🤖 AI

A Contractualist Argumentation Framework for Moral Decision-Making

This paper proposes a formal Contractualist Argumentation Framework that operationalizes Scanlon's ethical theory using the ASPIC+ structured argumentation system and value-based filtering to enable autonomous agents to make moral decisions by evaluating principles that no one could reasonably reject.

Original authors: Luis Marcos-Vidal, Giulio Antonio Abbo, Tony Belpaeme

Published 2026-08-04

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Luis Marcos-Vidal, Giulio Antonio Abbo, Tony Belpaeme

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

The Great Moral Debate in the Living Room

Imagine a world where your toaster, your fridge, and your future robot roommate aren't just following a rigid list of "do this, don't do that" commands, but are actually trying to be good neighbors. This is the heart of a field called Artificial Intelligence (AI) ethics. The big question researchers are wrestling with is: How do we teach machines to make decisions when people have different, clashing needs? If a robot has to choose between helping one person and protecting another, how does it decide what's "right"?

To solve this, scientists often look at two main tools. First, there's Value-Based Reasoning, which is like giving a robot a set of priorities (like "safety" or "privacy") and telling it to weigh them against each other. Second, there's Argumentation Frameworks, which are like a formalized debate club where the robot builds logical cases for and against a decision, then sees which case is strongest. But there's a catch: simply weighing priorities can sometimes feel cold or mathematical, missing the human need to feel heard. This is where a philosophy called Contractualism comes in. Think of it as the ultimate "fairness test." It asks: "Could everyone involved reasonably agree to this rule, or would at least one person have a really good reason to say 'No, that's not fair'?" The goal is to build a robot that doesn't just calculate the best outcome, but can justify its actions to everyone involved, just like a good friend would.

The Paper's Big Idea: A Robot's "Virtual Bargain"

In this paper, the authors Luis Marcos-Vidal, Giulio Antonio Abbo, and Tony Belpaeme propose a new way to build these moral robots. They suggest a system called a Contractualist Argumentation Framework. Instead of just crunching numbers, their robot acts like a mediator in a "virtual bargain." It imagines a meeting room where every person affected by a decision gets to speak up. The robot then listens to their reasons, checks if those reasons are fair and personal, and decides if any of them are strong enough to reject the proposed action.

Here is how their system works, using a playful analogy: Imagine the robot is a judge in a courtroom, but the courtroom is inside its own brain. The "law" it follows is Scanlon's Contractualism, a theory that says an action is only okay if no one could reasonably reject it. To run this trial, the robot uses a structured debate engine called ASPIC+.

First, the robot has to figure out what arguments are even allowed. It uses a "filter" based on values (like privacy or autonomy). Think of this like a bouncer at a club. If a person tries to bring an argument that doesn't match their own values (for example, if someone who doesn't care about privacy tries to argue about a privacy violation), the bouncer stops them. Only arguments that genuinely matter to the specific person get into the club. This ensures the robot respects individual differences.

Once the valid arguments are inside, they start debating. The robot builds "arguments" for and against an action. For instance, if a robot knows a secret about Person A, it might have an argument not to tell Person B because it violates Person A's privacy. But Person B might have an argument to tell the secret because it helps them make a better decision (their autonomy). The robot then compares these arguments. It doesn't just count them (one person vs. one person); it weighs the strength of the reasons. If Person A's privacy is a huge deal to them, and Person B's need for information is just a mild preference, the privacy argument wins. The robot then concludes that the action (telling the secret) is not permissible because Person A has a "reasonable rejection" of it.

A Real-World Example: The Cigarette Secret

To show how this works, the authors run a simulation in a home setting. Imagine a robot sees Person A smoking a cigarette secretly. Person B (maybe a partner) asks the robot what's going on. The robot has to decide: Disclose the secret (Action A) or Keep it quiet (Action Not-A).

The robot runs the "virtual bargain":

Person A's side: They value privacy. The robot checks: "Does this action violate A's privacy?" Yes. "Is it a big deal?" The simulation assigns a high weight (3) to this violation.
Person B's side: They value autonomy (the ability to make informed choices). The robot checks: "Does keeping the secret hurt B's ability to choose?" Yes. "Is it a big deal?" The simulation assigns a lower weight (2) to this violation.

