Abstract:
Recent advances in large language model (LLM) chatbots have led to their increasing use in computing education, particularly for feedback, assessment support, and introductory programming assistance. While these systems offer clear benefits in terms of scalability and personalised learning support, the literature highlights persistent limitations, including shallow pedagogical reasoning and the continued need for human oversight in educational settings. Current approaches increasingly favour retrieval-augmented generation (RAG) architectures and course-specific tutoring systems, which ground responses in curated instructional material and constrain model behaviour to improve reliability and contextual relevance.
Despite these technical advances, there is still limited understanding of how AI-supported learning environments shape student experience, particularly in relation to gender equity, confidence, and persistence in computing. This is significant given persistent underrepresentation of women in computing and growing evidence that AI systems can reproduce and amplify social stereotypes in ways that affect learner engagement and belonging.
This work-in-progress paper synthesises evidence across AI chatbot architecture, gender bias in model outputs and chatbot design, and mitigation approaches, and argues that technical grounding alone does not guarantee equitable learning outcomes. It proposes a socio-technical model for gender-responsive LLM-based AI teaching assistants that uses RAG architecture. The proposed model is centred on six design considerations: inclusive knowledge architecture, stereotype threat mitigation, neutral persona design, communication functionality, participatory co-design, and integration within wider intervention ecosystems.
The paper further argues that AI teaching assistants should be evaluated not only in terms of accuracy and learning gain, but also in terms of their impact on confidence, engagement, and persistence across diverse student groups. It concludes that equity in educational AI is not a post-hoc adjustment, but a property that must be designed into both the knowledge base and the interaction layer of the system.
Despite these technical advances, there is still limited understanding of how AI-supported learning environments shape student experience, particularly in relation to gender equity, confidence, and persistence in computing. This is significant given persistent underrepresentation of women in computing and growing evidence that AI systems can reproduce and amplify social stereotypes in ways that affect learner engagement and belonging.
This work-in-progress paper synthesises evidence across AI chatbot architecture, gender bias in model outputs and chatbot design, and mitigation approaches, and argues that technical grounding alone does not guarantee equitable learning outcomes. It proposes a socio-technical model for gender-responsive LLM-based AI teaching assistants that uses RAG architecture. The proposed model is centred on six design considerations: inclusive knowledge architecture, stereotype threat mitigation, neutral persona design, communication functionality, participatory co-design, and integration within wider intervention ecosystems.
The paper further argues that AI teaching assistants should be evaluated not only in terms of accuracy and learning gain, but also in terms of their impact on confidence, engagement, and persistence across diverse student groups. It concludes that equity in educational AI is not a post-hoc adjustment, but a property that must be designed into both the knowledge base and the interaction layer of the system.