🤖 AI

Towards an Argumentative Foundation for Evaluative AI

This position paper advocates for using computational argumentation as a formal foundation for Evaluative AI, enabling systems that support human decision-making by presenting competing hypotheses with evidence in an explainable and contestable manner.

Original authors: Xiang Yin, Tim Miller, Nico Potyka, Antonio Rago, Francesca Toni

Published 2026-08-11

📖 8 min read🧠 Deep dive

CC BY 4.0

Original authors: Xiang Yin, Tim Miller, Nico Potyka, Antonio Rago, Francesca Toni

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are walking into a giant, noisy library where the books are trying to help you solve a mystery. In the world of Artificial Intelligence, there's a big debate about how these digital helpers should talk to us. For a long time, the standard approach was like a strict librarian who points to a single book and says, "This is the answer. Trust me." This is called "recommendation," and while it's fast, it often makes us stop thinking for ourselves. We might blindly follow the advice or ignore it completely without understanding why. But there's a newer, more exciting idea called Evaluative AI. Instead of shouting out one final answer, this type of AI acts more like a friendly debate coach. It lays out several possible theories about what's happening, shows you the evidence for each one, and also the evidence against them. It invites you to step in, look at the clues, and decide for yourself which theory makes the most sense. This keeps you in the driver's seat, making the AI a partner rather than a boss.

This paper, written by a team of researchers from universities in the UK and Australia, suggests a specific, powerful way to build this kind of "debate coach" AI. They argue that the best tool for the job is something called Argumentation. Think of this not as a shouting match, but as a structured game of "connect the dots" where every piece of evidence is a node, and the links between them show how they support or attack each other. The authors propose using a mathematical framework called Weighted Quantitative Bipolar Argumentation Frameworks (wQBAFs). In plain English, this is a way to draw a map where every clue has a "confidence score" (how strong it is) and every connection has a "weight" (how much influence it has). By crunching these numbers, the AI can create a ranked list of the best theories, not just a single guess. The paper suggests that this method is superior to older ways of doing things because it's transparent—you can see exactly why a theory is ranked high—and, most importantly, it's contestable. If you, the human, think a clue is wrong or a connection is weak, you can challenge it, tweak the numbers, and watch the ranking change. It turns AI from a black box into a collaborative workshop where humans and machines can argue, refine, and agree on the best path forward.

The Core Idea: From "Here's the Answer" to "Let's Debate"

The authors start by pointing out a flaw in how we usually use AI. Most systems work on a "recommend-and-explain" model. The AI picks a winner, and then tries to justify it. The problem? Studies show this makes people stop thinking. They either get "cognitive fixation"—staring at the AI's choice like a hypnotized rabbit—or they dismiss it entirely without a second thought.

To fix this, the paper champions Evaluative AI (EAI). Instead of handing you a single recommendation, EAI presents a menu of plausible hypotheses (guesses) along with the "pros" and "cons" for each. Imagine a doctor using an AI to diagnose a patient. Instead of the AI saying, "It's definitely pneumonia," it might say, "Here are three possibilities: Viral Pneumonia, Bacterial Pneumonia, or Asthma. Here is the evidence supporting each, and here is the evidence that argues against them. You decide."

The Solution: The "Argument Map" Game

The paper proposes that the best way to build this system is using Argumentation. The authors suggest a specific mathematical structure called a wQBAF (Weighted Quantitative Bipolar Argumentation Framework). Let's break this down with a metaphor.

Imagine a giant, floating web of glowing nodes.

The Nodes: Some nodes are Evidence (like "High Fever" or "Cough"), and some are Hypotheses (like "Viral Pneumonia" or "Asthma").
The Connections: Lines connect these nodes. Some lines are Green (Support), meaning "This clue makes that theory more likely." Others are Red (Attack), meaning "This clue makes that theory less likely."
The Weights: Every node starts with a Base Score (how strong the clue is to begin with). Every line has a Weight (how much influence that connection has).

In the paper's example, a patient has a high fever and a cough. These are the starting nodes. The fever might strongly support "Viral Pneumonia" but weakly support "Asthma." The AI doesn't just guess; it runs a calculation. It looks at the base scores and the strength of the connections, then "recalculates" the strength of every hypothesis based on the tug-of-war between the green and red lines.

The result isn't a single answer, but a Ranked List. The hypothesis with the highest final strength sits at the top, followed by the next best, and so on. In the paper's healthcare example, the system might rank "Anti-viral therapy" and "Antibiotics" as equally strong (both with a strength score of 0.40), while "Bronchodilator" is lower (0.31).

