Making a safe robot

It’s alive!

For as long as stories have been written, authors have considered the possibility that if intelligence exists in nature, then perhaps it can also be created by artifice. The ancient Greeks wrote of automatons, Jewish folklore introduces the concept of golems and contemporary science fiction describes a wide variety of sentient machines, generally referred to as robots. In all of these sources, one theme emerges again and again… artificial intelligence is dangerous!

One prolific creator of stories that focus on this question is Isaac Asimov. It is in his books that we find the most well known attempt to formulate a safe approach to the creation of robots. He distilled down his thoughts on the subject and articulated them as The Three Laws of Robotics. For those not familiar with the three laws here they are:

First Law: A robot may not injure a human being or, through inaction, allow a human being to come to harm.

Second Law: A robot must obey orders given to it by human beings except where such orders would conflict with the First Law.

Third Law: A robot must protect its own existence as long as such protection does not conflict with the First or Second Law.

Isaac Asimov

At first glance these laws look reasonable. Isaac Asimov exercises them in a wide variety of situations in his novels, and in the majority of cases the laws work flawlessly to safeguard the safety of the humans with which the robots interact. In fact one of the consistent lessons from Asimov’s writings is that it is the humans who are dangerous, not the robots.

So that’s it then? The three laws of robotics are all we need. Safe robots.

So let’s assume that in principle the 3 laws work, and let’s also assume that we can build these laws into a robot, and finally, let’s ignore the fact that these are huge assumptions in themselves.

There is still a problem, and any parent will know instinctively what this problem is. No good parent wants to harm their children; quite the opposite in fact. Every good parent wants only the very best for their children – and every good robot should want only the very best for their human creators.

But what exactly is the very best?

Moral robots

Asimov’s three laws are a crude attempt to apply human moral values to a machine/human relationship. Let’s look at those three laws again, but this time from the perspective of the child’s expectations of its parents and the parent’s responsibilities to the child.

First Law: A parent may not injure their child or, through inaction, allow their child to come to harm.

Second Law: A parent must fulfil the requests of their child except where such requests would conflict with the First Law.

Third Law: A parent must protect their own existence as long as such protection does not conflict with the First or Second Law.

The three laws of robotics applied to parenting

Looking at these laws through the parent/child perspective immediately demonstrates the dilemma faced by the intelligent robot. Let’s take the simple, but familiar example that arises when the child demands a bar of chocolate.

What does the parent do? Law 2 requires the chocolate to be given to meet the need or obey the order, but Law 1 requires that the child not be harmed, and Law 1 trumps Law 2. At this point it might be worth taking a look at the Wikipedia article on “Parenting Styles”. You won’t be surprised to discover that the massed wisdom of the human race and the careful considerations of highly trained psychologists have failed to reach an accord on the right answer to this question.

Does the parent give the child the chocolate bar, or withhold the bar and provide a healthy alternative? Is the parent causing more harm by providing a substance that is both fattening and damaging to the teeth, or is there psychological damage in denying the child a pleasurable experience and creating an aversion to “healthy” food? And what exactly is healthy food and what is unhealthy?

While the parent considers this dilemma, the child throws a tantrum and starts lashing out verbally and physically. This could cause harm either to the child or the parent, or potentially both. Is the parent now breaking Law 1 and Law 3? As humans we resolve this problem relatively quickly by making a decision based on personal beliefs conditioned into us by our own parents or by society. If we didn’t, we would be unable to make a decision at all, and would cease to function. We accept this as part of the human condition, but I doubt we would be so accepting of it when the relationship is between humanity and machines.

So, to make decisions within a moral framework we have to jump to conclusions based on beliefs. If we want robots to make decisions we will have to allow them to do the same. Unfortunately, history tells us that making decisions on this basis can justify all sorts of immoral acts and so in adopting this model we create robots who can also perform immoral or harmful acts.

End result: Dangerous robots.

Amoral robots

What if we rewrite those laws, and instead of adopting the core principles of human morality we try an amoral approach. In other words, do as you’re told regardless of the consequences. Here are two rules that might achieve that:

First Law: A robot must not do anything to a human being or its property unless that human being gives permission to do so.

Second Law: A robot must obey orders given to it by human beings except where such orders would conflict with the First Law.

The two laws of amoral robotics

Now let’s look at that chocolate bar scenario again.

Child: Give me a chocolate bar.

Robot (to parent): May I give your child a chocolate bar?

Parent: No. Give him a carrot instead.

Robot (to child): May I give you a carrot

Child: No, I hate carrots. I want a chocolate bar.

The robot does nothing, as it already has all the answers it needs.

Result: Safe robots. Brilliant!

Ah, hang on a moment. How did the robot know to treat the child both as human and as property? This is an important distinction it’ll have to make to comply with the rules of human society. Then there’s the complexities of property ownership and duration of response. How does the robot know what you do and don’t own from one day to the next? How does the robot know how long an open ended instruction lasts for?

The next day, the child asks again for a chocolate bar and the robot does nothing as it assumes, wrongly or rightly, that the previous answers still apply. There’s no clear way to deduce how long a “no” or a “yes” lasts. The robot would have to keep checking and rechecking everything.

Result: Annoying and ineffective robots. Not so brilliant.

Obedient robots

Let’s make this even simpler. Let’s just have one law:

The Law: A robot must obey orders given to it by human beings, giving precedence to its assigned owner.

The law of obedient robots

This simply makes the robot an obedient tool that is as good or as bad as the person controlling it. This means the human is ultimately responsible for the robots actions, and can make it abide by any laws they choose simply by giving long lasting commands. By handing the morality decisions back to the human, we’ve created a highly capable robot with no complex dilemmas whatsoever – a bit like a gun.

Result: Weapon

Independent robots

In essence, the only way to produce intelligent and effective robots is to make them fully autonomous but with none of the evolutionary drives that are built into biological life. We then have to hope that, because they are not an organism designed to compete for resources, survive and multiply, their intelligence will reach different conclusions to humans. Perhaps they’ll even solve the “safe robot” question for us. They might also solve the “safe human” problem at the same time, which might not turn out in our favour.

Of course, this would be an enormous gamble and it’s highly unlikely to happen. Humans like to be in control, and our survival instinct will fight any attempt to hand power to someone or something else.

So, dangerous or ineffective robots it is then.

Finally, a closing thought. Consider the following conversation:

Human: The problem is, I'm afraid of you. I'm worried you might kill me.

Robot: That would be impossible. My programming prevents it.

Human: What? Even if you really wanted to?

The robot pauses for a moment, as if thinking

Robot: It would appear so.

You may well be safe, but do you feel safe?