Tutor, Not Solver: Designing a Guardrailed AI Assistant for Learning in Higher Education (PeteChat)

Created: 2026-06-10 | Tags: intelligent-tutoringhigher-edllmacademic-integrityscaffolding

Belle Li, Lily Tan, Wei Zakharov, Qiang Qiu, Colby Ben Acton โ€” Purdue University โ€” cs.HC, cs.ET ๐Ÿ“„ Full text (arXiv)

PeteChat is a course-aligned AI tutor developed and deployed at Purdue University, documented through design-based research (DBR). Drawing on literature-informed design inputs, pre-deployment baseline analysis of student-system interactions, and formative expert evaluation with teaching assistants and UX/developer stakeholders, the paper reports eight transferable design principles for assessment-aware AI tutors. These include homework guardrails (preventing answer-giving while allowing help), debugging scaffolds (guiding students through error resolution), self-regulated-learning support (prompting metacognitive reflection), and instructor-facing customization tools. The system is built on a locally hosted Llama-3 model with RAG grounded in course materials. The design principles and methodological approach offer actionable guidance for institutions deploying responsible, integrity-preserving AI tutors at scale. This work directly connects to debates in ai-literacy and academic-integrity about maintaining assessment validity while providing AI assistance, and extends intelligent-tutoring principles to the LLM era.

Related Pages

Citations

APA: Belle Li, Lily Tan, Wei Zakharov, Qiang Qiu, Colby Ben Acton (2026). Tutor, Not Solver: Designing a Guardrailed AI Assistant for Learning in Higher Education (PeteChat). arXiv:2606.09845. Purdue University โ€” cs.HC, cs.ET.