Designing for Uncertainty in AI: Supporting Decision-Making in the Process Industry - A User-Centered Approach to Communicating AI Uncertainty for Operators in the Process Industry

Abstract

This thesis investigates how uncertainty in artificial intelligence (AI) systems can be effectively communicated to support decision-making in the process industry. As machine learning-based decision support systems are increasingly integrated into industrial environments, operators are required not only to interpret system outputs but also to assess their reliability. However, current approaches to uncertainty communication are often insufficiently intuitive, limiting trust and appropriate reliance on AI systems. Adopting a human-centered design approach, this study explores how industrial operators perceive and reason about uncertainty in their everyday work, and how different forms of uncertainty communication influence trust, decision-making, and system use. The research is conducted within the context of the pulp and paper industry, with a particular focus on the bleaching process, a complex and safety critical operation. The study follows a Double Diamond design process, combining methods such as a literature review, semi-structured interviews, and observations. Based on these insights, a set of design guidelines for uncertainty communication is developed, validated through a design workshop, and implemented in a high-fidelity user interface prototype. The guidelines are then evaluated through the prototype with domain experts using think-aloud protocols, interviews, and acceptance measures. The findings show that operators rely heavily on experience-based and rule-based reasoning when handling uncertainty, and that transparency, contextual explanations, and historical performance data are essential for building trust in AI systems. Furthermore, effective uncertainty communication encourages more reflective decision-making and supports appropriate reliance on AI recommendations. This thesis contributes with empirically grounded design guidelines and a validated prototype that demonstrate how uncertainty can be communicated in a way that aligns with operators’ cognitive processes and work practices.