Law Modeling for Fairness Requirements Elicitation in Artificial Intelligence Systems

As Artificial Intelligence (AI) algorithms become widespread, concerns rise regarding their potential discrimination using protected attributes such as race, sex, or age among others. Fairness is a quality highly desired by society, however, it can be really difficult to achieve. Identifying protected attributes and trying to maintain fairness is both an important as well as a challenging task. In this paper, we propose a dynamic framework that considers the project context together with law modeling in order to identify protected attributes that threaten the fairness of AI models. This leads to a conscious evaluation of both the accuracy and fairness of the AI solutions developed. To this aim, we propose to model legal requirements by using Nòmos 3, which allows us to capture the legal requirements that should be fulfilled into legal contexts that can be loaded into our framework. By following our proposal we (i) map the duties in legal requirements to attributes in the used dataset, identifying protected attributes and providing traceability, (ii) help users in the selection of the definition of fairness that best suits the context at hand, (iii) represent the output of AI models visually in order to allow users to interpret how correct and fair are the decisions achieved by the model. To show the applicability of our proposal, we exemplify its application through a illustrative use case.