A Review of Bias and Fairness in Artificial Intelligence

Abstract

Automating decision systems has led to hidden biases in the use of artificial intelligence (AI). Consequently, explaining these decisions and identifying responsibilities has become a challenge. As a result, a new field of research on algorithmic fairness has emerged. In this area, detecting biases and mitigating them is essential to ensure fair and discrimination-free decisions. This paper contributes with: (1) a categorization of biases and how these are associated with different phases of an AI model’s development (including the data-generation phase); (2) a revision of fairness metrics to audit the data and AI models trained with them (considering agnostic models when focusing on fairness); and, (3) a novel taxonomy of the procedures to mitigate biases in the different phases of an AI model’s development (pre-processing, training, and post-processing) with the addition of transversal actions that help to produce fairer models.