To me the interesting question is: how did the AI acquire enough ontology and bridging to build a subagent whose goals are well-grounded? And grounded in what, so to speak? In the subagent's observable data, or in a fully deterministic ontology where all the uncertainty has been packed into the parameters?
A putative new idea for AI control; index here.
An underdefined idea connected with the challenge of getting an AI to safely move a strawberry onto a plate.
Now, specifying something in the physical world like that is a great challenge; you have to define ontologies and similar. But imagine that the AI had a goal -- any goal -- and that it had to program a subagent to protect itself while it was accomplishing that goal.
Then the subagent will certainly be programmed with a firm grasp of the physical world, and some decent bridging laws should it have an ontology change (if, for instance, quantum mechanics turns out to be incomplete).
This is just an illustration of a general fact: even if its goal is not properly grounded, the instrumental goals will include strongly grounded goals, resilient to ontology change.
This feels related to the fact that even AI's that are given goals in badly programmed natural language concepts ("Make humans* happy*", with the asterix denoting the poor grounding) will still need well-grounded concepts for "human", just to function.
So, is there a way to exploit this instrumental ideal? To somehow set human* equal to human in the motivation? I'm not sure, but it seems there might be something possible there... Will think more.