Skills That Fade When the Machine Does Them

In a 2023 experiment, 27 radiologists read 50 mammograms, each with a suggestion from what they were told was an AI system. On most cases the suggestion was right, and readers at every level of experience rated about 80 percent of those cases correctly.

On 12 of the mammograms, the suggestion was deliberately wrong. Accuracy collapsed. The least experienced readers got about one in five right. The most experienced did better, 46 percent, but even they were pulled toward the wrong answer more often than not.

Two things are going on there, and both matter for anyone putting AI into skilled work. People defer to a confident suggestion, even when they're experts. And the thing that protects them, independent skill, is exactly what a tool that's usually right slowly stops exercising.

When a tool does the effortful part of a task, people get better with it and can get worse without it. The skill that fades is often the one needed to notice when the tool is wrong. And the decline makes no noise: the person's numbers improve while their judgment weakens. So the skills that matter have to be chosen, practiced on a schedule and measured on purpose, because the job no longer does it for you.

Doing well and learning are different things

The psychologists Elizabeth and Robert Bjork describe two quantities that feel like one. Retrieval strength is how easily you can produce something right now. Storage strength is how deeply you've learned it. Performance during a task reflects the first. Lasting skill is the second.

The two can move in opposite directions. Conditions that make practice feel harder, like spacing it out, mixing problem types, or having to produce an answer rather than recognize one, slow you down in the moment and build more durable learning. The Bjorks call them desirable difficulties.

A capable AI tool removes difficulty, the desirable kind along with the rest. It drafts, so you don't generate. It retrieves, so you don't recall. Work done with the tool feels fluent, goes faster, and leaves less behind.

The cleanest evidence comes from a classroom. In a randomized trial with nearly 1,000 Turkish high school students, math classes practiced either with textbooks, with a GPT-4 chat interface like ChatGPT, or with a GPT-4 tutor instructed to give hints rather than answers. With the plain chatbot, practice scores rose 48 percent. On the exam afterwards, without any tool, those students scored 17 percent lower than students who had never had access. The chat logs showed why: they mostly asked for the answer and copied it. The hint-giving tutor raised practice scores even more and left exam scores about where they were. Students in neither group felt they'd learned less.

Same model. Told to leave the generating to the student, it did no harm.

The irony of supervising a good machine

In 1983 Lisanne Bainbridge wrote a short paper called "Ironies of Automation" about operators of automated industrial plants. Its central irony fits AI almost word for word.

Automation is brought in because it does the job better than the person. The person is then asked to monitor it and take over when it fails. But taking over requires the very skills the automation stopped exercising, in a situation harder than usual, because something has already gone wrong. And monitoring requires knowing what correct behavior looks like, which is the same knowledge. Her conclusion runs against intuition: the most reliable automated systems may need the greatest investment in training the people who supervise them.

The skill a tool stops exercising is often the skill needed to notice the tool is wrong.

Pilots show what fades first. In a 2014 study, 16 airline pilots who rely on automation for most of every flight flew scenarios by hand in a 747 simulator. Their stick-and-rudder skills were mostly intact. What had degraded was the mental work: tracking where the aircraft was without a map display, knowing which navigation step came next, recognizing instrument failures. Performance on those tasks went with how often pilots said their minds wandered while the autopilot flew. The hands remembered. The judgment that interprets a situation faded when nothing demanded it, and supervision runs on that judgment.

Add the pull of a confident suggestion, called automation bias, and you have two problems stacked: people defer to the tool, and over time their own ability to see its errors erodes.

Three ways a skill goes missing

Three different things get called deskilling, and they need different responses.

Deskilling in the strict sense is losing a skill you had, like the pilot who hand-flew for years and now mostly monitors. Something remains, so practice can rebuild it.

Never-skilling is not building the skill in the first place. A junior analyst whose first drafts come from a model skips the repetitions that would have taught her. Bainbridge saw this coming too, warning that operators were "riding on their skills, which later generations of operators cannot be expected to have." There's no earlier level to compare against, so it's invisible, and it shows up years later as people who look experienced on paper. The tasks automated first tend to be the ones juniors used to learn on (A Job Is a Bundle of Tasks).

Skill shift is the benign version: the skill moves somewhere else that matters, as calculators moved school arithmetic toward estimation and setting up the problem. Knowledge workers surveyed by Microsoft Research and Carnegie Mellon described this with generative AI, their critical thinking moving from producing material toward verifying and integrating it. A shift is a good outcome only if the new skill is enough for the job, including the part that happens when the tool fails.

The decline nobody sees

Here's why this is a management problem and not only a training one. A simple model shows how it could play out for one person. The parameters are plausible assumptions, not data.

A reviewer checks AI-drafted documents. Before the tool, she wrote similar documents herself every day. Now the AI drafts, she reviews, and she writes one herself about once a month. About two subtle errors a week reach her, the kind only someone who knows how to write the document well would catch. Her skill decays slowly without practice and is rebuilt a little by each document she drafts herself.

Scroll sideways to see the whole figure

Drafting daily, she catches about 75 percent of those subtle errors. Drafting monthly, her catch rate falls to 54 percent after six months and about 42 percent after two years. In her second year she misses about 59 of the 104 subtle errors that pass through her hands, more than twice as many as when she drafted daily.

