Towards an Argumentative Foundation for Evaluative AI
Authors: Xiang Yin, Tim Miller, Nico Potyka, Antonio Rago, Francesca Toni
Submitted: 25 April 2026 (v1)
Subjects: Artificial Intelligence (cs.AI); Multiagent Systems (cs.MA)
arXiv ID: 2608.07473 [cs.AI]
DOI: https://doi.org/10.48550/arXiv.2608.07473
Abstract
Evaluative AI (EAI) has emerged as a promising paradigm for supporting human decision-making. Rather than offering a single, prescriptive recommendation, EAI presents multiple competing hypotheses, each accompanied by evidence both for and against it. In this position paper, we advocate for computational argumentation as a particularly suitable framework for providing a formal, computable foundation for EAI that is both explainable and contestable. We argue that argumentation not only aligns with the core principles of EAI but also offers a robust mechanism for handling conflicting evidence and facilitating critical human oversight. We outline a long-term research agenda aimed at developing distributed, human-centred EAI systems built upon these argumentative principles, with potential applications spanning healthcare, public policy, and financial decision-making. As of 2026, this agenda remains timely, with growing regulatory pressure for explainable AI and increasing demand for systems that can justify their outputs in critical domains.
1. Introduction
In recent years, the limitations of conventional AI decision-support systems—which often output a straightforward best answer—have become increasingly apparent. Such systems struggle to handle uncertainty, conflicting objectives, or situations where stakeholders might reasonably disagree. Evaluative AI (EAI) addresses this by reframing the AI's role as one of evaluation rather than recommendation, presenting users with a spectrum of plausible outcomes and the reasoning behind each. This approach empowers decision-makers to weigh alternatives within their own contextual constraints.
However, for EAI to be practically useful and trustworthy, it must rest on a foundation that is both formal and computationally tractable. We propose that computational argumentation\u2014a subfield of AI concerned with the formal representation and automated analysis of arguments and counterarguments\u2014offers exactly such a foundation. In this position paper, we explore why argumentation is uniquely suited for EAI, how it can be integrated into existing decision-making workflows, and what challenges lie ahead.
2. The Promise of Evaluative AI
EAI represents a shift from a "black box" recommendation approach to an "open box" evaluative approach. Its core tenets include:
- Plurality: Instead of one answer, multiple hypotheses are presented.
- Evidential balance: For each hypothesis, supporting and opposing evidence are both shown.
- Human agency: The final decision rests with the human, who considers the presented alternatives.
This paradigm is particularly valuable in high-stakes settings, such as medical diagnosis, legal adjudication, and risk assessment, where a single AI recommendation may be ethically or pragmatically insufficient. As of 2026, EAI is gaining traction in these fields, yet its theoretical foundations remain underdeveloped.
3. Why Argumentation Fits EAI
Computational argumentation provides a natural fit for the EAI framework for several reasons:
- Formal semantics: Argumentation is built on rigorous logical and graph-based models, enabling precise definitions of concepts like attack, defeat, and acceptability.
- Handling conflict: It is designed to manage conflicting information, mirroring the real-world complexity that EAI must address.
- Explainability: The structure of an argumentative dialogue itself serves as an explanation, showing why a particular hypothesis is more or less credible.
- Contestability: In a well-designed argumentative system, users can challenge the AI's reasoning by introducing new arguments or questioning assumptions, aligning with emerging regulations such as the EU AI Act.
4. A Research Agenda for Argumentative EAI
We propose the following key directions for future research and development:
- Formal foundations: Develop a unified argumentation framework that supports EAI, including rigorous definitions of hypothesis strength and evidence quality.
- Scalable reasoning: Address challenges in computational efficiency, enabling argumentation-based EAI to operate in real-time and on large data sets.
- Human-AI interaction: Design interfaces and interaction protocols that allow users to engage with argumentative outputs intuitively, including mechanisms for iterative refinement.
- Distributed EAI: Extend argumentative principles to multi-agent settings, where multiple AI systems and humans collaborate, sharing evidence and arguments across diverse sources.
- Ethical and societal considerations: Ensure that argumentative EAI is transparent, fair, and respectful of user autonomy, embedding contestability as a core design principle.
5. Conclusion
Evaluative AI offers a compelling alternative to conventional decision-support systems, particularly in contexts where pluralism and human oversight are paramount. By embracing computational argumentation, we can provide EAI with a solid formal basis, paving the way for systems that are not only explainable but also genuinely contestable. As we look towards 2026 and beyond, the convergence of regulatory pressures, technological advancements, and ethical imperatives makes this agenda both timely and urgent. We invite the research community to join us in refining the argumentative foundations of EAI and building the next generation of human-centred AI systems.
via ArXiv AI