Predictability of AI Decisions

Abstract

Artificial intelligence evolves rapidly and will have a great impact on the society in the future. One important question which still cannot be addressed with satisfaction is whether the decision of an intelligent agent can be predicted. As a consequence of this, the general question arises if such agents can be controllable and future robotic applications can be safe. This chapter shows that unpredictable systems are very common in mathematics and physics although the underlying mathematical structure can be very simple. It also shows that such unpredictability can also emerge for intelligent agents in reinforcement learning, especially for complex tasks with various input parameters. An observer would not be capable to distinguish this unpredictability from a free will of the agent. This raises ethical questions and safety issues which are briefly presented.