Altering Power Relations in Defining AI Fairness: A Process of Negotiation from the Mammoth Project

The fairness of Artificial Intelligence (AI) systems is at the centre of many debates, but there are different perspectives on what it entails. At present, technical approaches dominate the scene. Discourses on AI fairness are frequently structured around the proposition of technical solutions designed to address algorithmic biases. This approach tend to privilege a computer science perspective on the problem and frequently fails to address the social context in which these systems are constructed and to recognise the diversity of opinions regarding what constitutes fairness when different stakeholders are involved in the discussion. As part of the EU-funded MAMMOth project, we investigated a) how the problem of fairness in AI is approached by different disciplines through 29 interviews with experts from both computer sciences and social sciences, and b) how fairness is perceived by those groups who are often discriminated and marginalised by AI systems through 6 co-creation workshops and a survey which collected 171 responses. While the results show different approaches to fairness across the participants, they also help to identify a common ground on which to build a lay-socio-technical view on the issue. The presentation will demonstrate how these disparate approaches and opinions have been reconciled within the MAMMOth project, and how fairness has been negotiated both between disciplines and in order to translate the concerns of vulnerable groups into technical requirements. The transition towards a lay-socio-technical perspective on AI fairness necessitates the redefinition of current power relations. These include the rebalancing of power between disciplines, where computer science currently holds a more dominant position than social sciences in determining the direction of technological development, and between experts and lay people, as people from vulnerable groups are often left out of the conversation.