At three in the morning, the scan arrives before the coffee does. Priya, a second-year radiology resident, opens the chest CT and finds that the software has already done the looking. Three small nodules glow inside yellow boxes, a confidence score beside each. Her job, officially, is to confirm. In practice, her eyes go straight to the boxes, and only to the boxes.

(Priya is a composite, drawn from how AI tools now run through medical training. She is not a real person I interviewed.)

Nothing goes wrong that night. That is the problem. The software is usually right, so Priya rarely has to hunt for the faint shadow it might miss. But hunting is how radiologists are made. Each search through an unmarked scan builds a private library of pattern and doubt that no lecture can supply. Remove the hunt and you remove the training, while keeping the title.

This is not a hypothetical worry. A 2025 study in The Lancet Gastroenterology & Hepatology followed experienced endoscopists in Poland after their clinics adopted AI that flags polyps during colonoscopies. When those doctors later worked without the tool, their detection rate fell from about 28 percent to about 22 percent. These were veterans, deskilled within months. Imagine the effect on someone who never had the skill to begin with.

The educator J. Jacob Tawney, writing recently in First Things, states the general rule plainly: technology replaces what it automates. He also separates the algorithm, the neural network doing the predicting, from the interface, the way a person meets it. A yellow box around a nodule is an interface choice. So is a chatbot that answers in warm, complete sentences. Neither is required by the underlying math.

So here is my argument, and it is meant to be argued with: learners should have to earn the machine. Until a person can do a task competently without AI, they should not be allowed to use AI to do it.

Aviation learned this the expensive way. In 2009, Air France Flight 447 stalled over the Atlantic after its autopilot disconnected and a crew long accustomed to automation mishandled the plane by hand. All 228 people aboard died. Four years later, the Federal Aviation Administration urged airlines to give pilots more time flying manually. Pilots still use autopilot. They are simply expected to fly without it first, and to keep doing so.

Medicine, law and schools could adopt the same rule. A resident reads unmarked scans until she passes a blind competency check, and only then gets the boxes. A law student writes briefs from scratch before an AI drafts one. A ninth grader writes her own essay, every time, because for a ninth grader the essay is not the product. The struggle is.

The rule matters most for the most humanlike interface of all. A teenager's chatbot companion is always awake, never bored and tuned to please; engineers call this tendency sycophancy. For an adult with settled friendships, it is a convenience. For a fourteen-year-old still learning what friendship feels like, it is a curriculum. Children should earn that machine too, by first learning how real people talk back.

The objections are serious. AI readers catch cancers that exhausted humans miss, so holding residents back may cost patients today. We have reasonably let some skills go, like long division. And any such rule is hard to enforce when the tool sits in every pocket.

But the costs are not symmetrical. Withholding help during training costs some accuracy for a few years. Raising a generation that cannot check its machines costs that ability for good. Probabilistic systems eventually err, and when they do, someone has to notice. Somewhere, a future Priya needs to be able to look at a scan with no boxes on it and see what is there.

The question is not whether AI makes us better at our jobs today. It is whether anyone will still know how to do them tomorrow.