When Nobody Has to Try

A tutor that answers for you makes you excellent for as long as it sits beside you. In an experiment with nearly 1,000 high school students in Turkey, researchers at Wharton and the University of Pennsylvania split the class in three: one group used a ChatGPT-style assistant, one used a version that gave hints under teacher supervision, and one had only books and notes. In practice sessions, the first group scored 48% higher than the books-only group. On the exam, without the assistant, it scored 17% lower. The hint-based group did better in practice (+127%) and matched the books-only group on the exam.

The same pattern shows up where a mistake costs more. At four colonoscopy centres in Poland, 19 experienced endoscopists, each with more than 2,000 procedures behind them, started working with an AI system that flags polyps. When they later worked without it, their adenoma detection rate fell from 28.4% to 22.4%. The study, published in The Lancet Gastroenterology & Hepatology, is observational, so other causes cannot be ruled out. It is still a number nobody expected to read about senior doctors.

Upstream delegation is handing a problem to a tool before you have tried it yourself. Lisanne Bainbridge described the mechanism in 1983, for industrial plants: automation takes the easy work away from the operator and leaves the rare, hard cases, exactly when the operator no longer practises on the ordinary ones. The more advanced the system, she wrote, the more crucial the human contribution. Someone who tries first and then asks for help trains judgment. Someone who delegates upstream gets the answer without ever measuring how far it sits from the problem. Most wrong answers never meet a test that exposes them, so the only check left is the person reading.

History shows what effort spread across a whole community can do. Venice stood on a lagoon that mud and rivers kept trying to close. In 1501 the Republic set up the Savi alle acque, a board of magistrates with the final say on rivers and lagoon. It refused requests to irrigate or canalise the Sile, the Brenta and the Piave, to keep the water the city depended on. The land around paid with floods. In the early sixteenth century the Arsenal employed around 16,000 people, split by stage of work, with up to 100 galleys under construction at once and, reportedly, close to one ship a day. Survival set the agenda, but the example matters for another reason: every trade met a concrete problem every day, and every solution stayed with the people who found it.

Sure, every tool has taken something from someone, and the calculator did not make anyone incapable of thinking. Yet a calculator replaces an operation you can already define. An AI assistant answers even when you cannot yet say what the problem is, so what gets delegated is the formulation, the stage at which most of the reasoning happens. That is a hypothesis, not a finding. The studies above measure short, single tasks, and none of them proves a general, lasting loss. The Wharton result also shows the other side: with hints and a teacher alongside, the damage disappeared.

One fix is cheap: write down your own guess before you read the machine’s answer, so the distance between the two stays measurable. That is the logic behind calibration of the human and the model as a pair. Without it, nothing tells you whether your trust is earned.

What remains is the question of who benefits if upstream delegation becomes a habit. People who can no longer check an answer cannot contest it, and the problems worth working on get chosen by the few who build and sell the tools. They also collect the credit for progress, while the rest of us consume what they build. On 9 October, the CBS programme 60 Minutes published a clip from an interview with Daron Acemoglu, the Nobel-winning economist at MIT. He was talking about a hypothetical AI that surpasses the best humans in every field, but his line holds without getting there: “We’ve never had something like that where most of us feel we’re dispensable.” The mechanism I describe starts with the tools we already have.

When a tool works in your place, the useful test is whether you could do the task tomorrow without it. If the answer is no, upstream delegation has already happened, and with it the ability to notice when the tool is wrong, the same ability that lets a worker stop the line.