian probing. This method allows platforms to assess whether AI models are specialized for harmful tasks by analyzing their internal workings, without generating any illicit material. This innovation marks a major step forward in AI safety, offering a practical, ethical, and scalable way for industry and regulators to proactively protect children and uphold responsible AI governance.