Using LLMs for tutoring poses several critical risks — most notably, hallucinations. But what if we leverage the feeling of uncertainty around whether and how LLMs might make mistakes to promote metacognitive monitoring? In this project, we design a LLM-powered learning platform that manipulates model behavior (accuracy) and model confidence (expression of confidence) to test whether and how having the feeling of uncertainty and unreliability towards AI might foster learners’ critical and metacognitive monitoring of AI response and their own learning processes.
