The policy only says that you have to make sure there are no copyright violations, not that AI is banned outright.
Reading the generated AI code and writing something based on it would be fine (just as we often have a look at gpl driver code in Linux for reference, but we don’t directly copy it).
Searching the web for cooe snippets to make sure they are original and not copypasted from somewhere else would also satisfy the policy. Don’t make it more extreme than it is.
Now, you will still have to pass code review and answer to all questions from Haiku developers. From my experience, that is much harder than getting the code to work, and personally I am not muchinterested if the review consists of me talking with the llm directly or indirectly: these things cannot learn from code review comments. If there is a human willing to learn from the code review, no problem with that, I think?