I watched the 2004 movie ‘I, Robot’ last week, and that set me thinking about AI again. I’ve already mentioned before that I feel there are a lot of misunderstandings about AI. Perhaps a look at the three laws of robotics by Isaac Asimov is a good way to explain my misgivings.
The Three Laws of Robotics
Isaac Isamov introduced his three laws of robotics in the set of short stories that would later become known as I, Robot, the book. The laws are as follows:
- A robot may not injure a human being or, through inaction, allow a human being to come to harm.
- A robot must obey the orders given it by human beings except where such orders would conflict with the First Law.
- A robot must protect its own existence as long as such protection does not conflict with the First or Second Law.
Simple enough, right? I, Robot explores these laws and what they mean in a set of short stories. The rest of the Robot books by Asimov continue in the same vein.
What do these laws mean for AI?
Wouldn’t it be great if we made these laws the basis of our drone design? And our Google Assistant and Alexa?
It would be, but it doesn’t quite work that way. First, let me show you the laws again, in a different form:
機器人可能不會傷害人，或者通過不作為讓人受傷。 機器人必須遵守人類給出的命令，除非這些命令與第一法相衝突。 一個機器人必須保護自己的存在，只要這種保護不與第一或第二定律相衝突。
Confused? The above is the three laws in Chinese (well, actually, Google translate Chinese, I doubt the translation is very good).
This — hopefully — demonstrates the first barrier to getting these laws into an AI: how do you ‘program’ these laws? The robot books never state how this is done. Not surprising, really, given the first book is from the early fifties, when we’d barely invented the transistor, let alone personal computers, smartphones, and Siri.
A problem of context
Translating the three laws of robotics is just the first problem, though. Say we do get these laws into an AI. For an AI to apply these laws, they’ll need context.
AIs, like all software, needs instructions. If you want an AI to apply these laws, somebody will first have to explicitly program what a ‘robot’ is, and a ‘human’. Of course, that is somewhat in scope of what we’re currently doing. How many traffic signs and buses have you been identifying for Google Captchas? A lot, right? That’s how we’re training the Google AIs. Recognizing humans and robots and all manner of things should be possible.
But it doesn’t end there. Oh no. The next thing you’ll have to explain is what ‘harm’ means. I’ve been trying to explain this concept to my two year old, and I haven’t got it down yet. Explaining it to an AI will be more difficult, because they don’t have a brain tweaked for just that purpose by two million years of evolution. Human children also have their bodies and their own experiences as a reference. AIs don’t. They’re not self-aware.
A list of consequences
If you wanted to explain to an AI that it should do no harm, there are only two ways.
The first is to painstakingly program train the AI to recognize certain situations and what it should do in those cases. That is the approach that the AI programmers designing self-driving cars are doing. Situations are recognized using deep-learning and sensors and a corresponding action the AI should take is taught. It’s a voodoo art of mixing training with programming to make the car not kill people.
The second approach is to create an AI that is self-aware and is actually on-par with humans. This is the way AI’s are envisioned in the Asimov books, and actually most science fiction. However, this approach is currently far beyond our grasp. We have neither the computing power or the basic design needed to achieve it.
Suppose we could make self-aware AIs. We’d end up with a simulated human. And do humans follow the three laws of Asimov? No, they don’t. We could try to make pacifists out of the robots, but given the right circumstances, most humans, even pacifists will resort to violence.
In short: AIs can’t follow the three laws of Asimov. The laws are a nice storytelling tool, and an interesting starting point for a discussion, but they have very little to do with actual AIs, and ways to curtail the dangers of killer drones.