RICE AlgebraBot: Lessons learned from designing and developing responsible conversational AI using induction, concretization, and exemplification to support algebra learning

The importance and challenge of Algebra learning is widely recognized, with students across the U.S. facing difficulties due to the subject's complexity. While extensive research has focused on enhancing Algebra learning in K-12 education, the reusability, scalability, and effectiveness of the strategies employed (e.g., manual interventions and digital tutoring platforms) remain limited. Conversational AI (ConvAI), enabled by the advancement of large language models (LLMs), emerges as a potential tool for automatic, personalized, and effective student support. However, ethical concerns surrounding diversity, safety, sentiment, and stereotype associated with ConvAI are prominent, and empirical studies examining its application in education are scarce. The purpose of this study is to develop a ConvAI system that mitigates the potential ethical concerns and empirically evaluate the effect of such a system for math learning. Specifically, we first examined computational strategies to mitigate the ethical concerns of ConvAI in educational setting with educational big data (npretraining = 2,097,139) and found that researchers could effectively enhance ConvAI responsibility through the investigated algorithmic strategies. Then, a ConvAI system was constructed using these strategies, guided by learning sciences principles. Lastly, we examined students' eye-tracking patterns, acceptance, and learning processes when using this ConvAI system to learn Algebra through a random experiment (nparticipant = 151). Participants using the developed ConvAI demonstrated generally increased visual attention levels as compared to the control group. Moreover, participants expressed a positive acceptance towards the ConvAI technology. Finally, participants' interaction patterns with the ConvAI technology influenced their Algebra learning. These results provide insights for both educational researchers and practitioners to integrate ConvAI in learning environments. •Algorithmic strategies, such as style control and counterfactual role reversal, effectively enhance large language models' capabilities in promoting diversity, safety, positive sentiment, and anti-stereotype content.•Users of conversational AI demonstrate greater visual attention compared to those in the control group.•Features grounded in induction, concretization, and exemplification are perceived as particularly helpful by participants.•High-performing participants used ConvAI to enhance Algebra learning through inter