AI Explanations as Boundary Objects: Toward Shared Language in High-Stakes Decision-Making

Explainable AI (XAI) research increasingly recognises that different stakeholders hold distinct explainability needs, yet much of this work designs explanations for individual groups in isolation. In high-stakes domains, however, AI-informed decisions must flow between parties with asymmetric knowledge, authority, and stakes, and translational breakdowns emerge when explanations are not designed with this communicative chain in mind. Drawing on team cognition, psycholinguistics, and boundary object theory, and grounded in a case study of diagnostic communication, we argue that the language of AI explanations is a foundational design parameter shaping cross-stakeholder understanding, trust, and accountability. We offer three recommendations for reframing explanation design as a consensus-building activity comprising participatory vocabulary development, iterative explanation negotiation, and ongoing governance. We further propose evaluation metrics including interpretive alignment, translational load, and contestability that shift focus from individual comprehension to cross-stakeholder communicative quality. We contend that the central challenge for human-centred XAI may not be making AI more explainable to individuals, but designing explanations that help people explain, deliberate, and decide together.