Plenary Session
Title: “Thinking Slow Is Not Thinking Well: Decision Science for Reasoning Machines”
Abstract
Artificial intelligence is undergoing a transition that decision scientists will recognise. Early AI systems either responded quickly and intuitively or reasoned rigorously within narrowly defined domains. Today’s reasoning models attempt something different: they generate, compare and evaluate alternative lines of reasoning before answering. In effect, they are learning to deliberate.
This talk argues that thinking slowly is not the same as thinking well. Slow reasoning without good structure can still be inconsistent or unfair, and decision science has spent decades studying the difference. I begin by showing that people remain inconsistent even when they deliberate carefully, for example when giving pairwise comparisons in methods such as the Analytic Hierarchy Process. Rather than treating inconsistency as error, decision science can interpret it as information about uncertainty in underlying preferences. Disagreement is not always a defect; sometimes it is the signal.
I then examine how the structure of deliberation influences the conclusions that people reach. Different elicitation processes, visualisations and decision frameworks do not merely support reasoning; they actively shape it. Understanding deliberation therefore requires understanding the mechanisms through which decisions are formed.
The same challenge now confronts artificial intelligence. If reasoning machines generate multiple possible answers, what can disagreement between those answers tell us? And if an algorithm produces an unfair outcome, how can we identify the sources of that unfairness? Using a threshold-dependent Shapley value method for attributing classifier unfairness, I show how explanations can reveal not only what a model predicts but which factors contribute to predictive disparity.
I close by arguing that many of the emerging challenges in reasoning AI, including uncertainty, disagreement, explanation and stopping decisions, are longstanding questions in decision science. AI has learnt to deliberate. Decision science can help it deliberate well.
Biography
Sajid Siraj is Associate Professor in Business Analytics at Leeds University Business School, University of Leeds, UK, where he directs the University’s eXplainable Artificial Intelligence Hub. His research brings together the fields of decision science and artificial intelligence, focusing on preference elicitation, uncertainty, decision support, explainable AI and algorithmic fairness. He holds a PhD in Computer Science from the University of Manchester. His research on pairwise comparisons and inconsistency has contributed to new approaches for understanding uncertainty in human judgement. More recently, his work has focused on explainable and fair AI, including Shapley value-based methods for attributing classifier unfairness and improving transparency in automated decision-making. He is the creator of PriEsT decision support software along with several other software tools; and is co-author of Smart Decisions: A Structured Approach to Decision Analysis (Wiley, 2024). His research appears in journals including the European Journal of Operational Research, IEEE Transactions on Artificial Intelligence, Engineering Applications of Artificial Intelligence and Applied Soft Computing. He co-chaired EURO 2025 Conference in Leeds, and has served on the editorial board of the Journal of Multi-Criteria Decision Analysis, Soft Computing Journal, and is also one of the guest-editors for European Journal of Operational Research.
