CroatiaA new method has been developed by researchers at MIT in collaboration with the nonprofit organization Thorn to identify AI models capable of generating illegal content such as hate speech and child sexual abuse material, without creating any illegal content during testing. The technique examines internal mechanisms of AI models to determine if they are specialized for producing harmful outputs, achieving 100% accuracy in identifying such models during tests. This approach allows platforms to detect, label, and remove dangerous AI variations before they spread widely. Researchers emphasize the scalability and cost-effectiveness of their solution, which addresses a critical gap in AI safety.
Bias read (Center): The article discusses a technological development related to AI safety and does not present any political viewpoints, biases, or controversial policy implications. It focuses on technical innovation and its application for detecting harmful AI models.
