[论文解读] Measuring Perceptions of Fairness in AI Systems: The Effects of Infra-marginality

Schrasing Tong, Minseok Jung+2 more|arXiv (Cornell University)|Mar 6, 2026

Ethics and Social Impacts of AI被引用 0

一句话总结

论文报告了一项用户研究，显示在低边际性（infra-marginality）下的公平判断取决于群体之间的分布差异与数据可用性，而不仅仅是平等，从而挑战仅仅以平等来衡量的公平性指标。

ABSTRACT

Differences in data distributions between demographic groups, known as the problem of infra-marginality, complicate how people evaluate fairness in machine learning models. We present a user study with 85 participants in a hypothetical medical decision-making scenario to examine two treatments: group-specific model performance and training data availability. Our results show that participants did not equate fairness with simple statistical parity. When group-specific performances were equal or unavailable, participants preferred models that produced equal outcomes; when performances differed, especially in ways consistent with data imbalances, they judged models that preserved those differences as more fair. These findings highlight that fairness judgments are shaped not only by outcomes, but also by beliefs about the causes of disparities. We discuss implications for popular group fairness definitions and system design, arguing that accounting for distributional context is critical to aligning algorithmic fairness metrics with human expectations in real-world applications.

研究动机与目标

• 研究低边际性如何影响人类对AI预测公平性的判断。
• 检验群体之间的分布差异与训练数据可用性是否影响感知公平性。
• 评估最终用户如何解读群体特定的表现及其与在医疗情景中公平决策的关系。

提出的方法

• 通过在线 Qualtrics 用户研究，85 名参与者在不同群体特异性表现下评估三种候选模型。
• 两项处理因素：群体特异性表现（7 个实例）和数据可用性（4 个实例）。
• 参与者对三种编码不同公平假设（平等 vs. 保留差异）的模型选项，在 7 点李克特量表上对公平性进行评分。
• 以准确率作为分布差异的代理变量，反映两群体癌症预测情景中的infra-marginality。
• 通过重复情景检查与试验测试确保理解与数据有效性；分析使用独立样本 t 检验。

Figure 1. Mean and standard errors of fairness perceptions on the 3 Options for the group-specific performance treatment. Group-specific accuracy denoted as (Race A and Race B) for the 7 subplots are NA/NA, 90/90, 70/70, 95/85, 75/65, 85/95, and 65/75. * signifies p $<$ 0.05, ** signifies p $<$ 0.01

实验结果

研究问题

• RQ1 在两群体之间存在分布差异的情况下，用户如何评价模型的公平性？
• RQ2 相对的训练数据可用性如何影响公平判断与对这些差异的解读？
• RQ3 用户是否在infra-marginality 下偏好基于平等的公平性还是保留观察到的群体差异的模型？

主要发现

• 参与者对公平性的判断会随群体特异性表现而变化：当表现相等或未知时偏好平等；当存在可见差异时倾向于保留差异（infra-marginality）。
• 数据可用性会影响公平推理：对于表现较高的群体拥有更多数据并不自动被视为公平；将差异归因于任务难度更易被接受。
• 公平判断相对基线而定；锚定于原始的群体特异性表现会影响对新模型的评估。
• 基于平等的指标（如等化误差率）在当差异反映分布差异时，可能与人们对公平的感知相冲突。
• 结果表明公平框架应纳入分布上下文，而非不惜一切代价强制平等。

Figure 2. Mean and standard errors of fairness perceptions on the 3 Options when Race A $>$ Race B in group-specific performance. Subplots show data of Race A relative to Race B: no info, 3x, 20x, and 1x respectively. * signifies p $<$ 0.05, ** signifies p $<$ 0.01, and *** signifies p $<$ 0.001.