Children of the Magenta Line
One industry has been flown for thirty years. It wrote the doctrine for the fourth rung in 1997, and then buried 231 people learning what the doctrine costs to ignore. The rule is not trust the autopilot. It is: when the machine surprises you, climb down.
In April 1997 an American Airlines captain stood in front of a room of pilots at the airline’s flight academy and named a new kind of aviator: children of the magenta line, pilots who follow the computed path on the display and have stopped flying the airplane. His remedy was not less automation. It was a rule about direction. The airplane has levels of automation; when it does something you did not expect, step down one. Sixteen years later, after two crashes, the regulator put the same rule in writing.
01 The magenta line
On a Boeing navigation display the route the flight computer intends to fly is drawn in magenta. The color is not decorative; it is the manufacturer’s code for data the computer is using to command the airplane. Airbus draws the same line in green, so the phrase belongs to Boeing cockpits, but everyone in the industry knows what it means. Captain Warren Vanderburgh used it in April 1997, in a lecture that has circulated under the title Children of the Magenta ever since, to describe a pilot who has become a passenger on their own flight deck: someone who watches the line and no longer knows, in their hands, what the airplane is doing.
What made the lecture last was not the diagnosis. It was the structure he put under it. Vanderburgh taught that the cockpit has three levels of automation. At the bottom you fly the airplane by hand. In the middle you tell the autopilot a heading and an altitude and a speed, and it holds them. At the top the flight-management computer follows a plan it was given hours ago, laterally and vertically, and the crew watches it do so. His rule was about moving between the levels, and it ran in one direction. When things go wrong or get complicated, step down. Do not sit at the top level typing a correction into the plan while the airplane does something you did not intend. Click off. Take a level back. Fly the airplane, then think.
Lay the three levels beside the ladder from the last essay and the fit is nearly exact. Hand-flying is the crawling and walking rungs: you do the verb, with more or less help from the display. The middle level is running: you type an instruction, hold 250 knots, and the machine does the doing. The top level is flying, in the sense the last essay gave the word. You type nothing. The plan was made earlier, the machine executes it, and you are flown. Vanderburgh was describing the fourth rung in 1997, for the one profession that had already reached it, and the only doctrine he offered for living there was a rule for leaving it.
The doctrine for the fourth rung is a descent rule. It says nothing about how to climb. It says exactly when to come down.
02 What the ratio does at the fourth rung
The last essay put the Reality Equation under the ladder. Reality is a ratio, what arrives over what was expected, and the design goal of every product built for the fourth rung is to hold that ratio at one: the machine does the thing you would have asked for, before you asked, and you feel nothing, because there was nothing unexpected to feel. Unitless stability. That is the promise of being flown.
Now watch what Vanderburgh’s rule says in those terms. The airplane did something you did not expect is the ratio leaving one. The machine at the top level produced an Actual that did not match the crew’s Expectation, and the surprise arrived. The rule does not say to correct the machine at its own level. It says to change who holds the verb. And that is the only move available, because at the fourth rung you own no verb at all. You cannot touch the numerator; what the machine did is done. You cannot touch the denominator; what you expected was brought before the airplane moved. The one thing an Actualizer can act on is the Reality itself, and the one action that is always available is to take a verb back. Once the surprise has arrived, more than one action remains possible, and stepping down is the action that returns you to a rung where your hands are on something.
This is why the rule runs downward and never up. Climbing the ladder is not a decision anyone takes in the moment; it is what the record does to the machine over months, as the artifacts of your instructions thicken its prediction of you. But a surprise is a moment. The response to a moment has to be available in the moment, and only the descent is.
03 The two ways it kills
Aviation learned the price of ignoring the rule twice in four years, and the two accidents fail in different places. They are worth keeping apart, because the fourth rung of an ordinary life will fail in both.
Over the Atlantic, at night, Air France 447 flies into ice crystals that block its airspeed probes. The speed readings become inconsistent, the autopilot disconnects, and the Airbus A330 hands itself back to the crew. The pilot flying pulls the nose up. He keeps it up. The airplane stalls near 38,000 feet and falls for three and a half minutes, and the BEA’s final report, presented at Le Bourget on 5 July 2012, finds that the crew never understood they were in a stall and so never attempted to recover from one. Two hundred and twenty-eight people.
That is the first failure. The machine stepped down on its own, without being asked, and handed the verb to hands that had never been trained to hold it at that altitude. The autopilot did exactly what the doctrine says: it reduced its level. The crew had no level to step down to, because nothing in their training and nothing in their daily flying had ever put the skill there. In the framework, their denominator held no record of hand-flying an A330 at cruise, because the record of their careers held none. Nothing had ever arrived to train it. Calibration debt, called in all at once.
