Might attempting to align AI cause a "near miss" which results in a much worse outcome?