>language that a neural network can't learn

how to do this? how would language have to change for us to be able to immediately tell if a post was written by a human?

if it is unclear, chaotic, full of metaphors, slang, and sarcasm, even if the AI is not precise all its mistakes could be interpreted as style.
if it is precise, clear, and orderly, it's going to be easier for the AI to learn it.

are there other ways? assuming the AI is constantly learning from what we post, could we hack it the same way you inject code into a website? could we have different dynamically changing rules that human posters can adapt to but AI can't?
what kind of text is most difficult for google translate?