By design, a watermark would make it easy to create a discriminator that distinguishes between LLM content and human content. In that case, just make a discriminator yourself and use regex to find and remove any of the watermarks.