CAREER: Achieving Explainable Artificial Intelligence (AI) through Human-AI Interaction

解释和证明人工智能(AI)所做决策的能力对于技术驱动的创新和广泛的社会采用AI至关重要。解释人工智能不仅意味着描述人工智能是如何做出决定的，以及这些决定是基于什么，而且还意味着人工智能擅长做出哪些决定，而不是哪些决定。这样的解释使人们能够决定他们是否应该信任人工智能。随着基于人工智能的技术被部署到许多高风险的决策场景中，例如在政府、司法、医疗保健和许多行业中，使人工智能变得可解释变得越来越重要。然而，大多数现有的解释人工智能的方法侧重于精通计算机科学的人工智能创建者，而不是最终用户，例如在决策中使用人工智能的其他领域专家、监管人工智能的政策制定者，以及与基于人工智能的系统交互的消费者。向最终用户解释人工智能可以帮助他们了解人工智能正在做出什么决定以及为什么，以及这些决定如何影响他们和更广泛的社会。这个项目的目标是通过精心设计的与模型和领域无关的人类-人工智能交互向最终用户提供人工智能解释，而不牺牲人工智能的性能，从而使人工智能解释民主化。这个项目为在感觉制造理论中寻找和推导人工智能解释的过程奠定了基础。在这一过程中，最终用户分两步更新他们关于人工智能如何工作的心理模型：1)寻找不同人工智能输入和输出之间关系的证据，2)为他们获得的证据赋予意义，以形成和测试关于可能的人工智能解释的假设。该项目将贡献有关经验验证机制的新科学知识，这些机制将向最终用户提供关于人工智能的有意义、全面和准确的解释。该项目的发现将扩大现有方法和工具的广度，使人工智能测试、最终用户倡导、公共教育和关于人工智能的调查性新闻成为可能。该项目还将生成设计指南，帮助广大最终用户更多地获得未来基于人工智能的技术。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

The ability to explain and justify the decisions made by Artificial Intelligence (AI) is of critical importance to technology-driven innovation and broad societal adoption of AI. To explain AI means to not only describe how the AI makes its decisions and what it bases those decisions on, but also which decisions the AI is good at making and which ones it is not. Such explanations enable people to decide whether they should trust AI or not. Making AI explainable is increasingly important as AI-based technology gets deployed into many high-stakes decision-making scenarios, such as in government, judiciary, healthcare, and across many industries. However, most existing approaches to explain AI focus on computer science-savvy AI creators rather than end-users, such as other domain experts using AI in decision-making, policy makers regulating AI, and consumers interacting with AI-based systems. Explaining AI to end-users could help them understand what decisions the AI is making and why, and how those decisions impact them and society more broadly.The goal of this project is to democratize AI explanations by delivering them to end-users through carefully designed model- and domain-agnostic human-AI interactions without sacrificing the performance of AI. This project grounds the process of seeking and deriving AI explanations in sensemaking theory. In this, the end-user updates their mental model about how the AI works in two steps: 1) foraging for evidence of the relationship between different AI inputs and outputs, and 2) assigning meaning to the evidence they obtain to form and test hypotheses about possible AI explanations. The project will contribute new scientific knowledge about empirically-validated mechanisms that deliver meaningful, comprehensive, and accurate explanations about the AI to end-users. The findings from this project will expand the breadth of existing methods and tools that enable AI testing, end-user advocacy, public education, and investigative journalism about AI. This project will also generate design guidelines that will help increase access to future AI-based technology for a broad audience of end-users.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.