Over the past couple of years, I’ve witnessed many contentious debates on the issue of whether it is possible to teach morality to AI machines. Wendel Wallach and Collin Allen argue that it is possible to design AI machines to develop morals using the bottom-up approach. He explains:

“In bottom-up approaches to machine morality, the emphasis is placed on creating an environment where an agent explores courses of action and learns and is rewarded for behavior that is morally praiseworthy. There are various models for bottom-up acquisition of moral capabilities. Childhood development provides one model.” (Wallach & Allen, Moral Machines)

This approach is akin to the old-fashioned way of how we teach children to grow up to become good and morally responsible adults. We give them rewards as positive reinforcement for good behaviour and punishments as negative reinforcement for bad behaviour. It is not a perfect approach, that’s for sure. There are lots of mistakes we parents make along the way. But it works through a continuous cycle of trial-and-error, anxiety and exhilaration.

The bottom-up approach to machine morality refers to the type of reinforcement learning used in AI machines that is powerful and effective. It provides continuous feedback on the ideal behaviour that we want from AI machines as they work their way through a specific environment. We put an AI machine in an environment that provides it with various alternative courses of action regarding an ethical dilemma. The AI machine evaluates the alternatives based on what it has learned from past experiences about what is right and what is wrong. It will then choose to pursue a specific course of action that it deems the best among the various moral choices. We humans will then give it feedback as to the wisdom of its choice. If it was the morally right choice, we give the AI machine a reward (high score) that will encourage it to make the same choice the next time it is faced with that similar type of ethical dilemma. If it chooses the less morally right or the immoral option, we will give the AI machine some sort of punishment (low score) that will deter it from choosing that specific course of action again.

If we can teach AI machines how to be moral, a thornier issue emerges: Whose morals will be used to teach AI machines what is right versus what is wrong? And so, the debate on moral machines rages on.