Judy Hanwen Shen, Alex Tamkin · 2026 · arXiv preprint (Anthropic) Read the source 52 mostly junior developers learned a new library with or without AI. The AI group scored 50% vs 67% on the follow-up quiz, worst on debugging. Those who asked for explanations or fixed their own errors kept their learning. Questions it bears on Offloading or atrophy Who pays the cognitive debt