Safe AI vs Safe Use of AI

Recent advances in the development of AI has, predictably, led to a massive increase in research on how to make AI safe. While this is a valid and important quest, this paper emphasizes the difference between requiring an AI to be safe, and that AI is used safely. Using AI to control a system with potentially serious safety risks, even with the goal of making the system safer through the use of AI, doesn't necessarily imply that the AI itself must be safe. Focusing too much on making AI safe could result in a two-fold problem: 1) We cannot fully utilize the potential of AI; (because) 2) We struggle with demonstrating adequate safety for the AI. It is our opinion that problem 2 often can be avoided, and problem 1 alleviated because problem 2 is not relevant. This, however, requires a somewhat different way of thinking about AI in safety-critical systems than seems often to be the case. In this paper we discuss these problems, illustrate them with examples, and show that there is already much knowledge on how to achieve documented safety of AI-enabled systems without having to provide safety assurance for the AI itself.

Keywords: AI, Safety, Assurance.