The operation and impacts of an AI system should be transparent, traceable, auditable and generally explainable to a degree appropriate to its use and potential risk profile so outcomes can be understood and challenged, particularly where they relate to people.