The AI-and-atrophy debate is stuck on the wrong question. Whether AI can erode your skills is settled: it can. The variable nobody prices is that erosion is a decision you make in how you use the tool, and the same tool, pointed the other way, makes you sharper.
Most of what I read about AI and skill decay is defensive. Protect the struggle. Don’t outsource thinking. Steer atrophy toward the skills that don’t matter. All true, and I’ve written versions of it myself. But it all shares a hidden assumption: that AI is a hazard to be contained, and the best you can do is lose slowly.
I no longer think that’s the ceiling. AI doesn’t decide whether you keep improving. You do. The tool is neutral. It will happily run either loop you build with it, and the two loops diverge harder every month.
The Two Loops
There is an erosion loop and a sharpening loop, and most people fall into the first by default because it’s the path of least friction.
The erosion loop is familiar: prompt, paste, accept. You describe the problem, the model returns 200 lines, you skim, you nod, you merge. It feels like velocity. What you’ve actually done is skip the step where your brain builds the connections between patterns. You got the output and missed the compounding. Do that for a year and you’ve had a year of not exercising the thing that made you valuable.
The sharpening loop uses the exact same tool for the opposite purpose. You make the model challenge your assumptions, expose the gaps in your reasoning, generate alternatives you wouldn’t have thought of, and drill you on the parts you’re weakest at. Same model, same session, and one path atrophies you while the other trains you. The difference is entirely in what you ask it to do.
Why AI Is an Unfair Deliberate-Practice Machine
Here’s the part the “contain the damage” framing misses entirely. Deliberate practice, in the Ericsson sense, needs three things that are historically expensive: you have to operate at the edge of your ability, you need immediate feedback, and you need targeted repetition on your specific weaknesses. A good coach supplies all three, which is why coaching is scarce and expensive.
An LLM supplies all three on demand, for free, at 2am. That is not a small thing. It’s the first time in history that a personal, infinitely patient sparring partner has been available to every engineer at once. Treating that purely as a threat to your skills is like treating a gym as a threat to your muscles.
The edge-of-ability part matters most. Left alone, I solve problems at the level I’m already comfortable with, because comfort is efficient. A model I’ve told to push me will hand me the harder version, poke the assumption I was leaning on, and refuse to let me coast. That’s the mechanism that actually builds capability, and it’s the one the erosion loop quietly removes.
How I Actually Run the Sharpening Loop
This is not theoretical. These are the moves I use to keep the tool on the training side of the line.
Make it steelman the position I’m about to reject. Before I commit to an architecture, I ask the model to argue the hardest case for the two approaches I’m not taking. Half the time it surfaces a constraint I’d waved away. The other half I understand my own choice better because I had to beat a real opponent, not a strawman.
Generate three, defend one. I ask for three genuinely different solutions, then I pick and I write the defense myself. The generation is cheap and divergent; the judgment is the rep. If I can’t articulate why the winner beats the other two, I don’t understand the problem yet, and that gap is the signal.
Have it quiz me, not answer me. When I’m learning something unfamiliar, I tell the model to withhold the answer and interrogate me instead. Ask what I think happens, tell me where I’m wrong, make me try again. It’s the Socratic tutor I could never otherwise afford.
Explain the code back before I merge it. If the model generated something, I explain it back, out loud or in the PR description, without looking. The moment I can’t is the moment I’ve found code I approved but don’t own, and I go back and fix that before it ships.
None of these is slower in any way that matters. They cost five extra minutes and buy the compounding the erosion loop throws away.
The Honest Part
I want to be clear that the erosion loop is the default, and defaults win. The sharpening loop takes deliberate friction, and friction is exactly what you have least of when you’re tired and the deadline is real. On those days I let the model write something I should have thought through, and I know I’m taking on a small debt when I do it.
So this is not a claim that AI automatically makes you better. The tool is an amplifier, and an amplifier points wherever you aim it. Aim it at your own laziness and it will erode you faster and more comfortably than anything before it. Aim it at your weaknesses and it will train you faster than any tool before it, too. The leverage cuts both ways with equal force, which is precisely why the choice is the whole game.
The people who come out of this era sharper won’t be the ones who used AI least, guarding their skills like a candle in the wind. They’ll be the ones who pointed the most powerful practice tool ever built directly at the edge of their own ability, on purpose, every day.
Skill erosion from AI is real. It is also optional. That’s the entire point.
Related
- The Struggle Is the Product – Why the difficulty is the thing that builds capability, and how to protect it
- AI Leverage Without Skill Atrophy – Strong vs weak usage, and the operator/engineer divergence
- Skill Atrophy: What to Keep, What to Let Go – Steering inevitable atrophy toward low-leverage skills
- Comprehension Debt Is Refinanceable – The debt you take on when you approve code you don’t understand
