I just read a wild study from Anthropic. They had developers learn a new Python library—half with AI help, half without.
The AI group finished 2 minutes faster.
But scored 17% lower on the test. Nearly two letter grades.
Here's what hit me: the biggest gap wasn't writing code. It was debugging. You know, that thing you need when AI hallucinates a API call.
What Separated the Winners
The devs who did learn didn't use AI less. They used it differently:
- Asked "why does this work?" after getting code
- Requested explanations alongside solutions
- Posed conceptual questions while coding solo
The ones who failed treated AI like Stack Overflow on steroids—copy, paste, pray.
The Real Talk
AI is incredible for stuff you know. Anthropic's own data shows 80% speedups on familiar tasks.
But learning? That's different. Learning requires friction. Getting stuck. That 20-minute rabbit hole that finally clicks.
AI skips the rabbit hole. You get the code, miss the lesson.
My Take
I'm not ditching Claude. But I'm changing how I use it.
Generate → Then comprehend.
Ask one extra question. Type it out yourself once. Suffer through the error message before asking for help.
Your debugging skills are the ceiling on how much you can trust AI. Don't let that ceiling drop.
What about you? Has AI made you faster, or has it made you think less?
Top comments (0)