FAI: Auditing and Ensuring Fairness in Hard-to-Identify Settings

用于高风险决策的人工智能（AI）系统的普及引发了迫切的道德和法律要求，以避免歧视并保证受保护阶层的公平。示例应用领域包括信用决策、个性化医疗、有针对性的政策制定以及判决和保释设置。在这些环境中，公平性通常通过决策对不同群体的不同影响来量化。然而，可用数据的根本限制使得审计差异和消除差异变得困难或不可能。例如，在信贷和保险索赔数据中，种族等敏感标签都缺失，人们永远不知道审前被拘留的被告是否会逃跑。在这两种情况下，这种缺失都会导致无法识别差异。这些识别问题不仅出现在几乎所有关注公平性的应用中，而且还打破了公平人工智能的大多数现有方法，并在公平人工智能的理论和实践之间造成了紧迫的差距。为了解决这一差距，该项目开发了一种强大的理论和方法，用于在公平指标难以或不可能确定的环境中评估和确保公平性。具体来说，该项目将开发：（a）公平性评估方法，在面临基本识别限制的情况下能够可靠地支持可信结论； (b) 即使公平性无法衡量，学习算法也能在设计阶段强有力地执行公平性。一个关键方法是认识到识别的局限性，并通过考虑算法在实践中可能或可能不会引起的可能差异范围来解决它。该项目确定了几个离散的环境，其中公平性很难或不可能衡量，并且需要单独处理：不可观察的受保护类别成员身份、个性化干预、非二元算法输出、决策支持算法和不可忽略的选择性标签。通过解决这些问题，该项目将改变对真实实际环境中差异的评估，其中公平性实际上在统计上很难确定，并相应地影响我们如何确保人工智能系统公平并平等地惠及每个人。这种影响也将通过与从业者的直接合作来直接实现。该奖项反映了 NSF 的法定使命，并通过使用基金会的智力价值和更广泛的影响审查标准进行评估，被认为值得支持。

The spread of artificial intelligence (AI) systems for high-stakes decision making gives rise to an urgent ethical and legal imperative to avoid discrimination and guarantee fairness with respect to protected classes. Example application domains include credit decisioning, personalized medicine, targeted policymaking, and sentence and bail setting. In these settings, fairness is often quantified by the decisions' disparate impacts on different groups. However, fundamental limits in the data available make both auditing disparities and eliminating them difficult or impossible. For instance, both in credit and insurance claims data, sensitive labels such as race are missing, and one never knows if a defendant detained pretrial would have fled. In both cases this missingness renders disparities unidentifiable. These identification issues not only arise in nearly every application where fairness is a concern, they also break most existing methods for fair AI and create an urgent gap between the theory and practice of fair AI.Addressing this gap, this project develops a robust theory and methodology for assessing and ensuring fairness in settings where fairness metrics are hard or impossible to pin down. Specifically, the project will develop: (a) fairness assessment methods that can reliably support credible conclusions in the face of fundamental identification limitations; (b) learning algorithms that robustly enforce fairness at the design stage even if fairness is unmeasurable. A key approach is recognizing the limits of identification and addressing it by considering the possible ranges of disparities that an algorithm may and may not induce in practice. The project identifies several discrete settings where fairness is hard or impossible to measure and that require separate treatment: unobserved protected class membership, personalized interventions, non-binary algorithmic outputs, decision-support algorithms, and non-ignorable selective labeling. By tackling these, the project stands to transform the assessment of disparity in real practical settings, where fairness is actually statistically difficult to pin down, and correspondingly impact how we ensure AI systems are fair and benefit everyone equally. This impact will also be directly achieved through direct collaborations with practitioners.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.