Evaluative AI gets argumentative foundation for contestable decisions

A new position paper from Xiang Yin, Tim Miller, Nico Potyka, Antonio Rago, and Francesca Toni makes the case for putting computational argumentation at the core of Evaluative AI (EAI). Unlike conventional AI decision support systems that output one recommendation, EAI is designed to present competing hypotheses alongside evidence for and against each option. The authors argue that this approach gives human decision-makers a fuller picture, letting them weigh alternatives rather than blindly trust a single suggestion. The paper, posted on arXiv as 2608.07473, positions argumentation as the natural formal framework for making EAI both explainable and contestable.

The core idea is that computational argumentation — a mature field from AI and multi-agent systems — offers rigorously defined semantics for constructing, comparing, and attacking arguments. This makes it possible to compute which hypotheses are justified under uncertain or conflicting evidence. Crucially, because the reasoning is formal and transparent, users can inspect why a particular hypothesis was supported or challenged, and even contest the conclusions. The authors outline a long-term research agenda toward distributed, human-centred EAI systems that support accountability in high-stakes settings. By grounding EAI in argumentation, they aim to move beyond black-box recommendations to a paradigm where AI and humans genuinely collaborate on decisions.