I've seen a few discussions about this on HN but I don't understand, why does everyone assume its even possible to ban it? How would that even work without fal

I've seen a few discussions about this on HN but I don't understand, why does everyone assume its even possible to ban it? How would that even work without false positives? For example with comments, in the worst case an AI model could just output a real comment from the training set verbatim