Probing formal/informal misalignment with the loophole task

摘要

Any autonomous agent deployed with some representation of rules to follow will face scenarios where the applicability of its given rules are not clear. In such scenarios, a malicious agent might successfully argue that some action which clearly goes against the spirit of the rules is allowed, under a strict interpretation of the rules. We argue that the task of finding such actions, which we call the loophole task, must be solved to some degree by an autonomous ethical agent, and thus is important for robot ethical standards. Currently, no artificially intelligent system comes close to solving the loophole task. We define this task, by characterizing it as exploiting a misalignment between informal and formal representational systems, and discuss our preliminary work towards creating an automated reasoner capable of solving it.