PhD Proposal: Towards Trustworthy AI-Mediated Communication Across Languages and Cultures

Artificial Intelligence (AI) tools are increasingly used worldwide to mediate communication across languages and cultures. They support direct interactions—such as helping two people without a shared language converse via translation systems—and indirect interactions—such as assisting users interpret unfamiliar idioms or cultural practices. Since these interactions actively shape how people perceive, interpret, and act upon in various decision-making scenarios, ensuring that AI-mediated cross-lingual and cross-cultural communication is trustworthy is critical. This proposal outlines a research agenda emphasizing two complementary requirements. First, multilingual AI systems must possess and appropriately use knowledge across all languages and cultures to help establish common ground between users from different backgrounds. Second, users must be supported in making informed decisions about when to rely on system outputs, especially when they lack the means to independently assess them. By integrating both model- and user-side perspectives, we aim to advance trustworthy AI-mediated communication across linguistic and cultural divides.
Part I of the proposal focuses on model-side workflows that evaluate and enable AI systems to act as trustworthy communicative mediators. This requires meeting two key conditions. First, models should ground their outputs in evidence that provides users with equal access to knowledge across languages. We design a method for measuring how models rely on multilingual evidence in retrieval-augmented generation and find a strong preference for English sources—even when they are not relevant to the user query. This suggests that evaluating outputs alone is insufficient; achieving knowledge parity requires understanding how models arrive at their predictions. Our proposed work aims to study intermediate reasoning processes across multilingual inputs, focusing on how models reason and converge on predictions in different languages. Second, models should adapt their responses to diverse cultural contexts. In practice, however, model outputs often by default align with the dominant cultures and languages represented in their training data. To address this, we propose a multi-turn, multi-model collaboration framework that moves beyond the conventional single-turn, single-model approach, fostering more equitable cultural competence.
Enhancing model-side workflows is necessary but not sufficient, as even improved models will produce imperfect outputs that users must make decisions on. Part II addresses user-side workflows, examining how users perceive, evaluate, and act upon system outputs, and designing decision-support mechanisms that promote more reliable interaction in cross-lingual or cross-cultural scenarios. We begin with a common yet challenging scenario in which monolingual users must judge whether machine translation (MT) outputs are good enough to share, despite not understanding the source language. To aid this process, we propose a new quality estimation approach that leverages question-answer (QA) pairs, making quality signals more actionable and interpretable for monolingual users. We then conduct controlled evaluations and human studies to assess how effectively such feedback guides real users' decision-making and reliance on MT outputs. Beyond language understanding, users navigating unfamiliar cultural contexts also need support. In our proposed work, we investigate how AI feedback can help non-native speakers communicate newly emerging words accurately and naturally with native speakers.
Collectively, our research agenda establishes a comprehensive framework for trustworthy AI-mediated communication across languages and cultures. By uniting model-side workflows that advance knowledge parity with user-side workflows that provide informed and reliable decision-support for acting upon AI outputs, we aim to build multilingual AI systems that effectively serve diverse real-world users.