How do you quantify friendly in a meaningful way? If AI systems were in control over human systems, and it's involvement would directly harm or benefit humans,

How do you quantify friendly in a meaningful way? If AI systems were in control over human systems, and it's involvement would directly harm or benefit humans, wouldn't the friendly/unfriendly designation vary based on perspective? What I mean is let's say the AI is taught morality from a pure utilitarian standpoint. Then, certain ethical decisions might harm a minority to benefit humanity overall. Is this AI friendly or not? Ethical systems, like all first-order logic, are inconsistent systems, and there will be ambiguities. Good and bad are human concepts that are deeply rooted to a given observer's perspective.