Hacker News Reader

Comment by kens | Hacker News Reader

Seriously, I think elevators are a good model for this, as they are autonomous systems that occasionally kill people. Who's at fault? If the manufacturer messed up the design, they're liable. If the owner ignores maintenance, they're liable. If someone intentionally sabotages the elevator to kill someone, that would be murder. If someone is elevator surfing and dies, then it's their own fault. It's the same with other autonomous systems; the blame could lie anywhere, depending on the circumstances. (I've simplified the legal issues somewhat, of course.)

View on HN

I think factory floor robots will be a grey area before elevators. As long as things are deterministic then "fault" can be ascribed - it was the workers fault for climbing over the barriers, or it was the manufacturers fault for not testing the arm would stop moving in < 1.2 seconds.

But as soon as AI / ML becomes involved, it stops being deterministic and becomes probability. "Judgement" is asked of the robot arm. Is that image a person or a car door.

I think this is likely to hit before cars simply because so many of us drive cars but none of us operate giant robot arms.

That "judgement" is a design decision. A ML algorithm misidentifying a person as a car door isn't that different from an elevator cable snapping. That happens more or less randomly too. It could be bad design, bad maintenance (e.g. real world data drifted away from training data but no retraining took place), or just bad luck (an element in almost any accident). Depending on the fault someone or nobody is at fault. And even if nobody is at fault we can still decide to advance the state of the art, and that in the future no single elevator cable snapping should cause more than a minor inconvenience (due to multiple cables, breaks etc).

Eh, this.

It all depends on if 'robots', however one defines them are 'conscious', however one defines that.

But until that gets settled, it's always a human pulling the trigger.

No difference from when someone's pitbull chews off someone else's head.

It might be the owner's fault, it might be not. But the dog in question can not be blamed.

But supposing the robot is "guilty". How (or why) would anyone punish it? Put it in jail? I don't think it will mind. So what would be the point of finding it guilty? Rather, it would make more sense to, in some way "fix" the robot.

Same would be true of people, were they easily "fixed".

I think we should treat it as if an animal killed a human. We don't call that murder. If a dog kills a human, I think it is usually put down? Not sure if a whale does it. So it may depend. But in any case, murder I think is reserved for conscious killing of a human by another human.

I'm not sure about an animal killing a human, but certain animals, like primates, definitely commit murder (i.e. the conscious killing of another member of its tribe for territory, social status, or whatever).

Everything living kills everything in sight for resources. From viruses to civilizations. To some extent this can be moderated, or for a time some optimizations might result in even less destruction, but this is the way this world is.

Killing is rational in a sense - for some subset of purposes, in context of some subset of environmental pressures, both of which change constantly, and often faster than adaptation to kill does.

Accept this. It will not ever change.

I agree they do but I don't think the word murder is applicable

If we could "fix" the animal, rather than killing it, wouldn't we?

People are self-reproducing robots that are made of meat. Robots made of meat are very hard to fix.

I actually think owners of pets should bear criminal liability for their pets actions.

A common metric already used in industrial automation liability examination is Safety Integrity Level. [1] This is a probabilistic measure of how likely a failure is.

[1] https://blog.msasafety.com/what-safety-integrity-level-means...

These discussions are always around the hypothetical of artificial intelligence, so real world analogies don't really apply. Today's automations will do whatever they are programmed to do, so the blame can always be traced back to a person. If the robot wakes up and independently decides "I'm going to kill someone today!" then all bets are off.

I would say that in a case like that we would still be well within the bounds of the existing legal frameworks to assign blame according to contributing factors, strength of causality, etc. There's no legal principle that only one person can be criminally liable for murder. Therefore, if the AI is a legal person, it could be criminally liable. If its designers designed it with murderous intention, or reckless disregard for life, so could they. It would just depend on the circumstances, as would the punishment. I'm sure the law would need to be updated to take into accounts the rights of created persons, if we ever get to that point, but it doesn't seem remotely insurmountable.

If you design a robot that can independently decide to kill people, that's on you.

If you birth a child who kills someone when they turn 18 is that on you as well?

The significant disanalogy is the amount of control parents have over their child's character compared to a programmer. Parents can pretty much do their best to raise a child, but the child may still end up screwed up. I wouldn't really put the blame at their feet. If they deliberately raise their child to be a killer, that's obviously another matter. By contrast, an AGI designer is in an unprecedented situation where they essentially have total control over the character of the being they're bringing into existence. If the AGI does bad things, it's very much on the programmer.

That's assuming the AGI designer even understands what they built. If the first AGI is built via some semi-automated bootstrapping process based on a lesser intelligence, you can't really predict how the resulting intelligence will behave.

For sure, which means it's the designer's responsibility to do due diligence in making certain they do understand it. Kickstarting such a process is not an ethically responsible thing to do, unless they are sure it and its future iterations are aligned.

No, but if you tell someone to kill someone else, and that person does, you'll be found guilty - assuming the normal standards of you reasonably expecting a person to do that, you being mentally competent, etc.

Those caveats apply to machines as well: if I tell a toaster to kill someone, and then the toaster electrocutes them, I would not be guilty.

The key word is "independently". If you build a robot that is capable of doing anything, and it chooses to kill someone (without you specifically telling it to), whose fault is it? It is pointless to speculate about it today, since the status of a hypothetical intelligent robot of the future could be that of a pet (you are responsible), a child (you are not responsible) or really anything else.

If I build a gun mount, that uses a basic computer vision algorithm to identify people out a window. It then randomly select a target - “true” random so no claims I set a seed - and then shoots them.

Do you think there is an argument that I should not be responsible?

If “random” is too independent, or you want it to be systematic, we could do something else:

* the person with the highest “red hat” score according to a 20 year old copy of opencv

Or we could do something modern and “AI” and give it a completely random neural net that’s super deep and has three outputs: angle, pitch, and whether to pull the trigger.

Or we could go all out and just replace this hypothetical machine that acts like an independent person with the next best thing: a regular human.

If a government, or a terrorist group, gives a group of people (or soldiers) a bunch of weapons and drops them off in a city, and they start shooting, do you think that the government would be responsible?

That’s easy. Your intent was for it to maybe shoot someone, and that is a criminal intent. If you had no such intent, it’s either accidental murder or no guilt (e.g. you built a robot that takes a heap of non-classy trash and throws it into a grinder; an intruder was grabbed by it accidentally because they weren’t well-dressed).

This discussion mistakes existing policies as some objective base, but it’s the other way round.

Well yeah kinda. Name any mass murderer we've had and tell me the parents didn't have to hold account - perhaps not always legally, but do they have jobs or friends? Even if they do, will it ever be the same?

They'd have decided it by running their code, which is not exactly independence.

> elevator surfing

I had no idea that was actually a thing. https://en.wikipedia.org/wiki/Elevator_surfing

Reply on news.ycombinator.com