Socially Acceptable AI and Fairness Trade-offs in Predictive Analytics

Fairness and non-discrimination are basic requirements for socially acceptable implementations of AI, as these are basic values of our society. However, the relation between statistical fairness concepts, the fairness perception of human stakeholders, and principles discussed in philosophical ethics is not well understood. The objective of our project is to develop a methodology to facilitate the development of fairness-by-design approaches for AI-based decision-making systems. The core of this methodology is the “Fairness Lab”, an IT environment for understanding, explaining and visualizing the fairness implications of a ML-based decision system. It will help companies to build socially accepted and ethically justifiable AI applications, educate fairness to students and developers, and support informed political decisions on regulating AI-based decision making. Conceptually, we integrate statistical approaches from computer science and philosophical theories of justice and discrimination into interdisciplinary theories of predictive fairness. With empirical research, we study the fairness perception of different stakeholders for aligning the theoretical approach. The utility of the Fairness Lab as a tool for helping to create “fairer” applications will be assessed in the context of participatory design. With respect to application areas, we focus on employment and education. Our project makes a significant contribution to the understanding of fairness in the digital transformation and to promoting improved conditions for the deployment of fair and socially accepted AI.

Machine Learning; Participatory Design; Ethics; ML-based Decision-Making; Artificial Intelligence; Fairness; Human Ressource Management