A Person-Centred Approach to Understanding Trust in Moral Machines

人工智能（AI）越来越多地用于执行具有道德维度的任务，例如优先考虑稀缺的医疗资源。关于人工智能的道德问题，我们应该如何编程道德人工智能，以及我们需要优先考虑哪些道德价值观的争论非常激烈。但机器道德既关乎人类道德心理学，也关乎构建人工代理的哲学和实践问题。为了获得人工智能的好处，利益相关者需要愿意使用、采用和依赖这些系统：他们必须信任人工智能代理。TRUST-AI利用心理学和哲学来探索人类如何以及何时信任充当“道德机器”的人工智能代理。这个为期五年的项目借鉴了经典的信任模型和道德心理学关于道德领域信任复杂性的最新理论工作，探讨了1）预测信任的人工智能代理的特征; 2）使我们或多或少信任人工智能代理的个体差异; 3）我们更有可能将道德决策“外包”给人工智能代理的情况;以及4）这些发现应该如何用于设计值得我们信任的人工智能代理。我的方法是方法论上的多元化，包括定性分析和自然语言处理，行为经济游戏和亲身行为实验。我开发了一个定制的数据收集平台，该平台将在全球范围内以至少3种不同的语言进行在线实验，并在9个不同的国家进行跨文化研究，以及一系列实验，估计有29，000名参与者参与了四个工作计划和20项研究。这些发现将帮助我们了解如何，何时，以及为什么人们信任AI代理人，这对研究道德机器中信任的前因和后果具有重要的理论和方法论意义。

Artificial intelligence (AI) is increasingly used to perform tasks with a moral dimension, such as prioritisingscarce medical resources. Debates rage about the ethical issues of AI, how we should programme ethical AI,and which ethical values we need to prioritize. But machine morality is as much about human moralpsychology as it is about the philosophical and practical issues of building artificial agents. To reap thebenefits of AI, stakeholders need to be willing to use, adopt, and rely on these systems: they must trust in theAI agents.TRUST-AI draws on psychology and philosophy to explore how and when humans trust AI agents that act as'moral machines'. Drawing from classic models of trust and recent theoretical work from moral psychology onthe complexity of trust in the moral domain, this five-year project explores 1) The characteristics of AI agentsthat predict trust; 2) the individual differences that make us more or less likely to trust AI agents; 3) thesituations where we are more likely to 'outsource' moral decisions to AI agents; and 4) how these findingsshould be used to design AI agents that warrant our trust.My approach is methodologically pluralistic and includes qualitative analysis and natural language processing,behavioural economic games, and in-person behavioural experiments. I develop a customised data collectionplatform that will run online experiments globally in at least 3 different languages, as well as cross-culturalstudies in 9 different countries and a range of experiments with an estimated 29,000 participants across fourwork programmes and 20 studies.These findings will help us understand the how, when, and why people trust AI agents with importanttheoretical and methodological implications for research on the antecedents and consequences of trust in moralmachines.