Coding expertise is going to collapse from AI reliance

TL;DR Highlight

Novices who leaned on AI for code ended up with confidence but little understanding

Who Should Read

A team lead setting AI coding policy for juniors, or a developer worried about their own skill atrophy

Core Mechanics

Only experts get the big AI payoff
- AI coding tools help most those with enough experience to steer, audit and verify output.
- The less experience you have, the easier it is to be misled.
Novices didn't use AI like a tutor
- In a novice-programmer study cited by JetBrains, heavy AI users skipped planning stages.
- They finished with an 'illusion of competence' rather than understanding.
Novices who limited AI did best
- Those who limited or ignored AI had 'negative expertise': the ability to ignore wrong suggestions.
- They used AI to speed up code they already intended to write.
UPenn: AI math learners did 17% worse
- In a 2025 study of 1,000 students, those who used AI as a crutch scored 17% worse than the textbook-only group.
- The GPT Tutor group did 127% better in AI-assisted practice but about the same on the test as textbook users.
Friction builds expertise

Community Reactions

Seen at enterprise level already
- One commenter said leadership mandates against hand-coding mean code is produced faster than humans can review it.
- Another said AI-inflated Jira tickets make filtering AI artifacts a big part of the job.
Guided coding beat agentic coding
- A developer with 15+ years said using an editor with a fast flash model while writing and reviewing code was as productive as vibe coding with much higher quality.
- They said headless agents accumulate defects until trivial changes cost millions of tokens.
Expert novices can't root-cause incidents
- A commenter said expert novices at their company cause serious incidents and can't root-cause them alone.
- They become a proxy for the LLM, so their fix ability is capped by the LLM's.
Counterpoint: it may not matter
- One commenter said they learned more from LLMs than they lost and now use a wider range of tools.
- Others argued the skill may fade like horsemanship, or that friction-seekers will just find it elsewhere.

How to Apply

Get quizzed by an LLM before opening a PR
- Give the LLM today's diff and ask it to question your understanding.
- Reread and rework any part where you stall.
Put a 3-question check before using AI
- Ask whether you could do the task without AI and whether you could explain the output under audit.
- If either answer is no, switch from generation to explain-and-question mode.
Use tutor mode when learning something new
- Tell the model to give only hints and questions, not full code.
- Verify results against official docs and by running them yourself.
Add hand-coding zones to junior onboarding
- In team guidelines, require writing unfamiliar layers by hand and using AI only for review.
- Explicitly allow delegating rote repetitive tasks to AI.

Try It

copy & run

# Paste into Claude or ChatGPT (simplified version of the article's 'Friction First' approach)

You are a Socratic tutor, not a code-writing tool.
Rules:
1. Never give complete code.
2. First ask for my approach plan and point out missing steps via questions.
3. If I'm stuck, give one hint at a time.
4. When I paste code, ask 3 questions about why each line is needed instead of giving answers.
5. At the end, tell me what to verify in the official docs.

What I want to learn: {{topic}}
My current attempt: {{code or plan}}

Next step: after finishing, ask "Test my understanding with 5 questions."

Terminology

skilled orchestrator paradoxThe skills needed to manage AI agents well are the ones that erode when you keep using those agents.

negative expertiseThe ability to recognize and ignore wrong or unhelpful AI suggestions, like a driver who notices when the GPS is wrong.

cognitive debt / offloadingDebt means handing over your judgment and decisions; offloading means delegating only mechanical or tedious work.

Socratic sparring partnerA practice partner that draws out your thinking through questions instead of giving answers.

Law of Leaky AbstractionsJoel Spolsky's law that every abstraction eventually leaks, so you need to understand the underlying mechanics to handle it.