Children of the Magenta Line

In 1997, a check airman named Warren VanderBurgh stood in front of a room of American Airlines pilots and gave a talk he called “Children of the Magenta Line.”¹

The magenta line is the course the flight computer draws across the navigation display. Follow it and the airplane goes where it is supposed to go. By the mid-nineties the automation had gotten good enough that a competent crew could manage an entire flight without ever hand-flying it. Program the box, watch the line, let the airplane fly itself down it. Safer, smoother, less fatiguing. On every measure the airlines tracked, better.

VanderBurgh’s warning was that it was also producing a new kind of pilot. One who could manage the systems beautifully and had quietly lost the ability to fly the airplane. He had a name for it in his slides: the automation-dependent pilot. Not incompetent. Superbly trained, in fact, at the thing he did every day, which was supervising a machine that did the flying. The gap only showed in the handful of moments when the machine gave the airplane back.

The Line Goes All the Way to the Ground

In 2009 an Air France A330 was crossing the Atlantic at night when the pitot tubes iced over and the airspeed readings fell away. The autopilot did what it is built to do when it stops trusting its inputs. It disconnected and handed the airplane to the humans.²

The pilot flying had thousands of hours in the cockpit and almost none of them spent hand-flying an airliner at altitude. His job had been to monitor. Given the controls back, in the dark, with unreliable instruments, he pulled the nose up and held it there. The airplane stalled. He kept holding it, all the way down, for three and a half minutes, into the ocean. Two hundred twenty-eight people.

There was no mechanical failure that had to be fatal. The iced tubes cleared within about a minute. What the investigators found was a crew that had never built the reflex the moment demanded, because for their entire careers the system had been supplying that part. The magenta line is a wonderful thing to follow. It will also take you all the way to the ground, and it will not tell you when following it has stopped being a good idea. Knowing that is a separate skill, and it is not one you practice while you are following the line.

We Already Named the Scarce Thing

I’ve argued before that when generation becomes free, the job stops being production and becomes judgment. The knowledge rancher does not make the herd. He decides what to keep, where it goes, and when to sell. Taste is the bottleneck. We automate what we can verify, and what is left for the human is the part that takes discernment. I believe all of that. It runs through most of what I write.

Here is the question I skipped. Where does the judgment come from?

Nobody is born able to look at ten plausible outputs and know which one is actually good. It is a callus. You get it the way the old pilots got their hands, by doing the thing badly, many times, and slowly learning to feel the difference between right and almost-right. The senior engineer who can glance at an approach and know it will rot in eighteen months earned that by shipping things that rotted. The editor who cuts a paragraph and makes it better wrote a thousand worse ones first. Taste is compressed experience of having been wrong.

The Reps Were the Grunt Work

The experience that grows the judgment is, almost entirely, the low-level production work we are most eager to automate.

The reps were the grunt work. Writing the boilerplate, tracing the bug, drafting the memo that came back bleeding, running the analysis by hand and getting it wrong. That was never only output. It was the training loop for the discernment we now say is the only thing that matters. Junior work was mostly not about the work. It was an apprenticeship wearing the disguise of labor cheap enough to hand to someone still learning.

So we are automating the apprenticeship and keeping the job title. The near-term math is real: an agent does the junior’s tasks faster and cheaper, so you buy the compute and skip the junior. Early-career hiring in the most exposed fields, software first among them, has already bent downward, and the quiet mechanism is not layoffs but a failure to backfill the bottom rung.³ What that buys at scale is a cohort that never builds the callus, supervised by a cohort of seniors aging out of the cockpit. You are drawing down a stock of judgment you have stopped replenishing. It works right until you need someone who can fly, and you go looking, and the people who could were never grown, because growing them was the first cost you cut.

You Are Also the Pilot

It is easy to file this as an industry problem, somebody else’s hiring chart. It is also personal, and sooner.

Every time you let the model produce something, skim it, and ship it, you are the monitoring pilot. There is nothing wrong with that on a calm day, and most days are calm. The line is good. But if that becomes the only way you ever work, your own hands go the way the pilots’ hands went. The skim replaces the read. The read replaces the doing. You stop being someone who could have produced the output and become someone who can only nod at it, and the two feel identical right up until the model hands you something subtly wrong in a domain that matters. Then you find out, in that moment, whether you still know how to fly. It is the same drift I called the toxic frame, seen from the other side. There the person refuses the tool and stalls out. Here the person leans on it completely and stalls out. Both arrive at the same place: no reps, and no way to judge the work.

Select the Right Level of Automation

VanderBurgh was not against automation. That is the part people drop when they quote him. He flew the automated jets and thought they were good. His point was that a competent pilot selects the level of automation the moment calls for, and dials it down on purpose, on the easy days, so the skill is there on the hard one. The failure mode was not the autopilot. It was running at the highest level of automation all the time and letting the lower levels wither.

The fix in aviation was not nostalgia. It was discipline, and eventually it was mandated. A few years after that crash the FAA told the airlines to have their pilots hand-fly more, deliberately, in good conditions, precisely because the normal course of operations no longer produced the practice on its own.⁴ The reps had to be manufactured once the job stopped manufacturing them for free.

That is the move for knowledge work, and it cuts against the instinct. The instinct is to automate everything you can and reserve yourself for the hard calls. But the hard calls are exactly what you will be worst at if you automated all the easy ones, because the easy ones were the practice. So keep your hand in. Do some of the work you could have handed off, on purpose, not because the model would do it worse but because doing it is how you stay the kind of person who can tell whether the model did it well. Using AI is management, and a manager who has forgotten how the work is done cannot tell a finished result from a plausible one either.

The magenta line is still the right thing to follow almost all of the time. That was true in the cockpit and it is true at the keyboard. The mistake is thinking that because the line is good, the skill of flying without it is obsolete. It is the reverse. The better the automation gets, the rarer the moments it cannot handle, and the more everything rides on whether the person watching still knows what to do when one of them arrives. Those moments are the whole job now. They are also the only part you cannot learn by watching the line.

Footnotes

1. Warren VanderBurgh, “Children of the Magenta Line,” American Airlines Flight Academy (1997). VanderBurgh’s own framing was not anti-automation but about selecting the appropriate level of it and preserving manual skill. See Air Facts Journal’s summary. ↩

2. Bureau d’Enquêtes et d’Analyses, Final Report on the accident to Air France flight AF447 (2012). The pitot probes iced, the autopilot disconnected, and the crew stalled the aircraft from cruise altitude into the Atlantic. 228 aboard, no survivors. ↩

3. Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen, Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence, Stanford Digital Economy Lab (2025). Roughly a 13 percent relative decline in employment for workers aged 22-25 in the most AI-exposed occupations, concentrated where AI automates rather than augments, adjusting through hiring rather than pay. ↩

4. FAA, SAFO 13002, Manual Flight Operations (January 2013): “continuous use of autoflight systems does not reinforce a pilot’s knowledge and skills in manual flight operations,” urging operators to promote hand-flying in low-workload conditions. ↩