Radioactive Watermarks for AI Images with Louis Kerner

What happens when AI models increasingly train on content generated by AI itself? Researchers fear that models could gradually lose quality over time. This phenomenon is known as “Model Collapse.” In this episode of CISPA TL;DR, CISPA researcher Louis Kerner joins us to discuss BitMark, a new watermark for AI-generated images developed by him and his colleagues. Under certain conditions, BitMark can remain detectable even when other AI models are later trained on marked images. Why the researchers describe BitMark as a “radioactive” watermark, how it works, and why approaches like this could be important for the future of generative AI are all topics we explore in this episode. This episode was recorded in German. Enjoy listening!