Everyone assumes AI becomes dangerous when it makes the wrong decision.
I think the bigger challenge begins when two capable AI agents make different decisions, and both have a reasonable case.
That's not a bug.
It's what happens whenever agreements involve context instead of