Why This Approach is Special

The authors argue that this argumentative approach has three superpowers that other methods lack:

Explainability: Because the system is built on a visible map of arguments, it can explain why a hypothesis is ranked high. It can point to the specific "High Fever" node and the strong green line connecting it to "Viral Pneumonia" and say, "This is why we think this is the top choice." It's not magic; it's a traceable path.
Contestability: This is the big one. In many AI systems, if you disagree, you're stuck. Here, because the system is just a map of weights and connections, you can change the map. If you think the "Cough" symptom isn't as important as the AI thinks, you can lower its weight. If you have a new piece of evidence, you can add a new node. The system instantly recalculates the ranking. This turns the user from a passive observer into an active participant who can challenge the AI's reasoning.
Multi-Agent Collaboration: The paper envisions a future where different "agents" (maybe one is a human doctor, another is a specialized database, and a third is a different AI model) each build their own argument maps. They can then combine their maps, argue with each other, and fuse their opinions to create a stronger, more robust final ranking. It's like a roundtable of experts where everyone brings their own notes, and they debate until they find the best solution together.

The Rules of the Game

To make sure this system behaves rationally, the authors propose a set of "Principles" that the ranking system should follow. They are like the rules of a fair game:

Monotonicity: If you add more evidence supporting a theory, its rank should never go down.
Balance: If you add equal amounts of "pro" and "con" evidence, the ranking shouldn't change.
Dominance: If one theory has more or stronger supporting evidence than another (with the same opposing evidence), it must be ranked higher.
Robustness: Small changes in the data shouldn't cause the whole ranking to flip upside down in a crazy way.
Explainability & Contestability: As mentioned, the system must be able to explain its logic and allow humans to challenge it.

The paper notes that many existing mathematical methods for argumentation already satisfy the first four rules, but the authors are suggesting that we need to keep developing new methods to ensure the system is also robust and truly contestable.

The Limits and the Future

The authors are careful not to claim they have solved everything. They admit that building these argument maps is hard. Right now, getting an AI to correctly identify all the arguments, the connections, and the weights is a major bottleneck. They suggest that newer "Large Language Models" (the smart chatbots we know) might help build these maps automatically, but it's not perfect yet.

They also warn about the "contestability" feature. If humans can change the weights, what stops someone from messing with the system on purpose? They suggest we might need "permissioned" systems where only trusted users can change the rules, or ways to weigh the trustworthiness of the person making the change.

Finally, they acknowledge that in a multi-agent world, different experts might disagree deeply. Sometimes, there won't be a single "correct" answer. The goal of their system isn't to force everyone to agree, but to manage that disagreement in a structured, transparent way so humans can navigate the conflict.

In short, this paper doesn't just propose a new algorithm; it proposes a new philosophy for AI. It suggests that the future of smart machines isn't about them being the smartest person in the room, but about them being the best facilitators of human debate, helping us weigh the evidence, challenge our assumptions, and arrive at decisions we can truly trust.

Technical Summary: Towards an Argumentative Foundation for Evaluative AI

1. Problem Definition

The paper addresses the limitations of current Explainable AI (XAI) paradigms, specifically the "recommend-and-explain" model where an AI produces a single output followed by justification. The authors argue this approach induces cognitive fixation, leading users to over-rely on or dismiss AI recommendations without sufficient deliberation.

To counter this, the paper focuses on Evaluative AI (EAI), a paradigm that supports human decision-making by presenting competing hypotheses alongside structured evidence for and against each, rather than a single recommendation. The core technical challenge identified is formalizing EAI as a ranking-based problem: given a set of hypotheses ( $H$ ) and evidence ( $E$ ), the system must determine a total preorder (ranking) over the hypotheses based on weighted pro and con dependencies, rather than selecting a single winner.

2. Methodology

The authors propose Weighted Quantitative Bipolar Argumentation Frameworks (wQBAFs) as the formal and computational foundation for ranking-based EAI.

Formal Problem Definition

The EAI problem is defined by a tuple $E = \langle E, H, \text{pro}, \text{con}, \tau, w \rangle$ :

$E$ and $H$ : Finite sets of evidence and hypotheses.
$\text{pro}, \text{con}$ : Disjoint binary relations representing support and attack dependencies. Crucially, these relations exist not only between evidence and hypotheses ( $E \times H$ ) but also among evidence ( $E \times E$ ) and among hypotheses ( $H \times H$ ), allowing for multi-step reasoning.
$\tau$ : A function mapping elements to initial weights in $[0, 1]$ , representing plausibility or severity before considering dependencies.
$w$ : A function mapping relations to weights in $[0, 1]$ , representing the strength of influence.

