The Skills We Keep

I was part of a conversation recently with a group of people who think hard about where technology is taking us, the kind of discussion that turns into less a meeting than an open exchange of ideas. It circled, as these conversations tend to now, around AI: how they are thinking about it, how we are thinking about it, where it all seems to be heading. It was a good conversation, and a few threads from it stuck with me afterward.

One question that came up has stayed with me since. As AI works its way deeper into technical roles, and into plenty of non-technical ones too, what happens to the underlying skills people used to build along the way? Are we trading capability for convenience? Is “skill degradation” something real and worth taking seriously, or is it the kind of worry that every impactful technology attracts and then outgrows?

The real issue, I think, is not whether AI causes some old skills to fade. Of course it will. Every useful technology does that. The harder question is whether we can tell the difference between skills that are safe to set down and skills that build the judgment we will still need when AI is doing more of the work.

I should say up front where I stand. If you placed everyone on a line running from AI pessimist to AI optimist, I sit closer to the optimist end. I think the upside is real and large. So this is not a piece about pumping the brakes, it is a piece about paying attention.

I am also offering observations here, not final answers. Some of what follows rests on fairly solid ground. Some of it rests on early studies and honest uncertainty. The goal is not to settle the question, but to think about it more clearly than the usual conversation tends to allow.

And the usual conversation, it seems to me, gets stuck on some version of the phrase “skill degradation,” which sounds like one big thing. The longer I sit with it, the more it seems like at least four different things using the same name.

Pulling “Skill Degradation” Apart

Part of why this conversation tends to go in circles, I think, is that we use a single phrase to describe several different situations. When someone says AI is eroding our skills, they could mean any of a few quite different things, and those things do not call for the same response. Some of them are not really problems at all. Lumping them together is how the conversation gets confused: one person is worried about something real, while another is waving off something that was never worth worrying about.

So before deciding how concerned to be, I think it can be helpful to separate them. When I do that, I come up with four kinds of skill loss tucked inside the one phrase. I have listed them roughly in order, from the kind I would lose no sleep over to the kind I will watch closely in myself.

• Letting Go: Skills we set down on purpose because something better came along. Most of the time this is simply progress, and we are right not to mourn them.
• Getting Rusty: Skills we still need but stop practicing, because the technology now handles the part we used to do by hand. The ability is real and needed, it just fades from disuse.
• Never Building It: Skills the next person never develops in the first place, because AI was there from the beginning and they never had to climb the hill themselves.
• Tuning Out: Skills we keep, and could still use, but stop bothering to apply because the technology is right often enough that checking starts to feel unnecessary.

The order of impact of these categories is part of the point. Much of the worry about AI and skills, I think, gets aimed at the first two, which are the most familiar and the most manageable. The latter two are quieter, harder to notice, and to me the ones actually worth our attention. The rest of this is really just a walk-through of each, with a real example to show what I mean.

Letting Go

We have been setting skills down for as long as we have been building technology to replace them, and often we are right to. I don’t do long division on paper anymore, I would be slow to balance a checkbook by hand, and I could not tell you the last phone number I needed to memorize. These were real skills I once needed, something better came along, and letting them go freed up attention for other things.

I think that a good deal of what people file under "AI is eroding our skills" is really this. If AI now drafts the boilerplate code, formats the comments, or remembers the syntax I used to keep in my head, I have lost something only in the narrowest sense. But there is a catch hiding in this category, and I think it is worth seeing clearly.

Consider the U.S. Naval Academy. For most of its history it taught midshipmen celestial navigation, the centuries-old practice of fixing your position at sea by measuring the angle of the stars with a tool called a "sextant." In the late 1990s the Academy phased out the detailed instruction, and by the mid-2000s the Navy had ended the training across the fleet. The reasoning was sound and seemed obvious. GPS could place you within a few feet of your true position. A skilled celestial navigator, on a good day, could get you within a mile or two. When one technology is that much better, keeping the old skill alive starts to look like nostalgia.

Then in 2015, the Academy quietly brought celestial navigation back. Not because anyone had fallen back in love with the sextant, but because GPS had turned out to carry a weakness we couldn’t fully engineer away. It can be jammed or spoofed, and requires electricity. A satellite signal can be attacked in ways a sextant simply cannot, and a ship that has lost GPS with no other way to find itself is in real trouble. The skill that looked obsolete turned out to be the backup for the exact moment the “better” technology falls short.

Personally, I think this story tells us something about making the decision itself: calling a skill obsolete is a judgment, and a judgment we can get wrong. And you tend to discover the mistake at the worst possible time, when the technology you trusted is unavailable and the skill you set down is what would have helped.

So to me, letting go is about understanding which skills are genuinely safe to retire, and which ones we are setting down only because the new technology has been reliable so far.

Getting Rusty

This is the category I think most people actually have in mind when they worry about AI and skills. The skill is real and still matters. But the technology takes over the part you used to do by hand, so you stop exercising the skill, and a skill that goes unexercised gets rusty. You never chose to give it up, it just faded while you weren't looking.

