Unbiased AI for a Sovereign Digital Future: A Bias Detection Framework

Abstract

The recent advancements in Artificial Intelligence (AI) have opened avenues on all fronts of life that were never even imagined before. The possibilities of incorporating AI tools to various domains of the computing and non-computing world to enhance efficiency, performance and reliability are skyrocketing in the current era of technology. While computing power is a crucial factor driving the development of these AI tools, the role of data is equally significant. One of the key reasons for AI tools to perform the way they perform is the presence of an enormous amount of data they can work with. However, the reliance on vast datasets also raises concerns about control over data and digital sovereignty, especially when such data impacts critical decision-making processes in recruitment, policy making, loan approvals, and beyond. If a feature in the data related to intersectionality, e.g., gender, race, cultural background, etc. dictates the outcome, the data is most probably biased. The bias of data can lead to injustice, inequality and unfairness and hence it is extremely important to tackle it. Ensuring that the data used is ethically managed, especially in line with national and regional data sovereignty regulations, is an integral aspect of mitigating these issues. The first step in the process is to identify the bias in data. This paper explores a methodology for detecting bias in data, based on a general AI-based framework that can be applied across various domains. The paper goes into the details of evaluating the identified bias for the gender feature and explains how this feature influences the outcome of a machine learning (ML) model.