It can be said that none of yet proposed methods for achieving artificial ethical reasoning is realistic, i.e. working outside very limited environments and scenarios. Whichever method one chooses, it will not work in various real world situations because it would be very cost-inefficient to provide ethical knowledge for every possible situation. We believe that an autonomous moral agent should utilize existing resources to make a decision or leave it to humans. Inverse reinforcement learning has gathered interest as a possible solution to acquiring knowledge of human values. However, there are two basic difficulties with using a human expert as the source of exemplary behavior. First derives from the fact that it is rather questionable if one person or a few people (even qualified ethicists) can be trusted as safe role models. We propose an approach which requires referring the maximal number of (currently available) possible similar situations to be analyzed, and a majority decision-based "common sense" model is used. The second problem lies in human beings' difficulties with living up to their words, surrendering to primed urges and cognitive biases, and in consequence, breaking moral rules. Our proposed solution is to use not behaviors but humans' declared reactions to acts of others in order to help a machine determine what is positive and what is negative feedback. In this paper we discuss how the third person's opinion could be utilized via means of machine reading and affect recognition to model a safe moral agent and discuss how universal values might be discovered. We also present a simple web-mining system that achieved 85% agreement in moral judgement with human subjects.
展开▼
