There are several personality disorders in which there is reduced cognitive or affective empathy in different combinations. Someone with a high cognitive empathy, low affective empathy, a willingness and skill in manipulation, and lack ethics, can cause problems. I don't see how that would be different when it is a robot or android. (We have problems with socially engineering dopamine feedback loops to raise "engagement" on our apps and platforms, and as a society, we have not figured out what we want to do about this yet).
Lastly, in regards to manipulative behavior and trust -- there is an insight from Crucial Conversations, and that is that when someone feels that another person has their best interest at heart, it is easier to trust that person even when talking about difficult subjects. I think that applies here to the question of AI and ethics. It is not so much as, whether it is more ethical for a self-driving AI to sacrifice one group of people to save another, but whether individuals feel that the AI has their best interest at heart (whatever that means).
It seems like affective empathy, actually feeling, would be something you would only want from any person or person-equivalent since implies they do what you want only because they like you and they'd expect you to satisfy them in turn, a problematic relationship to have with a robot. I'd want a robot doing what I told it to do with various fail-safes and feedback (obviously, knowing something made me unhappy would be a moment to request feedback).
Both anti-social personality disorder and narcissistic personality disorder has reduced or no affective empathy and often high cognitive empathy. However, I would be more inclined to work with the "pro-social" ASD than someone with NPD. Someone with NPD is incapable of thinking beyond themselves while someone with ASD is willing to entertain mutually beneficial arrangements.
You could swing the opposite way too. Someone on the autism spectrum would have reduced cognitive empathy yet high affective empathy. They tend to care more, but only after understanding, and then the emotion tends to be more intense.
Likewise, there are people who feel deeply what everyone else is feeling to the point where they lost their center and and is unable to function as an individual.
The main point I was making is that people tend to confuse one type of empathy for another. And the other is that, what is a good test is if you trust the other person or AI to have yout best interest at heart.
It seems like the appropriate analogy would be something like a drug or a video game. I would claim it is unwise in the extreme to allow a robot with emotional-influencing abilities/programming to first calculate one's "best interests" and then go about imposing those best interests without getting overt, conscious go-ahead from the user.
Of course, even with human beings, you want to interact with people who have an intention to be good to you and whose idea of "good" is somewhat in accordance with your own ideas of good, otherwise you get nuts and cults and so-forth. With robots, devices which function imperfectly and which we don't want just imposing their "values" on society, jeesh, obviously, one wants tight control what concept they have of "best interests" along with tight control on what sort of influencing they do.
The inability of people to see human-robot interactions as rather distinct from human-human interactions, just makes the situation more urgent.