Explaining Decisions of Agents in Mixed-Motive Games

Mixed-motive environments require new explanation approaches: Traditional explainability methods designed for purely cooperative or competitive settings fall short in real-world scenarios where agents must simultaneously cooperate and compete. This paper addresses a gap by developing explanation methods specifically tailored to these complex social dilemmas, which are increasingly common as AI systems interact with humans and other agents.

Communication and hidden strategies complicate AI transparency: The research identifies that agents can communicate through both explicit natural language and implicit actions (like strategic moves in games), making their decision-making opaque. The paper demonstrates these issues in games like no-press Diplomacy and prisoner's dilemma variants, showing that understanding AI behavior requires accounting for what agents say versus what they actually do.

Explanations improve human understanding of AI in strategic settings: The authors validated their explanation methods with human subjects in two different games, demonstrating that the explanations actually help people better understand why AI agents make particular decisions. This suggests that responsible AI deployment in competitive or mixed-motive contexts should include tailored explanations rather than generic ones.

One-size-fits-all explanations don't work across different game types: By testing their methods across three games with "vastly different properties," the research shows that explanation approaches must be flexible and context-aware. This has implications for how organizations should think about AI transparency—explanations need to be customized to the specific strategic environment rather than applied universally.