This is my article that was published in the Technology Section of our school magazine, "The Reason", 2022 edition.

You’ve seen the movies, right? Think about all those bloodthirsty robots controlled by artificial intelligence (AI). What do they all end up doing? That’s right: betray their programming and obliterate all human life in sight. Movies and media love to show AI betraying its original objective in order to kill humans. However, would this really happen in real life?

First, let’s talk about neural networks. Think of them like brains that help the AI determine a good action to perform, known as its output. What dictates the output is whatever information the AI is given, known as the input. Through some complicated maths, the AI moulds the input into what it thinks is an appropriate output. How good the output is depends on how well the neural network was trained. You yourself wouldn’t expect to do something well without practice first, right? In short, neural networks control the AI, but they need to be trained first. The AI is trained by its creator, usually a programmer. They set an objective for the AI, and it tries to learn how to best meet that objective through long and arduous training. However, the AI can only do that objective, and nothing else. You wouldn’t expect a household Roomba to suddenly pull out a knife and start chasing down the nearest human, would you? Even if the AI could in theory pull of such a stunt, its objective and programming prevents it from doing so.

Now an obvious question may arise: if the objective dictates what an AI does, can the programmer purposefully make it kill? The answer is yes. If a malicious programmer decides that they want to watch the world burn, they could train an AI to go on a slaughtering rampage. Obviously, the programmer would also have to build the AI with the physical capabilities to do such a thing. But it should be obvious by now that no AI will suddenly decide by itself that it doesn’t like humans and will begin to kill them. A normal objective, like cleaning a room, doesn’t let it do such a thing.

Of course, a programmer may be blissfully unaware that an objective they gave an AI is in fact deadly. In this case, this would still be the programmer’s fault. Why? Because of a little thing called “Thinking logically”. In programming, to think logically is to think of every possible outcome of the program and anything users can do with it. The purpose is to safeguard against dangerous or bad outcomes. We already know AI can’t change its objective on a dime, so if an AI becomes evil, the programmer is likely at fault.

Now I hear you ask “what about a self-evolving AI? Surely it could give itself a dangerous objective!” It’s true that a self-evolving AI can be programmed to decide its own objective. However even here, responsibility falls on the programmer. The whole point of an AI reaching for an objective is that it wants to maximise its rewards. These “rewards” are arbitrary numbers that bore us humans but are like ice-cream to the AI. The programmer is the one who decides which actions give the most rewards. So when an AI changes its objective, it will always do so to get more rewards. As a result, if the AI changes its objective to “kill all humans”, it did so because the programmer decided that this action gives it the most rewards.

No matter how you view things, the AI isn’t capable of making itself dangerous. A Roomba won’t go from vacuuming your carpet to blasting your head off. If an AI became dangerous, the only person to blame would be the programmer. Movies, though, like to pretend an AI can become dangerous by itself because saying “it’s the programmer’s fault” doesn’t make a good story.