There is no sure way to identify LLM code. I could write a project by hand with all the GPT "mannerisms" for fun. I could sign my commits with "Co-Authored-By: Claude" because my cat's name is Claude and I find it funny.
Now this rule will either be never applied or when it's applied it will be an arbitrary decision made with mostly inconclusive "evidence". This is not a good way to write rules.
Write rules based on merit. Excessive data usage or licensing problems might come to mind. These things are measurable. Otherwise it's really just posturing.