An example I know of comes from aviation. In 2009, an Air France flight over the Atlantic ran into a relatively ordinary problem: some sensors iced over, leading to unreliable airspeed readings, and the autopilot handed control back to the pilots. What followed was not an equipment failure, the aircraft was flyable the entire time. But the crew, who had spent years monitoring automation that didn’t ask as much of them, struggled with something a well-practiced pilot is trained to handle: recognizing and recovering from an aerodynamic stall. The plane was lost with everyone aboard.

The lesson was not as simple as "automation made the pilots less capable." Accidents rarely reduce that cleanly. But one of the lessons the aviation world drew from the crash was that highly automated systems can leave crews under-practiced for rare, high-stakes manual situations, especially when the handoff comes suddenly and under stress.

The same pattern is starting to surface with AI itself. A 2025 study in The Lancet Gastroenterology & Hepatology found that experienced endoscopists' unaided adenoma detection rates dropped after they had grown used to an AI assistant flagging suspect areas during colonoscopies, the first real-world documentation of a deskilling effect from clinical AI.

AI handles the routine ninety-something percent so capably that you rarely touch the skill yourself, which is fine right up until the rare case it cannot handle lands in your lap and the skill is not there anymore. Getting rusty is real, and it follows a consistent pattern across many more examples. The more reliable the technology, the less you exercise the skill underneath, and the less ready that skill is on the day the technology reaches its limit.

Never Building It

The first two kinds of skill loss happen to people who already have the skill. But this one is different; it lands on the people coming up behind them, the people who never build the skill at all, because AI was there from the first day and they never had to.

Think about how anyone actually becomes good at technical work. It rarely comes from the parts that go smoothly. It comes from the struggle: the bug that takes three days to find, the design that has to be torn up and rebuilt, the dead ends that teach you why the standard answer is the standard answer. Expertise is built from reps, and a lot of those reps are tedious and slow. They are also precisely the parts AI is now glad to do for you.

When a junior software engineer can hand the hard, formative problem to a model and get a working answer back, the problem gets solved and the learning that used to come with it does not. The output looks the same, but the person underneath is not.

And here is the part I find hardest to shake. An expert who leans on AI is trading away sharpness that, in principle, they can win back. A novice who never built the skill has nothing to return to. If the reps that turn junior engineers into senior engineers get handed to the machine, it is worth asking where the next generation of seniors is supposed to come from.

Of course this is hard to study in a workplace, where careers unfold over years. But at least one controlled version I know of has been run in a classroom. In 2025, researchers from the University of Pennsylvania published a field experiment in PNAS involving nearly a thousand high school students in Turkey who were split into three groups studying math. One group used a standard ChatGPT-style assistant as a tutor. A second used a version of the tutor built with guardrails, designed to offer hints rather than hand over answers. The third group had no AI at all. Afterward, all three took the same exam on their own, with no help.

During the study sessions, the students with the standard assistant looked strong, while AI was doing the work alongside them. On the exam, with the tutor gone, they scored about seventeen percent worse than the students who had never touched AI. They had been collecting answers without building the understanding underneath, and once the support was removed, there was less there than in the students who had worked through the problems the hard way.

One detail in that study matters enough that I will come back to it. The guardrailed version, the one that gave hints instead of answers, did not cause the same damage. Those students came out roughly even with the no-AI group. The problem was not exactly the presence of AI, but the way it was used.

I will be honest that the workplace version of this is still mostly ahead of us. The early classroom evidence is real, but careers are not semesters, and I have not seen a long study that follows a cohort of engineers who grew up handing their hardest early problems to a model. What I think we have is reasonable worry. Getting rusty at least leaves a "before," a level you once reached and could recover. But never building it leaves no before at all. The work still gets done, so nothing shows up on a dashboard. It surfaces years later, as a question that is hard to answer: why are there so few people who really understand how any of this works?

Tuning Out

The last kind is the strangest, because nothing is actually lost. You could still do the thing as well as you ever could. You just stop bothering, because the AI is right often enough that checking it has started to feel like wasted effort.

It creeps in: AI drafts something, you check it carefully, the output looks good. That happens again, and again, and each check feels a little more like a formality until one day you are not really checking anymore, you are clicking accept. The judgment you would have applied is still in you, but it has gone quiet.

This is the kind of skill degradation that I think is easiest to miss because it feels like competence. You are moving faster, shipping more, clearing the backlog. Every visible signal says you are doing better work. The one thing that has dimmed, your own effort put into scrutiny, is invisible because nothing has gone wrong yet.

You can see a related pattern beginning to show up in software, though I would treat the evidence carefully. GitClear, a firm that analyzes code changes over time, looked at 211 million lines of changed code from 2020 through 2024 and found some concerning shifts as AI coding assistants became more common. Refactoring appears to have declined, copy-pasted code increased, duplicated code outpaced reused code, and more code was reverted or rewritten shortly after being committed.

I would not use that data alone to prove that developers are tuning out. Some of what we are seeing may reflect how current coding assistants tend to work: they are often better at producing a fresh block of code than at understanding and improving the surrounding system. But the human part matters too. Duplicated logic, missed reuse, and shallow review are exactly the kinds of things a developer applying full judgment would normally catch. So while the data does not prove complacency, it does raise the right question: are we accepting more output because it is fast and plausible, while applying less of the scrutiny that used to come with writing it ourselves?

