TY - JOUR
T1 - Embedding, Copilot, or Agent? L2 Teachers Modes of Collaboration With Generative AI in Providing Feedback on Writing Assignments
T2 - Rubric Influences, Workload Dynamics, and Barriers to Integration
AU - Wei, Wei
AU - Zhang, Xin
AU - Zhu, Xinhua
N1 - Publisher Copyright:
© 2026 The Author(s). European Journal of Education published by John Wiley & Sons Ltd.
PY - 2026/9
Y1 - 2026/9
N2 - Drawing from the established human-AI collaboration framework (Mosqueira-Rey et al., 2023), this study investigated how L2 teachers at a research-intensive university collaborated with Generative AI chatbots to provide feedback on tertiary-level written assignments over 9 weeks. Data from thirteen teachers included 278 interactions with GenAI, 876 feedback instances, and reflective accounts of their experiences. Findings revealed three Teacher-GenAI collaboration modes—embedding (human-led), copilot (interactive), and agent (GenAI-led)—with a MANOVA indicating that rubric dimensions significantly influenced their frequency of use. Post hoc comparisons indicated that teachers used the agent and copilot modes more frequently in the “language use” dimension. Conversely, the embedding mode was used most frequently in the “content” and “overall evaluation” dimensions. Although the agent mode reduced teachers' workload in providing feedback on mistakes at the micro level, the embedding mode demanded more effort from teachers to give feedback at the macro level of students' assignments. Guided by Ertmer's framework (1999) on the barriers to educational technology integration, we found that teachers tried to claim more control over human-AI collaboration by adopting embedding and copilot modes for four concerns: GenAI agents' misalignment with task requirements, inconsistency with external expectations, inadequacy in addressing students' emotional needs, and cautious attitudes toward technology.
AB - Drawing from the established human-AI collaboration framework (Mosqueira-Rey et al., 2023), this study investigated how L2 teachers at a research-intensive university collaborated with Generative AI chatbots to provide feedback on tertiary-level written assignments over 9 weeks. Data from thirteen teachers included 278 interactions with GenAI, 876 feedback instances, and reflective accounts of their experiences. Findings revealed three Teacher-GenAI collaboration modes—embedding (human-led), copilot (interactive), and agent (GenAI-led)—with a MANOVA indicating that rubric dimensions significantly influenced their frequency of use. Post hoc comparisons indicated that teachers used the agent and copilot modes more frequently in the “language use” dimension. Conversely, the embedding mode was used most frequently in the “content” and “overall evaluation” dimensions. Although the agent mode reduced teachers' workload in providing feedback on mistakes at the micro level, the embedding mode demanded more effort from teachers to give feedback at the macro level of students' assignments. Guided by Ertmer's framework (1999) on the barriers to educational technology integration, we found that teachers tried to claim more control over human-AI collaboration by adopting embedding and copilot modes for four concerns: GenAI agents' misalignment with task requirements, inconsistency with external expectations, inadequacy in addressing students' emotional needs, and cautious attitudes toward technology.
KW - collaborative strategies
KW - GenAI chatbot
KW - human-AI collaboration
KW - LLM Chatbot
KW - teacher feedback
KW - written corrective feedback (WCF)
UR - https://www.scopus.com/pages/publications/105045969213
U2 - 10.1111/ejed.70795
DO - 10.1111/ejed.70795
M3 - Article
AN - SCOPUS:105045969213
SN - 0141-8211
VL - 61
JO - European Journal of Education
JF - European Journal of Education
IS - 3
M1 - e70795
ER -