Comment by estebarb

What happens if we train models (GPT or human students) using the outputs of a model with text havingbthose watermarks? Is there something preventing the watermark from being learnable?