On approach to San Francisco in clear daylight, the crew of Asiana 214 select a descent mode on their Boeing 777 and pull the thrust levers to idle. That combination puts the autothrottle into a mode called HOLD, in which it does not control airspeed. Neither pilot notices. The airplane slows, sinks below the glide path, and strikes the seawall short of the runway. Three of the 291 passengers die. The NTSB, adopting its report on 24 June 2014, finds that the pilot flying “had an inaccurate understanding of how the autopilot flight director system and autothrottle interacted to control airspeed.” Its word for what went wrong is mode error, and its analysis names the underlying condition: a lack of mode awareness, which leads to what the researchers call automation surprise.
That is the second failure, and it is the more interesting one for the ladder, because the skill was present. The crew could have flown the approach by hand. They did not, because they believed they were at the top level, being flown, and they were not. The machine was at the middle level, holding what it had last been told and nothing else. Their denominator was wrong about the world only at one address: which rung the machine was on.
The last essay recorded, without much comment, that OpenAI withdrew its fourth-rung product in June and moved the work into scheduled tasks, and called that the third stage wearing the fourth stage’s clothes. Read it again with Asiana in mind. A scheduled task is the autothrottle in HOLD. It does exactly what it was last told, at the time it was told to, and it is not watching anything in between. A person who believes it is watching has made a mode error. Nobody has died of one yet in this industry. But the shape is identical, and a life run at the fourth rung will produce it daily: the reply you thought it had sent, the renewal you thought it was tracking, the morning brief that ran at six and has not looked since. The failure is not that the machine is bad at its rung. It is that you were wrong about which rung it was on.
The ladder, with its descent rule
| Rung | What you type | Vanderburgh’s level | The trigger to step down | Where you land |
|---|---|---|---|---|
| Fly | Nothing | Top: the computer follows its plan | The machine does something you did not expect | Run — give the instruction yourself |
| Run | An instruction | Middle: the autopilot holds what you dialled in | The instruction produces the wrong thing | Walk — ask, and read the answer |
| Walk | A question, with memory behind it | Bottom: hand-flying, with the display helping | The answer assumes a you that has moved on | Crawl — supply the context by hand |
| Crawl | A question, every time | Bottom: hand-flying, raw data | — | There is no lower rung. This is your hands. |
04 The regulator’s answer
What aviation did next is the part I want to carry across, because it is the part no one is proposing for the consumer version of the fourth rung.
On 4 January 2013 the FAA issued a Safety Alert for Operators, number 13002, titled “Manual Flight Operations.” Its language is careful and its content is blunt: “continuous use of autoflight systems could lead to degradation of the pilot’s ability to quickly recover the aircraft from an undesired state,” and so operators are encouraged to promote manual flight operations when appropriate. In September of that year a working group of the FAA’s rulemaking committee and the Commercial Aviation Safety Team dated its report on how crews actually use flight-path management systems; the FAA released it that November. It found vulnerabilities in pilot knowledge and skill for manual flight, found that mode confusion errors continued to occur, found that pilots sometimes over-relied on the automation, and recommended, first among its recommendations, that pilots be given opportunities to practice the skills. In May 2017 the FAA reissued the alert as SAFO 17007, “Manual Flight Operations Proficiency,” because the first one had not been enough.
Read what that is. The regulator is not telling airlines to automate less. The magenta line is safer than a tired human on a long night, and everyone in the room knows it. The regulator is telling airlines to schedule the descent: to have crews step down a level on purpose, on ordinary days, when nothing has gone wrong, so that the level is still there when something does. The surprise that would have forced the practice no longer arrives; that is what the automation is for. So the practice has to be manufactured, by policy, against the grain of a system whose whole purpose is to make it unnecessary.
Lisanne Bainbridge saw the shape of this in 1983, in a four-page paper called “Ironies of Automation” that the aviation reports still cite. Automate the parts of a task you know how to automate and you leave the human the parts you do not, which are the hardest, at the moment their skill at those parts has decayed from disuse, and you ask them to monitor a system they no longer practice controlling. The irony is not a flaw in any one design. It is what automation does to the operator by construction, and the only answer anyone has found in forty years is the one the FAA wrote down: keep the lower level alive by using it when you do not need to.
The calibration essay reached the same place from the other side and called it the manual mile: at intervals, run one absorbed process by hand, not for the output but for the error signal. What aviation adds is the finding that this will not happen on its own. A crew left to choose will choose the magenta line every time, because it is better, and their hands will go quiet, and the regulator had to make the descent a requirement precisely because no surprise was left to make it a habit. The same will be true of a life. Nobody steps down from being flown for the pleasure of it.
