Interactive Explanations by Conflict Resolution via Argumentative Exchanges

Interactive explanations work better as conversations than static statements: Rather than simply telling users why an AI made a decision, the paper demonstrates that allowing humans and AI to exchange arguments back-and-forth—similar to a debate—more effectively resolves disagreements about AI predictions. This suggests explainability tools should be designed as dialogues rather than one-way information delivery.

The strongest argument doesn't always convince people: The research found experimentally that "the strongest argument may not always be the most effective," meaning that logically superior explanations don't necessarily lead to better understanding or agreement between humans and AI systems. This has practical implications for how explanations should be tailored to human reasoning rather than pure logical strength.

AI systems can learn to explain using counterfactual reasoning: The paper shows that AI agents can be programmed to explain decisions by presenting counterfactual patterns (e.g., "if this feature had been different, the prediction would have changed"), which mirrors how humans naturally explain their reasoning and may improve mutual understanding.

Human cognitive biases should be explicitly considered in AI explanations: The framework accounts for how human cognitive biases affect interpretation of explanations, suggesting that responsible AI systems should be designed with awareness of these biases rather than assuming humans will interpret information objectively.

Multi-agent argumentation frameworks can resolve human-AI conflicts: By formalizing explanations as structured arguments that agents can exchange, the approach provides a systematic method for addressing disagreements between humans and AI systems, moving beyond ad-hoc explanation methods.