Abstract. Policies are being increasingly used for automated system management and controlling the behavior of complex systems. The use of policies allows administrators to modify system behavior without changing source code or requiring the consent or cooperation of the components being governed. Early approaches to policy representation have been restrictive in many ways. However semantically-rich policy representations can reduce human error, simplify policy analysis, reduce policy conflicts, and facilitate interoperability. In this paper, we compare three approaches to policy representation, reasoning, and enforcement. We highlight similarities and differences between Ponder, KAoS, and Rei, and sketch out some general criteria and properties for more adequate approaches to policy semantics in the future.