Would it be ethical to implement AI for self-defence for public walking robots which are exposed to dangers such as violence and crime such as robbery (of parts), damage or abduction?
What would be pros and cons of such AI behavior? Is it realistic, or it won't be taken into account for some obvious reasons?
Like pushing back somebody when somebody start pushing it first (AI will say: he pushed me first), or running away on crowded street in case algorithm will detect risk of abduction.
The question mentions "walking robot", but it may be illustrative to re-frame the discussion in terms of self-driving cars, because:
So, should a self-driving car run someone over (likely killing them) if they try to steal it? I'm hoping that few people would argue that it should.
Should it attempt to do a lesser amount of damage (say, calculated to hopefully only break a leg)?
Again, I'd argue not. The main reason for saying this is that our decision-making algorithms are simply not sufficiently context aware to be able to decide whether theft or harm is the intent. To concretely illustrate this: a recent fatality arose because a self-driving Tesla was oblivious to context to the extent that it couldn't distinguish between a high-sided van and empty space.
Under those circumstances, it's probably best not to allow commercial autonomous systems to cause physical damage (even to inanimate objects).
'Running away' (or rather, 'driving away', in the case of the car) is another matter: driving is what it's designed to do.
It depends on whether the loss of the robot would end up causing harm to humans.
If the robot was supposed to be watching for a suspected terrorist attack to start taking place (so it could alert authorities or halt the attack), it would be very bad if somebody dismantled the robot or otherwise stopped it from carrying out its mission. In that case, the device would be certainly justified in stopping humans from injuring it in any meaningful way.
A robot carrying classified information should probably be similarly willing to protect itself, since the spread of such data could bring harm to a state or a lot of people.
If an AI-enabled device was just walking the streets in the course of carrying out some mundane task, I think it would be hard to justify allowing the robot to incapacitate a human attacker. After all, it was made - presumably - to serve humans.
No matter whether the AI was programmed to defend itself, people couldn't just impede or damage it with impunity. Intentional destruction of another person's property (including public property) is almost certainly a crime, as is intentional obstruction of law enforcement. It wouldn't have to be up to each robot to defend itself; it could just send information about the perpetrator to C&C before its demise.
