AI Ethics: First rule of the AI control problem: don’t talk about the AI control problem

Abstract:

How do you control a superintelligent artificial being given the chance that its goals or actions conflict with human values? This concern – the AI control problem – has remained a central challenge for research in AI safety and alignment.

In this talk, I will present and defend two arguments that independently recommend the following strategy for those who worry about the AI control problem: don’t talk about it.

The first is argument from counter-productivity, which states that unless kept secret, one cannot ensure that efforts to solve the control problem will not be used by a misaligned AGI to counter the same efforts; the second is argument from suspicion, stating that unless kept secret, work on the control problem will serve to make humanity appear threatening to an AGI, which increases the risk that the AGI sees humanity as a threat.

Bio:

Jakob Stenseke is a PhD candidate in Practical Philosophy at Lund University. His interdisciplinary research project – titled How to build nice robots: ethics from theory to machine implementation – broadly explores the possibilities and challenges for creating artificial moral agents, that is, AI systems capable of acting in reference to what is morally good or bad.

This includes the theoretical possibility (how and to what extent can artificial agents be moral?), normative desirability (why and in what way do we want or need ”ethical” machines?), and technical engineering (how do you actually build ethical AI?) of artificial moral agents.