Short answer: the honest research says something more specific and more interesting than a flat yes or no — AI coding tools appear to trade measurable skill for speed that, in at least one major study, didn't actually materialize.
Here's what's actually been measured, not guessed at.
Study one: the comprehension gap
In a randomized controlled trial published January 29, 2026, Anthropic split 52 professional developers into two groups learning a new asynchronous programming library — one group used AI assistance throughout, the other coded entirely by hand — then tested both on comprehension afterward. The AI-assisted group averaged 50% on the quiz. The manual group averaged 67% — with the steepest decline specifically in debugging, the skill most dependent on understanding why code behaves the way it does, not just that it runs. The speed advantage that's supposed to justify the trade barely existed — AI users finished about two minutes faster on average, not a statistically significant difference.
Study two: the perception gap
A separate study from METR — an independent AI research lab — found something arguably stranger. Between February and June 2025, researchers gave 16 experienced open-source developers real tasks from repositories they'd worked in for an average of five years, comparing task completion with and without AI assistance. Developers using AI took 19% longer to finish — not less experienced developers struggling with unfamiliar code, but experts working in codebases they already knew well. The stranger part: afterward, those same developers estimated they'd been about 20% faster with AI — a roughly 39-point gap between how fast they felt and how fast they actually were.
That perception gap is the part worth sitting with longer than the speed numbers. If experienced developers can't accurately tell whether AI made them faster or slower on their own familiar code, the same blind spot almost certainly applies to judging whether it's eroding a skill in the background. You generally don't feel a skill leaving. You just notice, later, that it's gone.
It's not just academic — developers are noticing themselves
A separate internal study at Anthropic found that engineers using Claude reported real concern about the decline of their own fundamental programming skills, even while their measured productivity was going up. And it's not confined to any one company's tools: a February 2026 study from METR found that most developers now refuse to work without AI coding tools even on a limited set of tasks — while the same body of research raises concerns that code quality may be declining as that reliance deepens.
So — worse, or just different?
The fair reading isn't "AI makes every developer worse at everything." It's narrower and more actionable: heavy, default reliance on AI for the effortful parts of coding — debugging, comprehension, architecture — correlates with measurably weaker performance on exactly those parts when the AI isn't there, and developers are bad at self-detecting it while it's happening.
That's a solvable problem, not a life sentence. It just requires something external to check the gap, because — per the research above — your own sense of how you're doing is precisely the thing that isn't reliable here.
Related reading: if this hit close to home specifically because you're early-career, Junior Developers and AI Dependency covers why the stakes are different at that stage. If you want the practical next step, How to Quit Vibe Coding is the how-to version of this article's conclusion. Related research angle: MIT Study: The Cognitive Dulling of Heavy AI Use.