Who Are You Explaining To? A Multi-Agent System for Audience-Aware XAI Narratives

Different audiences need fundamentally different explanations: The paper demonstrates that patients, clinicians, and data scientists cannot effectively understand the same AI prediction explanation in identical form. XstrAI's framework recognizes that medical AI explanations must be tailored to each reader's expertise level, goals, and risk of misinterpretation—a finding with direct implications for how AI systems should communicate their reasoning in high-stakes domains.

LLMs alone create dangerous explanations without proper safeguards: The research identifies critical risks when large language models simply verbalize model evidence, including "weak grounding, conflation of attribution with causal language, and outputs that are persuasive without being faithful to the underlying model evidence." This highlights why responsible AI deployment requires structured validation beyond standard language model outputs.

Explanations must maintain fidelity to actual model evidence: XstrAI uses an "immutable structured representation" of the underlying model evidence (SHAP values) that remains identical across all audiences, ensuring the core facts don't change based on how they're communicated. This separation of evidence from communication prevents the distortion of what the model actually found versus how it's being explained.

Validation mechanisms catch explanation failures: The framework employs specialized agents to check for "grounding, attribution consistency, communicative risk, and audience appropriateness" with automatic revision loops. This multi-layer validation approach outperformed 11 baseline methods, suggesting that responsible AI explanations require explicit quality control rather than relying on single-pass generation.

Responsible explanation requires domain-specific design: Rather than applying generic explanation methods across contexts, the paper shows that medical AI explanations benefit from specialized consideration of each audience's needs and potential harms—a principle applicable to other high-stakes AI domains like criminal justice, finance, or employment decisions.