That is tuning out. The capability is still there, waiting to be used. What erodes is the habit of using it, and a habit is an easy thing to lose without noticing. It also feeds the previous category: stop exercising judgment for long enough and tuning out begins to turn into getting rusty, the willingness going first and the ability following close behind.

The Lines Blur

Look closely at any real case and it rarely sits neatly in one category. The endoscopists whose detection rates slipped may have been getting rusty, or they may have been tuning out. The same coding assistant can let a veteran’s skills fade while keeping a beginner from building those skills in the first place. The categories are clean on paper, but tangled in practice.

That matters because skill loss often does not look like decline while it is happening. In the moment, the work still gets done, often faster than before. The dashboard stays green. The backlog moves. The output improves. What weakens is harder to see: the knowledge, judgment, and practiced attention underneath the work.

These categories can also compound. Tuning out becomes getting rusty. A senior person who has gone rusty becomes a weaker teacher. A junior person without enough hard-earned reps has less chance to build the judgment they will later be asked to apply. Each form of loss makes the next one more likely.

That is the larger risk I keep coming back to. Today’s AI systems are fluent, fast, and occasionally confidently wrong. What stands between that and real consequence is human judgment: people who know enough to notice when the answer is off. In a healthy organization, that judgment is distributed. Experienced people catch mistakes, and less experienced people become experienced by watching, struggling, and doing the work themselves.

But if the experienced people get rusty or tune out, they catch less. If the newer people never build the skill, they are not growing into the catching role at all. The risk is not that any one of us becomes slightly worse at a task. It is that a team, a company, or even a profession slowly loses some of its ability to tell when the machine is wrong, at the same time it is handing the machine more to do. To me, that is not a reason for panic, but it is a reason to be deliberate.

Using AI on Purpose

If the picture so far sounds grim, here is what pulls me back toward optimism, and it comes from the same research that worried me. Remember the math study with three groups of students. The AI tutor that handed over answers left students worse off once it was taken away. The version of the tutor that was built to give hints instead, to make them work a little before it helped, largely mitigated the damage. Same AI model, similar profile of students, same subject. The only thing that changed was how the technology was meant to be used, and that one change largely determined whether it acted like a crutch or a support.

I find that hopeful, because it moves the question from one we cannot control to one we can. The real question was never whether to use AI – it’s how. Skill loss is not an unavoidable tax bundled with the productivity gains.

The reps that build skill tend to be the slow, effortful, slightly annoying ones, which are exactly the reps AI is most eager to take off your hands. So, the discipline required is to notice which of those reps are actually "load-bearing", and to keep doing some of them yourself even when AI could do them for you. Because the point of those reps was never only the output, it was what doing them built in you.

That matters most for the skill steadily becoming the central one: judgment. As AI takes over more production, the human job shifts toward evaluation: knowing whether what came back is any good, and catching the moment it is confidently wrong.

But this is the catch that keeps the whole discussion honest: you mostly cannot evaluate well what you never learned to produce. The senior person who can spot the model’s bad answer can do it because they once wrote enough good and bad answers themselves to feel the difference. Verification does not bypass expertise. It runs on expertise that was built the hard way. So we cannot simply crown judgment the new skill and skip the work that used to create it.

For those of us building careers around this technology, I keep returning to a progression in my own use of AI: from tool user, to collaborator, to orchestrator of agents. The temptation is to vault straight to the top, to hand whole problems to a system and supervise from above. The trap is that good orchestration rests on judgment, and the earlier stages are where you build it. You can skip the climb in the short run. What you cannot skip is what the climb was teaching you, and you tend to discover the gap only when something breaks and you find you cannot tell what went wrong.

None of this is an argument to use less AI. I use a great deal of it, and I have no intention of slowing down. It is an argument to use it on purpose: to ask, of the work we are handing over, which parts we were getting better by doing, and to keep a hand in those. Adopt eagerly. Adopt deliberately. Those two are not in tension, and the entire difference is in holding both at once.

What I’m Watching For

I started with a question someone put to me, the kind that lodges itself in your brain and will not leave. I have not landed on a super clean or tidy answer, and I am a little suspicious of anyone who claims one. What I have instead is a clearer sense of what I will pay attention to in my own behavior and in our teams.

I am not worried about the skills we are glad to set down. I am watching the skills we still need but stop practicing, the skills the next generation never gets to build, and the judgment we keep but slowly stop applying.

What I keep coming back to is that losing important skills is not an inevitable outcome. It is the result of how we choose to use AI, and that leaves most of the outcome in our hands. The same study that shows AI can hollow out learning also shows that a small change in how it is used can prevent harm.

So I land more or less where I started: closer to the optimist end, and unmoved on that. The upside of AI is real and large, and I want us to reach for all of it. I would just rather we reach for it with our eyes open, keeping a hand in the work that quietly made us good in the first place. Because if AI becomes as capable as we keep promising each other it will be, we will still need people around who deeply understand what it is doing.