Argumentative Solution

The problem is mapped to a wQBAF $Q = \langle A, R^-, R^+, \tau, w \rangle$ , where:

$A = E \cup H$ (Arguments include both evidence and hypotheses).
$R^-$ and $R^+$ correspond to the $\text{con}$ and $\text{pro}$ relations.
Evaluation: Quantitative evaluation methods (e.g., O-QuAD semantics, MLP-based semantics) are used to recursively compute the final strength of each argument. These methods update initial weights based on the combined strengths of supporters and attackers.
Ranking: The final strengths of the hypotheses in $H$ are compared to generate the total preorder $\succeq_E$ .

Design Principles

The paper outlines seven principles that the argumentative solution must satisfy to be suitable for human-centred EAI:

Monotonicity: Rankings must align with rational intuition regarding evidence updates.
Balance: Rankings remain unchanged if pro and con evidence strengths are balanced.
Equivalence: Hypotheses with identical evidence profiles must be ranked equally.
Dominance: Hypotheses with superior pro evidence (in number or weight) must be ranked higher.
Robustness: Rankings should be stable against minor perturbations in evidence weights.
Explainability: The ranking mechanism must be transparent, supporting local (why A > B) and global explanations.
Contestability: Users must be able to challenge the ranking by modifying evidence weights, adding new evidence, or contesting hypotheses.

3. Key Contributions

The paper makes three primary contributions to the research agenda:

Formalization: It defines EAI formally as a ranking-based problem over hypotheses given weighted evidence, moving beyond the conceptual proposal to a tractable mathematical framework.
Paradigm Selection: It advocates for wQBAFs as the most suitable paradigm for EAI. This approach offers advantages over existing methods (such as Weight of Evidence or LLM-based conversational styles) by supporting richer dependency structures, providing intrinsic explainability, and enabling contestability.
Multi-Agent Vision: It proposes a multi-agent extension where diverse evaluative models (agents) with different reasoning styles can interact, exchange arguments, and aggregate their wQBAFs to form a group-level evaluation, thereby mitigating individual bias.

4. Results and Claims

As a position paper, the work does not present empirical experimental results or a deployed system. Instead, it provides a theoretical analysis and a "skeleton" of a solution (illustrated via a healthcare scenario involving symptoms, diagnoses, and treatments).

The authors claim the following based on their theoretical framework:

Explainability: Because the underlying model is argumentative (a wQBAF), explanations are "intrinsic" and "faithful." The graphical structure explicitly shows reasoning paths, and quantitative strengths can be explained via attribution or counterfactual methods.
Contestability: The framework naturally supports human-in-the-loop interaction. Users can contest specific components (e.g., the initial weight of a symptom or the existence of a relation) to alter the final ranking, positioning AI as a deliberative partner rather than a decision-maker.
Principled Behavior: The paper asserts that existing wQBAF evaluation semantics (e.g., O-QuAD) satisfy the first four principles (Monotonicity, Balance, Equivalence, Dominance). It notes that Robustness and Explainability require further investigation or specific semantic choices.

5. Significance and Limitations

Significance:
The paper's primary impact is transitioning EAI from a conceptual proposal to a formal research agenda. By grounding EAI in argumentation theory, it provides a foundation for systems that are:

Human-centred: Preserving human agency by requiring active deliberation between hypotheses.
Distributed: Capable of integrating heterogeneous sources and multi-agent deliberation.
Trustworthy: Through intrinsic explainability and the ability for users to contest and refine the system's reasoning.

Limitations:
The authors acknowledge several constraints and open challenges:

Construction Bottleneck: The framework assumes the wQBAF (arguments, relations, and weights) is correctly constructed. In practice, reliably extracting this structure from raw data or LLMs remains a challenge.
Manipulation Risks: The openness required for contestability introduces the risk of users providing misleading or low-quality evidence. The paper suggests permissioned contributions and trust-aware weighting as potential safeguards.
Irreducible Disagreement: In multi-agent settings, deep disagreements between agents may be irreducible. The framework suggests managing pluralism and structured disagreement rather than enforcing consensus.