Design for fairness in AI

Abstract

Artificial intelligence (AI) is an emerging field which unleashes massive new (business) opportunities. The potential growth and broad application of the AI technology has great economic benefits however also severe societal implications. Simultaneously, ethical challenges arise with its development. Questions of values and ethics are becoming urgent, as systems can be negatively biased and the decision processes are often not traceable, while impacting our lives. Abstract concepts such as fairness and values need to find their way into the fast and agile AI development processes. The contemporary (research and practice) fields tackle these challenges by technological feats, ethical AI principles and strategies. However, it are the decisions made by humans today and tomorrow that will shape our future. It is, therefore, alarming the translation of ethics to that day to day work of the AI development team is missing. Hence, the central aim of this thesis is to explore and design support for AI teams with the creation of more ethical AI systems, bridging the gap between ethical AI principles and current practice. By that, design for organizational capacity for the development of fairer AI by using strategic design and critical design approaches. In this thesis, due to the diversity and magnitude of ethical challenges in AI, particular attention is paid to two challenges, fairness and value-alignment, to benefit from a design perspective. Three streams of expertise are brought together to tackle these challenges: AI, applied ethics and design. Ethics bears critique, and this thesis argues that it can benefit from a design perspective, using imagination in the solution space and synthesized thinking for implementable ideas instead of solely discussion. The thesis focuses on ways how design approaches can supplement the ethical ones and thereby stimulate the ethical uptake in the AI field. Instead of defining what fairness is, this thesis takes a novel approach in unraveling ten unfairness sources in the AI development. It is aspired to reduce these sources of unfairness in AI, in project specific fashion. In AI practice, the ways ethics is incorporated and how value tensions are resolved is under-researched. In depth interviews, generative tools and provotypes are conducted and designed to research and critique the contemporary AI field in relation to ethics, both with IBM and their clients. Simultaneously to inquire novel value tensions in its development. Five main value tensions are unraveled in its relation to fairness. All above is consolidated a framework to design for organizational capacity and team support leading to the creation of fairer and value-aligned AI systems. With this framework an organizational role is designed, the ethical coach, to aid the AI team with cocreating fairer and value-aligned AI systems with an accompanying modular toolkit. The modular toolkit is iterated upon multiple times and uses the AI dish metaphor. Finally, two evaluation sessions with IBM and their clients as well as the conversations concerning of the implementation of the toolkit led to recommendations for further development including education and implementation directions.