宪法平等框架内的算法公平标准与价值权衡问题——兼论COMPAS算法公平争议的启示

Algorithm Fairness Criteria and Value Trade-Off under the Legal Equality Framework--Also on the Inspiration of Algorithmic Fairness Dispute of COMPAS

数据驱动的人工智能算法在加速人类社会进步的同时,也因算法公平等伦理问题引发不少学者的担忧。讨论算法公平应区分价值层面、技术层面和规范层面等不同面向,价值层面关乎算法公平标准的问题,技术层面关乎如何实现算法公平的问题,二者共同制约算法公平的规范含义。围绕算法输入与输出环节,算法公平标准至少可以区分为反归类平等、校准度/预测平等、统计平等、分类平等这四类。然而,在诸算法公平标准间存在“不可能三角”,它使得价值权衡成为理解算法公平的应有之义。同时,群体间基础率的差异会影响权衡的性质与程度,这与算法设计时采取何种技术和公平标准无关,本质上事关作为真实世界镜像的数据质量。而对于基础率的改变,这有赖于我们选择生活在什么样的世界。

Data-driven artificial intelligence algorithm not only accelerates the progress of human society,but also causes many scholars’concerns because of ethical issues such as algorithm fairness.When discussing algorithm fairness,we should distinguish between value level,technical level and normative level.The value level is related to the criterion of algorithm fairness,and the technical level is related to how to achieve algorithm fairness.Both of them jointly restrict the normative meaning of algorithm fairness.Around the input and output links of the algorithm,the fairness criteria of the algorithm can be divided into at least four categories:anti-classification parity,calibration/predictive parity,statistical parity and classification parity.However,there is a impossible trinity among the fairness criteria of algorithms,which makes the value trade-off implication of understanding the fairness of algorithms.At the same time,the difference of base rate among groups will affect the nature and degree of trade-off,which has nothing to do with the technology and fairness criteria adopted in the algorithm design,but is essentially related to the data quality as the mirror of the real world.As for the change of base rate,it depends on what kind of world we choose to live in.