CAREER: Redesigning the Human-AI Interaction Paradigm for Improving AI-Assisted Decision Making

基于人工智能的决策辅助工具越来越多地用于支持金融、执法、教育和网络安全等领域的工作人员。然而，人类与人工智能的协作决策性能往往低于预期。人与人工智能之间缺乏足够和有效的参与是导致这些令人失望的结果的根本原因。通常情况下，决策者不会以其完全的认知能力参与人工智能系统提供的建议，或者，当他们这样做时，他们会以导致对系统不适当的信任和依赖的方式参与其中-依赖不足和过度依赖都是一个问题。在目前的实践中，基于人工智能的决策辅助工具不足以鼓励人类参与，也没有考虑到人类的参与行为。本项目旨在为人与人工智能的交互创造全新的设计，其中人工智能是以执行为导向的，以改善人与人工智能在决策过程中的协作。本项目开发并评估了三种新颖的人与人工智能交互范例：（1）人与反思型人工智能交互，其中人工智能旨在通过引导决策者对决策过程进行批判性反思来提高决策者的参与水平;（2）人类自适应人工智能交互，其中人工智能辅助将以个性化的方式呈现，以推动决策者适当地参与人工智能建议;以及（3）人类行为感知人工智能交互，其中人工智能经过训练以预测决策者的参与行为并优化人类人工智能团队的表现。该项目将为如何将决策者的认知因素（包括他们的能力和局限性）纳入基于人工智能的决策辅助工具的设计中提供新的科学知识，以最好地支持决策者并提高人工智能团队的绩效。通过对新的人机交互模式对可用性、用户体验和人机决策性能的影响进行全面评估，将揭示不同方法的优缺点，并为其中的选择提供信息。该奖项反映了NSF的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

AI-based decision aids are increasingly used to support people working in domains such as finance, law enforcement, education, and cybersecurity. However, human-AI collaborative decision-making performance is often lower than expected. A lack of sufficient and effective engagement between the person and the AI is a fundamental contributor to these disappointing outcomes. Often, decision makers do not engage with the advice provided by the AI system with their full cognitive capacity or, when they do, they engage with it in ways that lead to inappropriate trust and reliance on the system---both under-reliance and over-reliance are a problem. In current practice, AI-based decision aids do not sufficiently encourage human engagement, nor do they take human engagement behaviors into account. This project aims to create radical new designs for human-AI interaction in which AI is engagement-oriented to improve human-AI collaborations in decision-making.This project develops and evaluates three novel human-AI interaction paradigms: (1) Human-Reflective AI interaction, in which AI is designed to increase the decision makers' engagement level by guiding them to critically reflect on the decision process; (2) Human-Adaptive AI interaction, in which AI assistance will be presented in a personalized manner to nudge decision-makers into appropriately engaging with the AI recommendations; and (3) Human-Behavior-aware AI interaction, in which AI is trained to anticipate decision-makers' engagement behavior and optimize for the human-AI team performance. This project will contribute new scientific knowledge about ways to incorporate cognitive factors of decision-makers, including their capabilities and limitations, into the designs of AI-based decision aids to best support decision-makers and increase AI-human team performance. Comprehensive evaluations on the impacts of new human-AI interaction paradigms on usability, user experience, and the human-AI decision-making performance will reveal the strengths and weaknesses of different approaches and inform the selection among them.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.