Add one deliberate practice session a week, drafting a document independently and then comparing it with the AI's version and an expert's notes, and her catch rate settles around 70 percent. One weekly session recovers most of what five daily ones provided, because her skill settles where practice balances decay, and a little well-designed practice tips that balance a long way.

The junior on the right of the figure starts with a 20 percent catch rate. Drafting monthly, she reaches about 36 percent after two years. With a weekly practice session, about 69. That's never-skilling in numbers: the experienced reviewer spends down savings, the junior never makes the deposit.

Now the uncomfortable part. None of this shows up anywhere. Missed errors are, by definition, unseen. Her approval rate goes up, her throughput goes up, and her manager sees a reliable process. The first evidence is a failure downstream, months later, blamed on the AI.

Her numbers improve while her judgment weakens. The first signal is a failure, months later, blamed on the machine.

Which skills to protect

Most people can no longer do long division quickly, and it doesn't matter. The calculator is almost never wrong, and a rough estimate catches the errors it makes. Three questions sort the cases where a faded skill does matter.

Will people need to act without the tool? A pilot has to fly when the autopilot drops out. An endoscopist works in rooms without the AI detector. If the answer is never, the faded skill costs little.

Does checking the tool's output require the faded skill? This is Bainbridge's irony as a question. Checking a calculator takes estimation, a much cheaper skill than computation. Checking a contract a model drafted, or code an agent wrote, takes much of the skill it took to produce them, because the errors that matter are the plausible ones. A model's wrong answers sound exactly like its right ones (Confidently Wrong).

How rare and how costly are the failures? Frequent failures keep people practiced and alert. Rare ones let skill and vigilance decay in between. Rare, expensive failures are the worst combination, and they're precisely what a well-functioning automated system produces.

Scroll sideways to see the whole figure

Most office AI sits in the lower right of that grid. People rarely work without the tool, but judging its output needs the skill it replaced. The goal there is enough skill to judge, short of full skill to produce.

The calculator objection

The strongest objection: people said all this about calculators, spell-checkers and GPS. Teachers feared calculators would destroy arithmetic, and a 1986 review of 79 studies found that students using them alongside regular instruction kept or improved their pencil-and-paper skills at nearly every grade studied. Most of the time people adapt, and the skill that matters moves.

That's right about most skills a tool absorbs, and panic about each new technology has a poor track record. But look at why the calculator story worked. Calculators arrived alongside teaching that still covered arithmetic, and their failures, like a mistyped number, are caught by a rough estimate. Checking the tool didn't require the skill it replaced.

AI tools differ on exactly that condition. Their errors are plausible rather than obvious, because fluency is what training rewards. Their reliability is jagged: high on some tasks, low on neighbors that look similar. And the tasks they take over, drafting, analysis, code, diagnosis, are ones where checking needs most of the judgment it took to produce. A rough estimate catches a calculator's error. Nothing equivalent catches a subtly wrong legal argument.

So my position is selective. Where AI output can be checked cheaply by other means, tests that run, numbers that reconcile, a second independent source, the calculator story may well repeat, and the skill can shift to designing those checks. Where checking still depends on the skill being replaced, it won't repeat by itself. How fast skills fade over years of daily AI use, and how fast they come back, nobody has measured yet. That uncertainty argues for cheap insurance, not for waiting.

What to do

Write down the short list of skills you must keep. The ones needed to check the tool's output and to recover when it fails. Let the rest go deliberately, not by drift.

Schedule practice without the tool. Regulators asked airlines to make room for hand flying; Bainbridge suggested hands-on time each shift. In office work that means periodic independent work on real cases (draft the memo, write the function, run the analysis), compared afterwards with the AI's version, rotated so more than one person stays able to do it. The model above suggests the dose can be modest if it's regular and comes with feedback.

Configure tools to leave the hard part to the person when learning is the point. For training and new hires, a tool that gives hints instead of answers. In a 2026 experiment with developers learning a new code library, those who asked the assistant conceptual questions learned about as much as those who worked by hand, while those who delegated the coding learned least. It costs speed. Pay it where building the skill matters.

Ask for a judgment before showing the AI's answer. The mammography experiment shows how hard a suggestion pulls. Forming a view first keeps the skill in use and makes disagreement visible. It's slower, so use it on a sample or on the riskiest cases (Human Review Where It Matters).

Measure the reviewer, not only the tool. Seed a few known errors into the review queue and count how many get caught. That turns an invisible decline into a number someone can watch, the same logic as testing a model on your own work, applied to the person.

Give juniors the work that teaches, even when AI would be faster. Never-skilling is mostly an organizational choice. A firm can save money today by letting AI write the first drafts juniors used to write, and pay for it in five years when nobody can review the AI's drafts. Budget some junior work as training, and call it that.

Check yourself

Why did the radiologists' accuracy collapse when the suggestion was wrong?

People defer to a confident suggestion even when they are experts, and independent skill is what a tool that is usually right slowly stops exercising.

Why does the decline of a reviewer's skill go unnoticed?

Missed errors are unseen by definition, while her approval rate and throughput go up. The first evidence is a failure downstream, months later, blamed on the AI.

Why did the calculator story work out when the AI story may not?

Calculator errors are caught by a rough estimate, so checking didn't need the skill it replaced. AI errors are plausible, and checking drafting, analysis or code needs most of the judgment it took to produce them.