Comment by fluoridation

It seems to me like the opposite solution is more robust. Rather than putting digital watermarks on AI-generated content, put them on human-generated content, and you can treat anything that doesn't have one as possibly AI-generated.