Text AI watermarks will always be trivial to remove

Would using Stable Diffusion Image to Image with a low noise ratio disable most steganographic techniques? I imagine the pixels would be scattered enough to erase most embeddings. I've never had to scrub images in my career, but that's what I would try.

I presume all the Foundation AI companies have been doing this for years, to avoid training models on their own generated data?