Decide when your AI use needs disclosing

Your team has no written rule on disclosing AI help. In one sprint, you used an assistant 4 ways:

1. It wrote a throwaway Python script that backfilled one column, once.
2. It wrote a non-trivial concurrency change in your Java payment service.
3. It suggested the wording of a customer-facing error message.
4. It drafted a design doc that you then rewrote heavily.

For which of these does disclosure matter, and what principle decides?