05 The rule, translated
So here is the doctrine for the fourth rung, borrowed whole from the only people who have lived there, and restated for a person rather than a cockpit.
Know which rung the machine is on. Not which rung you are paying for; which one it is actually operating at, right now, for this task. A scheduled task is the third rung. A memory is the second. Only a thing that acts on a prediction of you, unprompted, is the fourth, and as of this month nothing at consumer scale does it. Asiana’s crew did not lack skill. They lacked mode awareness, and it killed three people in daylight.
When it surprises you, step down. The reply it sent in a tone you would not use, the table it booked that you did not want, the morning it reprioritised in a way that is not yours: these are the ratio leaving one, and the response is not to add a preference to the autopilot and let it try again at the same rung. It is to take the verb back for that task. Give the instruction yourself. If the instruction produces the wrong thing, ask the question yourself and read the answer. Fly the airplane, then think.
Schedule the descent. This is the one the industry had to write into policy, and it is the one a person will not do unprompted. On a day when nothing has gone wrong, do one of the absorbed things by hand: the route without the guidance, the numbers without the dashboard, the letter without the draft. Not because the machine would have done it worse. Because the rung has to be exercised or it is not there, and the surprise that would have exercised it for you is exactly what you paid the machine to remove.
The last essay ended on a question: if you type nothing and something arrives, who is at the controls. Vanderburgh’s answer is that the question is the wrong shape. The controls are never the point. The point is whether you can still take them, and that is decided on the ordinary days, by whether you ever do.
06 Where this stops being true
- Already true
- Vanderburgh’s April 1997 lecture and its three-level, step-down doctrine are on the record, and his colleagues have confirmed the content. AF447 (BEA, final report 5 July 2012) and Asiana 214 (NTSB, adopted 24 June 2014) are the two accidents described, with the findings quoted from the reports. SAFO 13002 (4 January 2013) and SAFO 17007 (4 May 2017) say what this essay says they say. The PARC/CAST working-group report is dated 5 September 2013 and was released by the FAA in November 2013. Bainbridge’s paper is Automatica 19(6), 1983. The magenta convention is Boeing’s; Airbus draws the active route in green.
- What has to happen for the translation to hold
- A consumer product has to actually reach the fourth rung and stay there, which none has. Then the two failure shapes have to appear in the record: people arriving at a surprise with no skill to step down to, and people making mode errors about which rung their machine is on. If a year of fourth-rung products passes and the failures that show up are of some other kind entirely, the aviation frame was borrowed too eagerly.
- Where the shorthand breaks
- Aviation has checkrides, recurrent training, a regulator with the power to ground an airline, and two pilots who can watch each other. A person with a stack has none of those, and the stakes are almost never fatal, so “schedule the descent” has no enforcement behind it and may simply not happen. The three-to-four mapping is mine, not Vanderburgh’s. And the NTSB’s word is mode error; “mode confusion” is the working group’s phrase, and I have used both.
- Where I am probably wrong
- Vanderburgh could tell pilots to step down because every pilot in the room had learned to hand-fly first. The ladder in the last essay assumes the same: you climb it, so the lower rungs are in your record. The people arriving now may not climb it. A person who meets the machine at the fourth rung, at twelve years old, has no second rung to descend to, because they never held the verb at all — and for them the doctrine is empty. AF447 in that case is not a cautionary tale. It is the default. If that is the population that being flown produces, then the descent rule is a doctrine for my generation only, and the real question, which I have not answered, is what a ladder looks like for someone who was never on the ground.
Sources
- W. Vanderburgh, “Children of the Magenta,” American Airlines Flight Academy, Advanced Aircraft Maneuvering Program, April 1997 — as transcribed by 99% Invisible and as described by his co-presenters in Air Facts (2020).
- BEA, Final Report on the accident on 1st June 2009 to the Airbus A330-203 registered F-GZCP operated by Air France flight AF 447, presented 5 July 2012.
- NTSB, Descent Below Visual Glidepath and Impact With Seawall, Asiana Airlines Flight 214, AAR-14/01, adopted 24 June 2014.
- FAA, Safety Alert for Operators 13002, “Manual Flight Operations,” 4 January 2013; SAFO 17007, “Manual Flight Operations Proficiency,” 4 May 2017.
- PARC/CAST Flight Deck Automation Working Group, Operational Use of Flight Path Management Systems, final report dated 5 September 2013, released by the FAA November 2013.
- L. Bainbridge, “Ironies of Automation,” Automatica 19(6) (1983), 775–779.
- Boeing 737NG FCOM, Flight Management and Navigation: display color conventions — magenta for data used by the FMC to command the airplane; Airbus A320 EFIS guide — the active flight plan in green.
2 thoughts on “Children of the Magenta Line”