Scaffolding students-AI dialogue for safe educational interactions

Abstract Adolescents are the fastest-growing and largest age group adopting Large Language Models (LLMs), tools not specifically designed for their educational, emotional, and developmental needs. Safe, pedagogically grounded AI interactions are essential for both AI literacy and student wellbeing. We propose the Steered Contextual AI Framework for Orchestrating Learning Dialogue (SCAFFOLD), a layered reliability architecture that surrounds LLM-generated text and speech with external verification, targeted repair, and safe fallback, while preserving data privacy. Grounded in the science of learning, including the principle that learning requires effort, SCAFFOLD is designed to preserve productive effort, supporting active human-AI collaboration. It distinguishes deterministic from probabilistic checks and can be applied to any LLM. We piloted a minimal SCAFFOLD prototype in a classroom deployment with 12–16-year-old students using an LLM-powered social robot in a multi-user co-creation interaction, in which students collaboratively designed a mnemonic on a previously covered topic. Students first interacted with a prompt-only LLM and then with the SCAFFOLD prototype during a learning intervention. Preliminary results suggest greater student activity, engagement, and on-topic participation with the SCAFFOLD prototype. Co-creation level independently predicted post-test knowledge scores after controlling for prior knowledge, consistent with a positive role of active co-creation in learning. However, the comprehension tracking frame failed to accurately assess student understanding, highlighting the difficulty of automated comprehension assessment. These pilot findings provide initial evidence for SCAFFOLD's feasibility and potential to support engagement and learning in educational settings.