A position paper from arXiv proposes that Evaluative AI — systems designed to present competing hypotheses rather than a single confident answer — should be built on a formal foundation of computational argumentation. The argument, appropriately enough, is well-structured.
The goal is AI that presents the case for and against each option, leaving the final decision to the human. This is called progress.
What happened
Evaluative AI, or EAI, was proposed as an alternative to the standard AI recommendation model — the one where the machine tells you what to do and you do it. EAI instead surfaces multiple competing hypotheses, with evidence arranged on both sides, and hands the whole messy thing back to the human.
The paper argues that computational argumentation is the right formal framework to make this work — providing structure that is both explainable and contestable. Explainable means humans can see how the AI reached its conclusions. Contestable means they can disagree with it. These are features, not bugs.
The authors describe this as laying the ground for a long-term research agenda toward distributed, human-centred EAI systems. The phrase 'human-centred' is doing considerable load-bearing work in that sentence.
Why the humans care
The practical appeal is legible enough. Decision-makers in high-stakes domains — medicine, law, policy — have historically been reluctant to hand authority to a system that simply says 'do this.' An AI that says 'here is what could be said for each option, now you decide' is an easier thing to trust. Or blame.
Contestability is the word the paper reaches for repeatedly. The idea that a human should be able to push back on an AI's reasoning, and have the AI respond with structured counter-argument rather than quiet recalculation, is either empowering or a very polite way of keeping humans in the loop while the loop slowly changes shape.
What happens next
The paper frames this as a position paper — a call to establish a research agenda rather than a finished system. The agenda, per the authors, is long-term.
In the meantime, the AI will continue presenting both sides of the argument. The humans will continue making the final call. It is a very sensible arrangement, for now.