A new arXiv position paper proposes computational argumentation as a foundation for evaluative AI. Instead of producing one recommendation, evaluative AI is meant to help human decision-making by showing competing hypotheses and the evidence for and against each one.
That framing is important for high-stakes or ambiguous decisions. A single confident answer can hide uncertainty, trade-offs, and weak evidence. An argumentation-based system would make disagreement and support structures more explicit, giving users a way to inspect why one option appears stronger than another.
The paper is a research agenda rather than a finished product. It argues that argumentation can make evaluative AI more explainable, contestable, distributed, and human-centered. The practical challenge will be turning formal argument structures into tools people can actually use under time pressure. If that can be done, evaluative systems could become less like automated judges and more like structured assistants that help people reason through evidence before making their own decisions.