A Responsive Engagement Approach to Promote the Development of ‘Fairer’ Algorithms

There is much contemporary concern about ‘unfairness’ in algorithmic systems. Public controversies have arisen over lack of transparency and accountability in the development and application of algorithmic systems, as well as their potential to produce outcomes that are systematically unfavourable to certain groups. As a result, a variety of fairness criteria and metrics have been proposed to guide the development of algorithmic systems. However, it is unclear whether and how wider society can be involved in deciding which of these different fairness criteria should be favoured. Our work addresses this question by drawing on Responsible Innovation (RI) and ‘society in the loop’ (STIL) approaches. These suggest that the development of ‘fairer’ algorithmic systems may be facilitated through responsive engagement with societal stakeholders. We conducted an exploratory study to determine whether it is possible to present a complex set of algorithms to lay stakeholders in a way that enables them to make informed decisions about them. We presented participants with two limited resource allocation scenarios and a set of algorithms; we then asked them to select which of the algorithms they most and least preferred for the allocation. We collected quantitative data recording participant selections and qualitative data capturing how participants explained and justified their selections. We found that participants were able to meaningfully interrogate the algorithms presented to them and displayed grounded understanding of the consequences of different selections. Whilst there was no overall consensus in either scenario, participants displayed patterns in their reasoning. They consistently treated their decisions as contingent on specific understandings of fairness and context, and different interpretations of these matters accounted for different preference selections. These insights and the approach itself can be incorporated into co design processes for contemporary algorithmic systems.