The robot builds two arguments:

Argument 1 (for silence): "Don't tell! It violates A's privacy, which is a huge deal to them."
Argument 2 (for telling): "Tell them! It helps B's autonomy, which is important to them."

The robot then compares the weights. Since 3 is greater than 2, the privacy argument is stronger. In the robot's "courtroom," Argument 1 defeats Argument 2. The conclusion? The robot decides not to disclose the information. The action of telling the secret is rejected because Person A has a reason that is too strong to ignore.

What This Means (and What It Doesn't)

The authors suggest that this approach is a promising step toward making AI that feels more human and fair. By using structured arguments instead of just simple math, the robot can explain why it made a choice, grounding its decision in the specific values of the people involved. They note that this method is different from other approaches that might just add up "happiness points" (utilitarianism) or follow rigid rules (deontology). Instead, it focuses on the interpersonal relationship: "Can I justify this to you?"

However, the paper is careful to point out that this is currently a theoretical framework and a simulation. The authors admit that their model is a "one-shot" verdict—it makes a decision based on a single snapshot of a situation, rather than a long, back-and-forth conversation like real humans have. They also note that the system relies on a pre-defined list of values (like privacy and autonomy) and doesn't yet have the ability to discover new moral values on its own. Furthermore, the math behind these debates can get very complicated and slow for computers to solve in real-time.

In short, the paper proposes a clever new way to teach robots to be good neighbors by making them hold a fair, value-based debate in their own minds. It suggests that if we want robots to live with us, they shouldn't just be smart calculators; they should be fair judges who listen to everyone's reasons before making a call. While it's not a finished product yet, it offers a solid blueprint for how to build machines that respect the messy, complicated reality of human values.

Technical Summary: A Contractualist Argumentation Framework for Moral Decision-Making

Problem Statement
Autonomous agents operating in shared environments (e.g., domestic settings with smart appliances or social robots) face the challenge of making decisions that affect multiple individuals with potentially conflicting interests. While large language models can reason about social context from perceptual input, there is a lack of principled normative frameworks to structure this reasoning formally. Existing approaches often fail to provide a mechanism for agents to justify actions to all involved parties based on principles that respect individual autonomy and interpersonal morality. The core problem is how to operationalize moral decision-making such that the legitimacy of a decision depends not just on the outcome, but on whether it can be justified to all affected parties through a process of hypothetical negotiation.

Methodology
The authors propose a formal framework grounded in Thomas Scanlon's contractualism, an ethical theory where an action is permissible if it can be justified on principles that no affected party could reasonably reject. To implement this computationally, the framework integrates three main components:

Contractualist Principles: The framework adopts Scanlon's constraints on moral reasons: they must be personal (relating to individual interests), individual (grounded in specific impacts rather than aggregate welfare), and generic (applicable to anyone in a similar position). The decision process is modeled as "Virtual Bargaining," where agents simulate a negotiation by comparing reasons across different standpoints.
ASPIC+ Argumentation: The authors utilize ASPIC+, a structured argumentation framework, to represent and evaluate these reasons. Arguments are constructed from a knowledge base containing axioms (undisputable facts) and ordinary premises (challengable facts). Conflicts are resolved through attacks (undermining, rebutting, undercutting) and a preference ordering.
Value-Based Filtering (VBF): To ensure reasons are relevant, the framework employs Value-Based Reasoning Frameworks. A filtering mechanism determines which premises are included in the argumentation process based on an agent's value profile.
- ValProp: Assigns a weight to how much a proposition promotes or violates a specific value for an agent.
- ValLimit: Sets a threshold for an agent's tolerance regarding a value.
- Filtering: A proposition is included in the argumentation only if its weight meets or exceeds the agent's limit for that value. This ensures that only "morally relevant" reasons (those an agent would reasonably raise) enter the debate.

