[Let’s zoom out from OpenClaw] Two minutes saved, 17 points lost. Anthropic study shows AI assistance causes a drop in skill mastery with almost no gain in speed
They found three ways to use AI that preserve learning, and three that hurt it.
Judy Hanwen Shen and Alex Tamkin ran a randomized trial where 52 developers had 35 minutes to learn Python's Trio library with or without AI assistance, then immediately took a quiz without AI.
They found the AI assistance patterns that preserve learning:
- Generation-Then-Comprehension: participants asked AI to explain code after generating it (scored 86% on the quiz)
- Hybrid Code-Explanation: requesting code with explanations (68%)
- Conceptual Inquiry: asking only clarifying questions (65%)
These AI helps hurt:
- AI Delegation: copying AI output directly (scored 39% on the quiz)
- Progressive AI Reliance: starting independently then delegating (35%)
- Iterative AI Debugging: repeatedly asking AI to fix errors (24%)
My take:
- AI is like outsourcing. Great for familiar tasks. Dangerous when you outsource what you've never done yourself.
- Junior security engineers who let AI fix their errors skip the struggle that builds debugging and security intuition. Guess the outcome.
- Learning becomes a luxury in today's reality when AI sets the pace and the 996 norm enforces it.
How AI Impacts Skill Formation
