What the Cockpit Showed
Why editing an AI-written draft leaves a finished document looking the same whether you could have started it yourself or not.

A magenta line shows the course on the flight display. The pilot is head-down with the technology, losing track of the airplane. The question becomes: "what's it doing now?"
That was the pattern American Airlines captain Warren Vanderburgh addressed at the airline's training academy in Dallas. In April 1997, he called his class "Children of the Magenta Line."
His team had examined accidents, incidents and violations. They attributed 68% of those they analyzed to automation mismanagement. Pilots could become so busy managing the systems that they lost track of the flight.
The computers were meant to reduce their workload. But choosing buttons and modes could absorb the attention they were supposed to free. When things grew complicated, a pilot could keep working on the automation instead of taking back control.
Vanderburgh taught pilots to choose between levels of automation. Managing a flight through a computer was one level; flying by hand was another.
At least the difference was visible. A pilot following the magenta line could tell that the autopilot was flying.
Still doing the work
When I let a tool generate first and only edit afterward, I notice less practice in the judgment I rely on. But nothing about the finished document makes that obvious.
I still open the document, still change the words, still send it under my name. There is work in that. I have to read closely enough to decide what should stay. The person who reads it never sees the draft the tool wrote. They see what I decided was good enough to send.
What I cannot tell from those changes is whether I could have built the draft myself. The tool has already chosen a starting point. I can improve what follows without having to find that starting point.
I think that is how skill loss could hide inside a working day. My attention stays occupied. The document gets finished, and the visible evidence says I have been writing.
The risk grows when I accept the starting point the tool has already drafted before forming my own. Even rejecting a sentence leaves me responding to something already there. I have not had to face the empty page.
The cockpit showed which system was flying. My draft does not show which part of the thinking I stopped practicing.
When the square is gone
In 2025, Krzysztof BudzyĆ and colleagues reported on 19 experienced endoscopists at four centers in Poland. Each had performed more than 2,000 colonoscopies. The AI drew a square around suspected adenomas, precancerous growths that colonoscopy can remove before they turn into colon cancer, helping bring them to the doctor's attention.
The doctors continued doing colonoscopies after the centers introduced routine AI use. Some procedures used the tool; others did not. The researchers compared their unassisted detection rates before and after that introduction.
Before routine AI use, the doctors found adenomas in 28.4% of unassisted procedures. After three months of exposure, that rate was 22.4%.
After three months of exposure, that rate was 22.4%.

Click a numbered dot to see what it stands for. Start with 1.
This was an observational study. It found an association, not proof that AI caused the drop. Other factors could have affected the results, and the researchers called for further trials.
Still, these were experienced doctors continuing to perform the procedure.
In 2025, Hamsa Bastani and colleagues reported a randomized controlled trial involving nearly 1,000 high-school math students in Turkey. During practice, students using GPT Base, a chat tool like ChatGPT that gave full answers, performed 48% better than students without AI. Some asked for solutions and copied them.
Then the tool was removed for an exam. The GPT Base group performed 17% worse than the control group.
That experiment covered short-term learning in one subject at one school, not professionals losing an established skill. But it did separate two things that the practice scores had made easy to confuse: completing the problems and learning to solve them.
The GPT Base group performed 17% worse than the control group.

Click a numbered dot to see what it stands for. Start with 1.
So why might I miss the difference in my own work?
The students who used AI were overly optimistic about how much they had learned. Their stronger practice results had not given them a sound measure of what they could do alone.
I understand the appeal of that reassurance. A finished draft gives me something I can inspect and send. The judgment I did not practice leaves nothing on the page for me to inspect.
I also count reviewing AI output as real work. Spotting a weak argument takes judgment. I may become better at that while using the tool, and I would not want to dismiss the gain.
Automation has made flying safer, too. Vanderburgh was teaching pilots to choose the right level, not asking them to reject automation.
The Turkish high-school trial also tested GPT Tutor, a second version that gave hints instead of answers. It removed the exam penalty, though it produced no exam gain over the control group.
I think the way I use the tool matters more than whether I use it. But I cannot settle that by trusting how sure I feel after a good editing session. Reviewing a draft may exercise judgment without exercising everything I need to start one.
Step down a level
When things went wrong, Vanderburgh's advice was to step down a level of automation. Sarah Deener, an editor at AOPA's Flight Training magazine, wrote in 2023 that pilots should hand-fly regularly to keep their stick-and-rudder skills sharp.
My version is a day of core work with the tool off. On other days, it means making the first move before opening the tool: choosing the starting point myself rather than beginning with its draft. After that the tool can draft underneath, and I decide which draft is good enough to reach the reader.
I would pay attention to what feels foreign. That feeling would be a reason to look more closely, not proof that I had lost a skill.
I expect regular tool-off practice to help preserve what I can do alone. I would be wrong if people using AI on every task do as well unaided a year later as those who keep practicing without it.
The magenta line shows the way. It cannot show whether I still know it.
I write for people who want to keep their own hands on the work.
Part of The First Move, on why judgment depends on who moves first, and on who decides what reaches the customer.