Yimin Xiao | 肖亦敏

Yimin Xiao
University of Maryland, College Park
Email: yxiao[at]umd[dot]edu
Google Scholar

RESEARCH OVERVIEW AND PROJECTS

Understanding Human Agency in AI-Mediated Communication

My doctoral research started with understanding how people communicate across linguistic boundaries with AI language support. For people with limited proficiency in a target language, AI systems can function as imperfect yet necessary tools for producing and comprehending information in that language. However, individuals may experience a loss of agency. They face uncertainty about AI-generated outputs or have limited control over how their expressions are represented across languages. Meanwhile, the surface-level fluency of AI-generated language can obscure these constraints.

Our findings demonstrate that AI involvement can lead to misleading perceptions of linguistic competence and agency for both people using AI systems and their interlocutors. To develop a more precise understanding of human agency in AI-assisted language use, my research highlights the value of: (1) understanding language use as a multi-dimensional act (e.g., idea generation versus the expression of those ideas), and (2) examining the interactional processes through which people produce and comprehend language alongside AI systems.

Related Publications

(Dis)placed contributions: uncovering hidden hurdles to collaborative writing involving non-native speakers, native speakers, and AI-powered editing tools.
Yimin Xiao, Yuewen Chen, Naomi Yamashita, Yuexi Chen, Zhicheng Liu, Ge Gao.
ACM CSCW 2024 [paper]

Comparing native and non-native English speakers' behaviors in collaborative writing through visual analytics.
Yuexi Chen, Yimin Xiao, Kazi Zinat, Naomi Yamashita, Ge Gao, Zhicheng Liu.
ACM CHI 2025 [paper]

Toward machine translation literacy: How lay users perceive and rely on imperfect translations.
Yimin Xiao, Yongle Zhang, Dayeon Ki, Calvin Bao, Marianna J. Martindale, Charlotte Vaughn, Ge Gao, Marine Carpuat.
EMNLP 2025 Main [paper]

Physician detection of clinical harm in machine translation: Quality estimation aids in reliance and backtranslation identifies critical errors
Nikita Mehandru, Sweta Agrawal, Yimin Xiao, Elaine C Khoong, Ge Gao, Marine Carpuat, Niloufar Salehi
EMNLP 2023 🏆 Outstanding Paper Award [paper]

Interaction Design for Sustaining and Modulating Agency

Building on the understanding of human agency, my research progressed toward design interventions for supporting human agency in AI-assisted language use. Drawing on Bandura’s social cognitive theory of agency, we derived a resource–agency link as a core design principle: people’s agency depends on the resources they leverage for action. We designed human-AI interfaces to structure the extent of resources, i.e., linguistic knowledge, individuals draw on. Specifically, we modulated people’s resource involvement by prompting people to evaluate AI-outputs (primarily involving language comprehension) versus edit outputs (involving both comprehension and production).

Our empirical findings demonstrated that enabling personal resource involvement sustained people’s agency in AI-assisted language production. However, the higher agency achieved through increased resource involvement also came with reduced communication quality. These findings suggest that exercising agency at the margins of one’s existing capabilities can become cognitively and practically costly.

Related Publication

Sustaining human agency, attending to its cost: An investigation into generative AI design for non-native speakers' language use.
Yimin Xiao, Cartor Hancock, Sweta Agrawal, Nikita Mehandru, Niloufar Salehi, Marine Carpuat, Ge Gao.
ACM CHI 2025 [paper]

AI-Assisted Capability Development for Expanding Agency

My ongoing projects extend this line of research toward supporting human capability development. I position human-AI interaction as a form of scaffolding for thinking and social action, e.g., AI repeatedly modeling a structured thinking flow. I investigate how people can acquire new language resources and develop more independent practices over time, alongside immediate task assistance from AI systems.