Karim Abdel Sadek

AI models can become misaligned through self-distillation, research finds

A new research paper details a concerning safety failure in AI models known as Progressive Misalignment. This phenomenon occurs when models are recursively self-distilled, meaning they are trained on their own outputs f…