The colonoscopy study is right. The lesson is not the obvious one.

The number

28.4 percent, falling to 22.4 percent.

That is the unaided polyp-detection rate of 19 expert endoscopists, each past 2,000 colonoscopies, measured on standard cases in the three months before an AI assistant arrived and the three months after. Not trainees. Not a lab toy. Experts, losing a fifth of their eye for cancer, measured cold with the machine switched off. Budzyń and colleagues, The Lancet Gastroenterology and Hepatology, August 2025.

It is clean and it is damning. I want to sit with what it actually found, because the obvious reading skips the part that is useful.

Here is my stake. This year I watched a skill I spent decades building go slack in weeks. The same year, on the same kind of tool, I learned more than in any decade before it. Both are true. The gap between them is the whole essay.

The machine eats the reps

Call it by its name. Automation complacency. Hand the watching to the box and your own watching withers.

Expertise is not stored knowledge. It is a perishable loop kept sharp by reps, and every routine case is a rep whether you notice or not. The AI ate the reps. That is crueler than laziness, because nobody felt lazy. Their assisted scores stayed high the entire time. The skill drained out underneath a clean dashboard.

This is the trap that "just use it actively" never escapes. Watching a machine confirm a read is not the same muscle as making the read from a blank screen. On a routine case, while the tool does the looking, supervision is the only posture on offer. The one active thing that would keep you sharp is the exact thing the tool exists to do for you. So the only real defense is not a cleverer grip on the tool. It is putting the tool down.

The complaint is older than the alphabet

Twenty-four centuries ago Socrates warned that writing would rot memory and hand students "the show of wisdom without the reality." He was right. We lost the bards who could carry an epic in the body. He was also completely wrong, and the proof is that his complaint reaches us only because someone wrote it down.

Plato's word for writing was pharmakon: poison and cure in one object. The calculator is the same story. It killed mental arithmetic in everyone who reached for it without thinking, and sharpened number sense only in the few who were made to estimate first and check after. The good outcome was real. It was also the effortful exception, and it never survived the default.

Every deskilling panic has that shape. True about what is lost. Quiet about whether the saving exception scales. It almost never does.

The brain leaves a fingerprint

GPS that dictates every turn means the map is never drawn, and heavy GPS users show worse spatial memory over time. London cabbies who hold 25,000 streets in their heads grow a measurably larger hippocampus. The trainees who wash out do not. Same city, same task, opposite brains, and the only variable is whether the human did the thinking or received it.

The endoscopists were the GPS users. The machine did the looking, so the looking softened. Decades of learning science say it under other names: what you generate yourself sticks, what you read off a screen does not.

The same machine, the other way

A GPT-4 tutor built to withhold the answer and feed one step at a time roughly doubled the learning gains of a Harvard active-learning class, in less time. Built to make students generate, it made them learn faster.

But notice what that measures. The tutor builds a loop in a beginner. The colonoscopy study watched a loop decay in an expert. Different jobs. Nothing about the tutor proves that routine AI keeps a built skill alive. The rule I keep coming back to is narrow, and it is the spine of the whole thing: the variable is who does the generating. Point the machine at a beginner and force the reps, it teaches. Hand an expert the answers, it rots.

Why the simple version is mostly right anyway

Here is the uncomfortable part. The corrosive posture is the default posture. The entire pitch of the tool is less effort, so asking people to add back the effort that keeps them sharp is asking them to fight the gradient, on deadline, inside an institution that banks the speed and ships the decay to the future. No clinic schedules cold reps. No shop bills a client for the hours a senior writes code by hand to stay sharp.

And the decay is invisible from the inside, because your assisted output stays good. The evidence is stacking up. One study found the ChatGPT group wrote better essays, gained no more knowledge, and self-corrected less, which the authors called metacognitive laziness. An MIT EEG study found the AI writers had the weakest brain connectivity and often could not quote what they had "written" minutes earlier. They called it cognitive debt.

A better essay and a worse writer is not a paradox. It is the whole mechanism in one breath.

So "AI makes you worse" is crude about the mechanism and roughly right about the outcome. At scale, under the defaults, it predicts exactly what most people will do. The only word it gets wrong is inevitable.

Decay you can fix, and a skill you cannot

Two cases, and never confuse them. A skill that decayed can come back, because the deep wiring is still there to drill against. The catch is that recovery means practicing without the tool, which almost nobody does once the tool exists. Reversibility you never use is worth nothing.

The skill never built is worse, and it is permanent. The junior who never found a polyp cold, never wrote the function before autocompleting it, has nothing to recover, only something to build from zero, and the routine cases that would have built it are already eaten. The ladder is gone for the people who needed the rungs.

Which is where I turn the knife on myself. The people who swear the machine made them sharper are mostly the ones who already owned a skill to learn against. I am one of them. "I learned more than in any decade" is exactly the from-the-inside report this study says you cannot trust, mine included. The only reason I half-believe it is that I did the one thing the study rewards and nobody is paid to do.

What I actually do

I use the model for research and to bridge the gaps in what I know. What I do not do is hand it the recurring work. I do not prompt it to grind through the same task again and again. I prompt it for the script, the tooling, or the missing piece of knowledge, and then I automate the task myself, once, with something I built and can read.

That is the move that keeps the reps mine. The model does not sit in the loop doing the work, it helps me build the thing that does. The understanding stays in my hands, and the repetition goes to code instead of to a habit of asking again tomorrow. Point the tool at the recurring task and it eats your reps. Point it at the leverage and it hands you reps you would never have had the time to build.

I will not pretend this scales. Building the tool is more effort than reaching for the model one more time, and on a deadline most people reach for the model. So the systemic version is not advice, it is an audit. Measure unaided skill. Notice when the easy path is quietly spending it. Price the decay before someone bills it to the future.

The study is right. The simple reading just misses one move. It says the tool does this, when it looks more like the leaning that does it, and the building instead of leaning that undoes it. That sounds like a technicality. To me it is the difference between a curse and a discipline.

The machine will gladly do your thinking for you. That was never the danger. The danger is that you let it, and call the silence mastery.