So the watermark can be removed by rearranging words and choice of words. This seems trivial to bypass with a local model. If I understand this correctly.