That moment stuck with me longer than it should have, because it wasn't really about the regex. It was about how easily I had stopped reaching for my own brain first. And the more I paid attention after that, the more I noticed the same pattern everywhere, in my own work and in the work of nearly every developer, writer, and student around me. We are not becoming less capable overnight. We are becoming quietly, gradually, less practiced.

This Isn't Just a Feeling

It would be easy to write this off as nostalgia, the same complaint every generation makes about the next one relying too much on whatever new tool showed up. Calculators, spell check, GPS, Google itself, all accused of making us "lazier" at some point. But the research on AI specifically is starting to paint a more specific picture than that old complaint, and it's worth taking seriously.

A 2025 study out of SBS Swiss Business School, led by researcher Michael Gerlich, surveyed hundreds of participants across different ages and education levels and found a clear negative correlation between how often someone used AI tools and how well they performed on standardized critical thinking assessments. The mechanism behind it has a name now: cognitive offloading, the tendency to hand a mental task over to a tool instead of thinking for yourself. The study found this offloading habit was strongly linked to the drop in critical thinking scores. Notably, the effect wasn't linear. Light to moderate AI use didn't show much impact. It was heavy, habitual reliance where the decline showed up clearly, and younger participants showed the strongest dependence of all.

Around the same time, a joint study from Microsoft and Carnegie Mellon University looked specifically at knowledge workers, people whose entire job is built around thinking and producing ideas. The pattern they found was almost uncomfortably intuitive once you hear it: the more confidence someone had in an AI tool's ability to handle a task, the less critical thinking effort they reported putting in themselves. Trust in the tool and effort from the person moved in opposite directions.

Even MIT Media Lab has weighed in, describing what it called early signs of "cognitive atrophy" tied to excessive reliance on AI-assisted work. The study is small and hasn't been peer-reviewed yet, but the language itself is telling. Atrophy is what happens to a muscle you stop using. Apparently the same word is starting to show up in conversations about thinking.

What This Looks Like in Software Development Specifically

Developers have a front-row seat to this shift, maybe more than almost any other profession right now. AI code assistants are genuinely good. They can scaffold a component, write a test suite, explain an error message, and refactor a messy function faster than most of us can type the prompt describing what we want. That capability is real, and I'm not here to pretend it isn't useful.

But there's a version of using these tools that looks less like collaboration and more like outsourcing your own understanding. It shows up as accepting a suggested block of code without reading it closely enough to know what it actually does. It shows up as debugging by pasting an error into a chat window instead of reading the stack trace yourself, the way developers were forced to for decades before autocomplete got this good. It shows up, most quietly of all, in junior engineers who can produce working features remarkably fast without ever building the mental model of why the code works, which is the exact skill that used to separate a junior from a senior engineer over the course of a few years.

None of this is really the AI's fault. A calculator doesn't make you bad at arithmetic on its own; only skipping the arithmetic entirely does that. The tool isn't the problem. The absence of friction is the problem, because friction, the mildly uncomfortable experience of actually struggling with something, is where a surprising amount of real learning happens.

The Skills Most Likely to Quietly Erode

A few categories seem particularly exposed, based on both the research and what's visible just by watching how people work day to day.

Debugging and root cause reasoning. Asking an AI to fix a bug is fast. Understanding why the bug happened in the first place, and building the instinct to spot similar bugs before they occur, requires actually sitting with the problem, which is exactly the step that gets skipped when the fix arrives in five seconds.

Writing and argument construction. Drafting with AI assistance can produce clean, competent prose almost instantly, but the actual skill of writing has always been inseparable from the skill of thinking clearly. When the sentence arrives before the thought has been fully worked out, the thinking sometimes never quite gets done.

Basic quantitative reasoning. This one predates AI, but generative tools accelerate it. When every estimate, every ballpark calculation, every "does this number make sense" check gets outsourced, the underlying number sense doesn't get exercised enough to stay sharp.

Sitting with not knowing something. Perhaps the least discussed skill on this list. Tolerating the uncomfortable state of not immediately having an answer is itself a cognitive skill, and it is arguably the one being eroded fastest, because every uncertainty now has an instant, confident-sounding response available within seconds.

Why This Actually Matters, Beyond the Abstract Worry

It's tempting to treat this as a philosophical concern with no real stakes. It has real stakes. A developer who can produce code they don't fully understand is fine right up until that code breaks in production at 2 am, in an environment the AI has no context for, under time pressure the AI can't feel. A writer who can produce a clean paragraph but can't defend the argument inside it under real scrutiny finds that out the first time someone pushes back in a live conversation instead of a comment thread.

There's also a slower, quieter cost that doesn't show up in any single incident. Skills that go unused don't just sit there waiting. They degrade, the same way a language you stop speaking gets harder to recall, or a muscle you stop training gets weaker even though you technically still "have" it. The Harvard Gazette's coverage of the MIT research put it plainly: letting AI do your thinking, whether through autocomplete or through the increasingly common habit of "let AI write the first draft, and I'll just edit it," quietly undercuts both critical thinking and creativity over time, not through one dramatic failure but through thousands of small skipped repetitions.

What Seems to Actually Help

The research isn't uniformly bleak, and this is worth sitting with too. The Microsoft and Carnegie Mellon findings, along with several follow-up studies, suggest the relationship between AI use and thinking skill isn't simply "more AI equals worse thinking." Structured, deliberate use of AI, where someone remains actively engaged rather than passively accepting outputs, showed a very different pattern than passive, high-volume use. Higher education levels and higher self-confidence in one's own judgment both seemed to act as a buffer against the decline, which suggests this isn't inevitable so much as it is a habit that can be built deliberately in either direction.

A few practices seem to make a real difference, based on both the research and honestly, my own trial and error since that regex moment.

Attempt it first, even badly. Even a rough, wrong first attempt at a problem builds more of the underlying skill than reading a correct answer someone else generated. The struggle is doing something, even when it doesn't feel productive in the moment.

Use AI to check your reasoning, not replace it. There's a meaningful difference between asking "write this function for me" and "here's my approach, what am I missing?" The second version keeps you in the loop as the one actually thinking.

Read what you accept, every time. If AI generates code, a paragraph, or an answer, actually understanding it before using it is the difference between borrowing a tool and outsourcing your judgment.

Protect specific moments of friction on purpose. Not every task needs to be optimized for speed. Deliberately doing certain things the slower, manual way, on a schedule you control rather than by accident, keeps the underlying skill alive even while AI handles the bulk of the routine work.

Where This Leaves Us

I don't think the answer here is to abandon these tools, and I don't think that's realistic advice for anyone actually shipping software or writing for a living in 2026 anyway. The efficiency gains are real, and pretending otherwise isn't useful. But efficiency and capability are not the same thing, and it's worth being honest that we are, collectively, running a real-time experiment on which cognitive muscles atrophy when we stop using them by default.

The regex I didn't write myself wasn't a big deal on its own. It's the thousandth version of that same small choice, repeated without noticing, that adds up to something worth paying attention to. The tool isn't going anywhere, and it shouldn't. The question worth asking, for developers and for basically everyone else working alongside AI now, is simply whether we're staying in the loop as the ones thinking, or quietly handing that part over one convenient shortcut at a time.