The Bias Network Approach: a sociotechnical approach to aid AI developers to contextualise and address biases

In this thesis, I will introduce the Bias Network Approach (BNA) as a novel sociotechnical intervention designed to aid AI developers in identifying and addressing biases more comprehensively. The methodology of this thesis is mixed. Although primarily philosophical, it also includes an empirical case study demonstrating how to use the BNA in real life. Hence, in the second half of the thesis, I will discuss the case study findings to analyse how they support my theoretical proposal and the philosophical arguments presented in the first half of the thesis.

In Chapter 1, I will criticise “the problems of bias”: technocentrism, conceptual ambiguity, and isolated approaches to identify and mitigate bias. In Chapter 2, I will argue in favour of a sociotechnical approach, reviewing sociotechnical proposals by other researchers, and drawing key elements from them to ground my BNA proposal. In Chapter 3, I will address the conceptual ambiguity problem in more detail, contrasting “negative” and “positive” views of bias in AI, to then untangle a working definition for bias to be used in the BNA. In Chapter 4, I will present the BNA proposal through an empirical case study, analysing its main findings. In Chapter 5, I will provide guidance for developers and prompters wishing to adopt the BNA as a transitional intervention to promote ethical reflection. In Chapter 6, I will explain how responsibility should be attributed to developers, prompters, and organisations applying the BNA, including insights about responsibility as a moral obligation, forward- backward- and active responsibility, as well as the ethical agency of AI developers.

Finally, there is an Afterword, in which I will discuss an avenue for future work and hypothesise on how the core idea of a network approach could be extended to other ethical concerns in AI ethics.