“Current agents cannot truly internalize or understand the relative importance of different objectives in the way humans naturally do through experience and context. Instead, they operate more like highly sophisticated pattern-matching systems, trying to find solutions that match their training data rather than truly understanding and resolving the underlying conflicts.”