Collaboration Reimagined: How Can AI Transform Group Learning?

Abstract

Group learning fosters critical thinking, deeper understanding, and creative problemsolving. However, its success depends on balanced engagement among students. This is often a challenge in educational settings, where varying levels of prior knowledge and participation can lead to unequal contributions and diminished learning outcomes. The rise of AI in education, specifically for computer science education, presents new opportunities to address these challenges, but most research has focused on individual learning, leaving a gap in understanding how AI can actively support collaboration in group dynamics. This work explores how AI can transform collaborative learning by identifying and addressing imbalances in group engagement. Using Pytutor, a large language model (LLM) developed at MIT to support students in learning computer science, this research examines how students utilize AI specifically in the collaborative learning process. By analyzing these interaction patterns and evaluating student learning, this work will contribute to the development of adaptive AI systems that foster more inclusive and effective group learning environments. Findings suggest that students do enjoy collaborating on programming, especially when partnerships are balanced and mutually participatory. Many students reported that their partner helped to clarify their thinking, kept them engaged and reduced the stress of programming. However, students also reported frustrations when collaboration felt uneven or when the division of labor was unclear. AI was used by students in a variety of ways, often to offload individual debugging or syntax questions while maintaining an active peer discussion. Some students highly integrated the AI chat into their collaboration process, while others used it more parallely. Students generally found AI to be most helpful when it gave them hints, next steps, or debugging advice, rather than complete answers. Notably, students who collaborated actively rated both AI and their peer as helpful, suggesting that social dynamics amplify the effectiveness of the AI tutor.