While people have varying stances, I understand the policy at least for the main Haiku repo code. Allowing AI-written contributions means you will get code that neither the contributor nor the teammembers fully understand, and asking questions about the wider context becomes futile. It would require a lot of distrust to get consistently acceptable results.
I think the fact that there are multiple people who have deep understanding of what they are working on and of the system's structure goes a very long way to establish this trust and ensure quality.
In this environment, I think (as I said, the opinions are varying), having the contributor learn about their contribution (even asking AI as a starting point, if they want to) and then (re-)write it themselves is a really minimum gate to make sure everyone knows what is happening.