Designing Against Deskilling: Metacognitive Feedback Reduces Cognitive Offloading to LLM Assistants

TL;DR - A preregistered study found that metacognitive feedback reduced users’ cognitive offloading to an LLM tutor and improved later unaided performance, suggesting AI assistants can be designed to mitigate deskilling without restricting access.

• The online experiment involved 704 participants practicing fraction arithmetic, followed by an unaided test.
• Metacognitive feedback reduced answer offloading to the LLM assistant (OR = 0.47).
• The feedback also improved test performance (OR = 1.51).
• Effort-based rewards showed no evidence of affecting offloading or test outcomes.