Building Ethics into Artificial Intelligence

• Han Yu, Zhiqi Shen, Chunyan Miao, Cyril Leung, Victor R. Lesser, Qiang Yang

Abstract

• taxonomy which divides the field into four areas: 1) exploring ethical dilemmas; 2) individual ethical decision frameworks; 3) collective ethical decision frameworks; and 4) ethics in human-AI interactions

Types

• Consequentialist ethics
• Utilitarian ethics
• Deontological ethics
• Virtue ethics
• Ethical dilemmas

Exploring Ethical Dilemmas

• explore the ethical dilemmas in the target application scenarios [Anderson and Anderson, 2014]
• GenEth
• Moral Machine project

Individual Ethical Decision Frameworks

• AI research community largely agrees that generalized frameworks are preferred over ad-hoc rules
• if updates are provided by people, some review mechanisms should be put in place to prevent abuse
• moral decision-making by humans not only involves utilitarian considerations, but also moral rules.
• Such rules often involve protected values (a.k.a. sacred values)
• MoralDM
• Belief-Desire-Intention
• blind ethical judgement
• partially informed ethical judgement
• fully informed ethical judgement
• Moral decision making frameworks for artificial intelligence
• Preferences and ethical principles in decision making
• A declarative modular framework for representing and applying ethical principles.
• A low-cost ethics shaping approach for designing reinforcement learning agents
• Even angels need the rules AI, roboethics, and the law
• Norms as a basis for governing sociotechnical systems
• Embedding ethical principles in collective decision support systems
• A voting-based system for ethical decision making
• swap-dominance
• satisfying consequentialist ethics Ethics in Human-AI Interactions Belmont Report
• [Luckin, 2017; Yu et al., 2017b]
• 1. people’s personal autonomy should not be violated (they should be able to maintain their free will when interacting with the technology); 2) benefits brought
• about by the technology should outweigh risks; and 3) the benefits
• and risks should be distributed fairly among the users (people should not be discriminated based on their personal backgrounds such as race, gender and religion)
• persuasion agents
• [Kang et al., 2015; Rosenfeld and Kraus, 2016]
• [Stock et al., 2016]
• large-scale study to investigate human perceptions on the ethics of persuasion by an AI agent
• trolley scenario
• authors tested three persuasive strategies: 1) appealing to the participants emotionally; 2) presenting the participants with utilitarian arguments; and 3) lying
• participants hold a strong preconceived negative attitude towards the persuasion agent, and argumentation-based and lying-based persuasion strategies work better than emotional persuasion strategies
• did not show significant variation across genders or cultures
• adoption of persuasion strategies should take into account differences in individual personality, ethical attitude and expertise in the given domain.
• Coping Theory
• Argumentation-based explainable AI
• [Fan and Toni, 2015; Langley et al., 2017] well suited to the consequentialist ethics
• depending on how the explanations are used, researchers need to strike a balance on the level of details to be included
• Full transparency may be too overwhelming if the objective is to persuade a user to follow a time-critical recommendation
• useful as a mechanism to trace the AI decision process afterwards not enough transparency may hamper users’ trust in the AI