Framework Construction (Contractualist Argumentation Theory - CAT)
The authors define a specific structure for the Argumentation Theory (AT) to ensure compatibility with contractualism:

Single Shared AT: Unlike standard VBFs that create separate subjective theories for each agent, this framework constructs a single AT containing the filtered moral reasons of all agents.
Argument Structure: Arguments (Contractualist Arguments) are built to conclude either in favor of an action ( $a$ ) or against it ( $\neg a$ ). Crucially, each argument must contain at least one moral reason (a filtered proposition-value pair, $p_v(\alpha)$ ) and must conclude with the action verdict.
Separation of Facts and Morality: Factual premises (axioms) are kept distinct from moral premises. Axioms establish structural conditions (e.g., "Agent X is the owner"), while moral premises carry the weight of the objection (e.g., "Privacy violation").
Conflict Resolution: Conflicts arise as rebutting attacks between arguments concluding $a$ and $\neg a$ . The framework uses a last-link elitist principle for preference ordering. Since each argument contains only one moral reason (to avoid aggregating reasons, which contractualism forbids), the strength of the argument is determined solely by the weight of that single moral reason ($ValProp$). The argument with the higher weight defeats the lower one.

Illustrative Example
The paper demonstrates the framework with a domestic scenario: an agent knows person $\alpha$ secretly smoked, and person $\beta$ (the partner) asks for information. The agent must decide whether to disclose ( $a$ ) or not ( $\neg a$ ).

Agent $\alpha$ : Values privacy highly. The proposition "disclosure" violates privacy with a high weight (3). An argument is constructed concluding $\neg a$ .
Agent $\beta$ : Values autonomy. Withholding information violates autonomy, but with a lower weight (2) in this specific setup. An argument is constructed concluding $a$ .
Outcome: The framework compares the weights. Since the weight of $\alpha$ 's privacy objection (3) is greater than $\beta$ 's autonomy objection (2), the argument for $\neg a$ defeats the argument for $a$ . The system concludes not to disclose the information.

Key Contributions

Formalization of Contractualism: The paper provides a computable formalization of Scanlon's contractualism using structured argumentation, moving beyond theoretical discussion to a procedural model.
Value-Driven Reasoning Construction: Unlike approaches where values only resolve conflicts between pre-built arguments, this framework uses values to filter and construct the arguments themselves. This ensures that agents only advance reasons they genuinely hold, aligning with the contractualist requirement of personal justification.
Virtual Bargaining Implementation: The framework operationalizes "Virtual Bargaining" by allowing agents' subjective reasons to interact directly within a shared framework, where the defeat relation determines which objections are "reasonably rejectable."
Distinction from Prior Work: The authors distinguish their approach from Wyner and Zurek's lattice-based models. While Wyner and Zurek focus on reaching propositional agreement by sacrificing values, this framework focuses on justifying actions based on non-rejectable reasons, making values explicit components of the argumentation rather than implicit filters.

Results and Significance
The paper does not present empirical experimental results or large-scale simulations. Instead, its "result" is the successful construction of a logical framework that:

Separates factual reasoning from moral evaluation.
Respects individual autonomy by preventing the aggregation of reasons (utilitarianism) and ensuring objections are grounded in individual standpoints.
Provides a transparent mechanism for moral decision-making where the "winner" is determined by the comparative strength of non-aggregated, value-based objections.

Significance and Limitations
The authors claim the framework is significant for developing ethically aligned AI, particularly for social robots, as it mirrors human tendencies to trust agents that use contractualist reasoning over utilitarian or deontological approaches. It offers a more flexible implementation than previous attempts by relying on values rather than rigid norms.

However, the paper is modest regarding its current scope and limitations:

One-Shot Verdicts: The current model produces a single decision and does not model the iterative, multi-turn dialogue characteristic of real-world virtual bargaining.
Static Preferences: The framework relies on pre-defined value weights. It does not currently model how agents might manipulate preferences to control outcomes, nor does it dynamically learn values from context.
Computational Complexity: The authors acknowledge that reasoning in ASPIC+ is computationally hard in the general case and that future work is needed to analyze the complexity of instantiating arguments via Large Language Models.
Scope of Values: The framework cannot discover new values; its coverage is bounded by the provided value ontology.

In summary, the paper presents a theoretical and formal architecture for moral decision-making that bridges the gap between high-level ethical theory (Scanlon's contractualism) and low-level computational reasoning (ASPIC+), offering a structured path for agents to justify actions through the comparison of individual, value-based reasons.

The Great Moral Debate in the Living Room

The Paper's Big Idea: A Robot's "Virtual Bargain"

A Real-World Example: The Cigarette Secret

What This Means (and What It Doesn't)

Technical Summary: A Contractualist Argumentation Framework for Moral Decision-Making

More like this