1　Introduction

Generative Artificial Intelligence (GAI) has exerted transformative impacts on the ecology of foreign language education. It is noted that GAI Chatbots provide technical support for teaching, including real-time interaction, multilingual recognition, and multimodal generation, thereby promoting the development of teaching toward personalization, immediacy, and contextualization. On this basis, it is proposed that foreign language education needs to break through monolingual discourse patterns and respect learners' individual needs. Addressing the problem of limited learner agency in traditional teaching, this study clarifies how GAI can expand translanguaging spaces and stimulate learners' subjective agency, points out the research gap regarding insufficient attention to learners' translanguaging forms and their agency manifestations, and thus establishes the focus of this study on domestic English Medium Instruction (EMI) classrooms to explore the ways students engage in translanguaging practices with GAI Chatbots and the characteristics of their agency.

2　Literature Review

This section reviews three core topics: translanguaging theory has evolved from describing bilingual phenomena to becoming a critical theory, emphasizing language as a dynamic multimodal repertoire that facilitates integrated language and content learning in EMI classrooms, though its application remains controversial in academic circles; learner agency possesses characteristics such as sociocultural mediation and situatedness, with diversified mediational resources promoting its development, while research on its integration with information technology remains in the exploratory stage; GAI's multilingual and multimodal characteristics are highly compatible with translanguaging, capable of supporting learners' autonomous translanguaging practices, opening up non-teacher-led learning spaces, and promoting personalized learning processes and stimulating agency, yet how to effectively realize its educational potential still requires classroom-based empirical investigation.

3　Research Design

This study presents a four-month research design in an EMI course at a university, centered on two core questions: the ways of translanguaging practices in learners' interactions with GAI Chatbots, and the forms of learner agency manifested in classroom activities. The research subjects were 50 undergraduate students, from whom three types of qualitative data were collected during four classroom activities—approximately 550,000 words of human-machine interaction records, 40 open-ended feedback responses, and 28 reflective journals. Thematic analysis was employed for data analysis: for research question (1), a deductive-inductive hybrid coding approach was applied, using "multilingualism" and "multimodality" as preset frameworks for translanguaging analysis; for research question (2), inductive coding was conducted to extract agency characteristics from students' expressions of actively setting goals and mobilizing resources.

4　Findings and Discussion

This section presents the specific manifestations of translanguaging practices and the multidimensional characteristics of learner agency in GAI-empowered English Medium Instruction (EMI) classrooms. At the multilingual interaction level, four phenomena were identified: language compensation—students using their native language to substitute for target language expressions when content is restricted to ensure communicative coherence; terminology transfer—directly citing English professional terminology to maintain conceptual accuracy; emphasis highlighting—deliberately embedding foreign language adjectives or adverbs to strengthen emotional and informational focus; and cognitive simplification—unconsciously switching between languages to reduce cognitive load, aligning with the "cognitive miser" theory. At the multimodal resource utilization level, emojis serve emotional support functions, images carry local cultural elements, while video links and music expand communicative dimensions. Students achieve meaning construction through multimodal collaboration, such as guiding GAI to generate bilingual poetry integrating Chaoshan cultural imagery. These translanguaging practices collectively constitute the "GAI translanguaging space," breaking through teacher-dominated models, weakening inequalities arising from language proficiency differences, and providing students with an equal and open autonomous learning environment that promotes integrated language and content learning.

This section also explores four dimensions of learner agency: autonomous planning and regulation—students setting learning plans according to personal goals and preferences and making dynamic adjustments; critical thinking and feedback—students examining and evaluating the applicability of GAI-generated content, questioning scoring deviations, and proposing improvement suggestions; resource integration and utilization—students actively mobilizing mediational resources such as music, videos, and games, and even exploring multiple GAI tools to meet their needs; and situational awareness and adaptation—students acutely perceiving language and cultural differences and adjusting interaction strategies to adapt to specific communicative situations. These findings indicate that learners in GAI environments have transformed from passive recipients of information to subjects who autonomously construct knowledge, providing theoretical expansion and practical reference for understanding agency development in technology-empowered education.

5　Conclusion

This study presents the findings from four activities integrating GAI Chatbots into EMI classrooms, revealing that they construct an open and inclusive GAI translanguaging space. Within this space, students autonomously engage in translanguaging practices, demonstrating four types of multilingual interaction phenomena: language compensation, terminology transfer, emphasis highlighting, and cognitive simplification, while utilizing multimodal resources for meaning construction. Learner agency is manifested in four dimensions: autonomous planning and regulation, critical thinking and feedback, resource integration and utilization, and situational awareness and adaptation. Three pedagogical implications are proposed: utilizing the GAI translanguaging space to alleviate the contradiction between EMI monolingual policies and practical needs, stimulating learner agency and establishing interactive archiving feedback mechanisms, and being vigilant against negative technological impacts while emphasizing the irreplaceability of human original capabilities. Finally, the limitations of not measuring participants' language proficiency and collecting only four sets of qualitative data are acknowledged, with longitudinal tracking studies recommended for future research.