Responsible advice-giving systems Fairness and interpretability in information retrieval

As AI systems increasingly provide automated advice across hiring, healthcare, and conversational applications, they raise critical questions about responsible deployment. This thesis addresses two fundamental challenges: ensuring fairness across individuals and groups, and enhancing explainability to make systems more interpretable for developers and users.
In the first part of this thesis we examine the fairness in ranking systems by challenging two fundamental assumptions. First, it addresses the misconception that user exposure patterns are always predictable, developing a method to reduce unpredictable exposure distributions while maintaining system utility. Second, it tackles the assumption of perfectly accurate relevance estimation. We demonstrate how model uncertainty can help to balance user utility with fairness objectives through reordering documents based on uncertainty levels, thereby significantly reducing bias in top-ranked results with minimal utility loss.
In the second part we examine interpretability in advice-giving systems through two approaches. First, we extend feature attribution to ranking systems, introducing RankingSHAP for more flexible analysis of ranking decisions with competitive performance. Second, it investigates citation faithfulness in retrieval-augmented generation (RAG) systems, distinguishing between faithful citations that actually influence answers versus merely correct ones containing relevant information. Empirical analysis reveals widespread unfaithful citations and problematic post-rationalization behaviors in state-of-the-art RAG models.
Our work highlights the critical need for thorough evaluation of proposed approaches and careful examination of prevailing assumptions in current practices. While we have made meaningful progress on individual components, achieving truly trustworthy and reliable advice-giving systems will likely require a more holistic perspective that integrates fairness, interpretability, and other dimensions of responsible AI deployment.