When ChatGPT joins the team: a mixed-methods study of AI-mediated collaborative lesson design

Introduction The application and influence of artificial intelligence (AI), and specifically Large Language Models (LLMs), in educational processes is widely discussed. However, there remains a gap in research on using LLMs as peer-like contributors in collaborative learning contexts.Methods This article reports a mixed-methods quasi-experimental study investigating how positioning ChatGPT as a peer-like feedback provider shapes student-teachers' learning and collaboration during group lesson-design activities. The study employed a counterbalanced crossover structure for knowledge assessment and a sequential two-task design for authentic artifact production. A total of 102 teachers in training (M_age = 38.87, SD = 8.01), organized into 21 groups, completed two authentic design tasks within a single session.Results Across the session, students progressively adapted to AI interaction, refining how they queried the model and how they evaluated and integrated its suggestions. Results indicate a Post-Withdrawal Sustained Performance (PWSP) effect: improvements observed during AI-available phases were not followed by a detectable decline in the immediately subsequent AI-withdrawn phase within the study timeframe. This pattern was clearest for technology-related knowledge and was consistent with stable artifact quality after AI removal. While ChatGPT support increased efficiency and contributed to technology-focused insights, qualitative evidence also pointed to tensions, including reduced peer-to-peer idea-building in some groups and concerns about creativity.Discussion Overall, the findings suggest that integrating LLMs as a feedback team-mate can support collaborative design work without immediate post-withdrawal performance costs, particularly when learners are scaffolded to engage critically with AI output rather than accept it unreflectively. These results carry implications for the design of AI-enhanced collaborative activities, highlighting the need to balance AI efficiency gains with sustained opportunities for authentic peer dialogue.