AI makes you faster but not smarter — unless you use it as a tutor

An Anthropic-backed randomized experiment (arXiv:2601.20245, also covered on Anthropic’s research blog) tested developers learning Python’s Trio async library with and without AI assistance. The results are striking: AI users scored 17% lower on conceptual understanding and didn’t even finish faster.

Debugging skills took the biggest hit — AI removed the productive friction that forces you to actually understand what’s happening.

The 6 usage patterns they identified, ranked:

High scorers (quiz > 70%):

1. Ask only conceptual questions — never “write this for me”
2. Generate code, then ask “explain this line by line”
3. Mix generation with explanation requests

Low scorers (quiz < 40%):

1. Heavy code generation, paste directly
2. Let AI fix all errors (iterative AI-led debugging)
3. Minimal engagement — just copy output

The rule: Use AI as a tutor (ask why), not a ghost writer (just paste). The difference between learning and outsourcing is one follow-up question: “Why does this work?”

AI-enhanced productivity is not a shortcut to competence.