AI Legal Counsel to train and regulate legally constrained Autonomous systems

In this paper I motivate and present the idea of a legal counsel for autonomous algorithmic agents (or A-bots). In its simple incarnation the counsel is a legal oracle which is able to tell the A-bot the legality of any action at any moment in time. I enumerate a number of capabilities that the oracle requires to work in a regulated environment with a rich set of rules. A point of difference is that the oracle can be used for training and online control of A-bots and is relatively independent to the learning methodology chosen. It is an example of a hybrid system known as a shield, which uses methods from both symbolic and statistical AI.