A model of fair and explainable artificial intelligence

Artificial intelligence (AI) is making more and more algorithmic decisions for humans. However, the “intelligence” function in AI relies heavily on learning technologies, which suffer two major flaws leading to legal and technical challenges: potentially discriminative/biased decisions, and unable to explain why and how a machine makes such decisions. Therefore, building a fair and explainable AI model is important and urgent. This article presents a novel theory-based individual-level dynamic learning method that performs learning using data of an individual subject without employing others’ information, and identifies causal mechanism from unobserved data generating process that each subject exhibits. Thus, data selection bias is avoided and a fair and interpretable decision is achieved. We empirically test our method using a real-world dataset on risk assessment for lending decisions. Our results show that the proposed method outperforms conventional learning methods in terms of fairness in treating data subjects, decision accuracy and interpretability.