Metacognitive feedback cut answer offloading to an LLM assistant roughly in half.

arXiv 2609.20143 runs a preregistered 2x2 plus no-AI control with 704 participants practicing fraction arithmetic with an assistant that gave solutions only on explicit request, then tested them unaided. Feedback making the implications of offloading explicit reduced answer offloading (OR 0.47) and improved test performance (OR 1.51). An effort-based reward incentivizing less extensive assistance showed no evidence of affecting either. arXiv The null result on incentives is the more useful half for anyone designing assistant UX.