This dissertation is an interdisciplinary investigation and analysis of trustworthy machine learning (TML) and responsible artificial intelligence (RAI) research and development. Each chapter presents a different research approach for examining safety, AI ethics, and TML education. To appropriately address socio-technical problems and assess societal risk and harms, practitioners must be trained on AI ethics, safety, responsibility, and trustworthiness early on. This work connects safety, AI ethics, and education theory to investigate useful methods for teaching RAI to undergraduates early in their academic careers. Its contributions are an abolitionist framework for critical engagement in AI ethics, a framework for trustworthy implementation development, and a curriculum to teach RAI for undergraduates.

Computer science is an important but challenging subject for many students, while the number of teachers available has been limited for many years as enrollment in computing classes continues to grow. Automated intelligent support for learning can save instructors time and provide insights about students’ knowledge, but building these tools requires modeling what students do and do not know as they work through learning content. This dissertation focuses on methods to incorporate programming code and educational theory information into deep neural networks for tasks such as bug detection and knowledge modeling. It shows how students’ code submissions can be used to build more effective and interpretable deep student models, paving the road to interpretable deep knowledge tracing on complex, multi-skill problems without massive expert effort.

The widespread growth of massive open online courses has improved access to education, making it crucial to develop technologies that serve the diverse needs of students and instructors. This dissertation proposes QUADL, a pragmatic method for generating questions aligned with specific learning objectives. In evaluation studies, questions generated by QUADL were rated on par with human-generated questions in relevance to learning objectives and outperformed a state-of-the-art model that generates questions without considering learning objectives. QUADL was developed as part of PASTEL, a framework of evidence-based learning engineering technologies for online courseware.

Students, referred to as tutors, learn by assisting a teachable agent. Although learning-by-teaching has been proven impactful, novice tutors sometimes see limited benefit due to reluctance to engage in knowledge-building. This dissertation proposes the Constructive Tutee Inquiry model, a follow-up question framework employed by the teachable agent. Through a Wizard of Oz study, a trained response classifier, and an adaptive follow-up question framework leveraging large language models, the work shows that consistently posing constructive follow-up questions trains tutors to engage in knowledge-building autonomously over time and improves both procedural and conceptual learning.

AI ethics and fairness is a growing area of research focused on interdisciplinary approaches to assessing AI’s impact on society and mitigating the harms AI can cause. Across 649 papers published at two key conferences (AIES and FAccT) from 2018 to 2022, only 13 percent explicitly mention a system of oppression, and only 16 percent of those meaningfully engage with it. This demonstrates a gap in understanding that impacts researchers’ ability to situate AI ethics and fairness research in socio-technical reality. This research proposes the ACESOR rubric to help AI ethics and fairness researchers bridge that gap and situate their work in explicit acknowledgment of systems of oppression, following a design research methodology.