Skill atrophy from AI is real — a 2026 study found AI-assisted developers scored lower on comprehension and spent more time debugging than coding. The fix is deliberate practice: write 1-2 hours of code per week with no AI, review your own diffs without Copilot, and run spaced-repetition drills on the fundamentals you delegate most, as Cal Newport prescribes for any cognitive skill you want to keep.
I run an engineering team and use Claude Code, Cursor, and Copilot every working day. I am also watching the engineers I respect most slowly become worse at the parts of the job they have offloaded. Both things are true at once, and the honest answer to "how do you stop AI from quietly atrophying your engineering skills" is that you have to design against it on purpose. Nobody is going to do it for you, and the productivity gain is real enough that you will not feel the loss until it is already expensive.
The data is no longer ambiguous. A widely-cited 2026 study summarised on dev.to under the title "The Skill Atrophy Crisis" reported that experienced developers using AI assistance scored measurably lower on code comprehension tests, and spent more total time debugging AI-generated code than they would have spent writing the equivalent by hand. Anthropic's own research on AI-assistance and coding skills, published earlier this year, found a similar pattern with junior developers — the more they leaned on the model during learning, the worse they performed on transfer tasks where the model was absent. Ethan Mollick, in Co-Intelligence, calls this the "jagged frontier" problem from the user side: the tool is brilliant at things you cannot evaluate, so you stop evaluating, and your taste decays without you noticing.
What I actually do, and what I ask my team to do, is borrow Cal Newport's framing from Deep Work. Newport's point is that any cognitive ability you want to keep requires deliberate, distraction-free practice on the edge of your current capability. When AI handles the easy 80%, the practice volume on the hard 20% silently collapses, because you stop encountering it. So I block two unbroken hours each week — usually Friday morning — where I write or refactor something non-trivial with no AI assistant open. No Cursor, no Claude tab, no autocomplete beyond the editor's built-in symbol completion. The goal is not productivity. The goal is to notice what I have forgotten.
The second move is review hygiene. AI-generated code reviews like a brochure: it looks correct, it reads smoothly, and the eye slides over it. So I read AI-authored diffs the slow way — out loud, function by function, with the tab containing the AI conversation closed. If I cannot explain in a sentence why a line is there, the line gets rewritten or deleted. This is exactly the inversion Daniel Kahneman warns about in Thinking, Fast and Slow: when System 1 fluency makes the output feel obviously right, you have to deliberately invoke System 2, because System 2 will not show up on its own.
The third move is the one I had not seen tooling for until very recently. A small launch from this month, Fata (fata.dev), pitches itself bluntly as "spaced repetition to fight skill rot from AI coding." The premise is correct even if you never use the product: identify the concepts you delegate most often — async error handling, SQL window functions, whatever your stack's hard parts are — and put them on a spaced-repetition schedule with hand-written drills. The forgetting curve does not care that an LLM remembers for you. I run a simple version of this with a markdown deck and Anki for myself, and I have started asking new engineering hires to keep their own deck for the first ninety days, because that is when the dependency forms.
The failure mode I want to name explicitly is the one Mollick is too polite to belabour. You will not feel the atrophy. You will feel faster, more confident, and more productive — right up until the moment you face a problem the model gets wrong and you cannot tell. The signal that the routine is working is not that you got more done this quarter. It is that, in the quiet two hours each Friday, the unaided version of you is still recognisably good at the work. If that version is getting visibly worse, no shipping metric is going to compensate. Honest practice with the tools off is the only thing that does.
Related: How to Find Your Passion · Best Self-Improvement Books · How to Make Better Decisions · What University Will Not Teach You