Fairness in Design: A Framework for Facilitating Ethical Artificial Intelligence Designs

Abstract

As Artificial Intelligence (AI) and Digital Transformation (DT) technologies become increasingly ubiquitous in modern society, the flaws in their designs are starting to attract attention. AI models have been shown to be susceptible to biases in the training data, especially against underrepresented groups. Although an increasing call for AI solution designers to take fairness into account, the field lacks a design methodology to help AI design teams of members from different backgrounds brainstorm and surface potential fairness issues during the design stage. To address this problem, we propose the Fairness in Design (FID) framework to help AI software designers surface and explore complex fairness-related issues, that otherwise can be overlooked. We explore literature in the field of fairness in AI to narrow down the field into ten major fairness principles, which assist designers in brainstorming around metrics and guide thinking processes about fairness. FID facilitates discussions among design team members, through a game-like approach that is based on a set of prompt cards, to identify and discuss potential concerns from the perspective of various stakeholders. Extensive user studies show that FID is effective at assisting participants in making better decisions about fairness, especially complex issues that involve algorithmic decisions. It has also been found to decrease the barrier of entry for software teams, in terms of the pre-requisite knowledge about fairness, to address fairness issues so that they can make more appropriate related design decisions. The FID methodological framework contributes a novel toolkit to aid in the design and conception process of AI systems, decrease barriers to entry, and assist critical thinking around complex issues surrounding algorithmic systems. The framework is integrated into a step-by-step card game for AI system designers to employ during the design and conception stage of the life-cycle process. FID is a unique decision support framework for software teams interested to create fairness-aware AI solutions.