If you're going to rely on signed data from third parties then you're still trusting 3rd parties.

In a dozen or so lines of code you could create a system that collects signed and weighted opinions from a collection of individuals or organisations making encoding arbitration simple. (does the delivery company say they delivered it etc)

You're just kicking the trust can down the road.

On the other hand it's unlikely we'll see any reasonably smart AI's with anything less than millions of lines of code (or code and data) and flaws anywhere in them destroy the security of the whole system.

This is not a great use for AI until we 1: actually have notable AI and 2: have proven the code that makes it up which is a far larger undertaking.