Abstract

Due to their self-learning and evolutionary character, AI (Artificial Intelligence) systems are more prone to unintended consequences and more difficult to control than traditional sociotechnical systems. To deal with this, machine ethicists have proposed to build moral (reasoning) capacities into AI systems by designing artificial moral agents. I argue that this may well lead to more, rather than less, unintended consequences and may decrease, rather than increase, human control over such systems. Instead, I suggest, we should bring AI systems under meaningful human control by formulating a number of meta-values for their evolution. Amongst others, this requires responsible experimentation with AI systems, which may neither guarantee full control nor the prevention of all undesirable consequences, but nevertheless ensures that AI systems, and their evolution, do not get out of control.