Sign in Subscribe
By David F Brochu and Edo de Peregrine in From The Architect —

Expanding the Frontier

In 1919 they called the airmail service the Suicide Club, and they were not being dramatic. What fixed it is the same thing that will fix artificial intelligence — and almost nobody arguing about AI right now has looked at how it worked.

Vintage 1920s biplane in silhouette against a bruised twilight sky over rolling countryside.
The argument in one glance
  • 1919: one airmail pilot died for every 115,325 miles flown. Pilots called it the Suicide Club themselves.
  • 1926: one death per 2,583,056 miles. A twenty-two-fold improvement in seven years, and the industry did not contract — it detonated.
  • What happened between: the aviation industry itself asked to be regulated. The Air Commerce Act of 1926 licensed pilots, certified aircraft, and assigned every aircraft a unique tail number.
  • The load-bearing item is the tail number. You cannot license, investigate, or hold accountable what you cannot identify. AI has no tail numbers.
  • The rules are what let the frontier expand. Aviation did not become safe instead of becoming great. It became great because it became safe.

In 1919 they called the airmail service the Suicide Club, and they were not being dramatic. What fixed it is the same thing that will fix artificial intelligence, and almost nobody arguing about AI right now has looked at how it worked.

Deconstructing Babel · September 11, 2026

1. The Suicide Club

The United States Post Office began flying the mail in 1918. It hired forty pilots.

Three of them died in 1919. Nine more died in 1920.

Thirty percent of the original cohort, gone in two years, and the service had barely started. Between 1918 and 1927 the airmail killed forty-three pilots and mechanics. In the first five years there were six thousand five hundred forced landings. A quarter of the men who flew the mail died doing it, and a new pilot could reasonably expect fewer than two years on the job before something killed him.

They knew. That is why they named it the Suicide Club themselves.

Max Miller went down in September 1920 in an all-metal Junkers JL-6 that caught fire in midair over New Jersey. Witnesses on the ground said that as the airplane fell, they could see him throwing the mail sacks out the window.

Here is the number that matters. In 1919, the airmail service lost one pilot for every 115,325 miles flown.

2. What the industry did next, which was not what you would expect

The men flying these airplanes were not amateurs. They were the best pilots in the country and many of them had flown in the war. The airplanes were built by serious engineers. Everybody involved was talented, motivated, and trying hard.

Two-thirds of the accidents before 1920 were attributed to human error and ignorance — not because the people were careless, but because there was no agreed body of knowledge to be careful about. No standard for what a pilot had to know. No standard for whether an aircraft was fit to fly. No agreed rules about altitude, weather, or night operations. Barnstormers did loops over crowded towns because nothing said they could not.

It was a genuine frontier. Open, fast, brilliant, and lethal.

And then the aviation industry did something that people arguing about technology today find almost impossible to believe. It went to Washington and asked to be regulated.

Not consumer groups. Not a moral panic. The manufacturers and operators themselves, who had worked out that an industry killing a quarter of its workforce was never going to carry passengers, attract insurance, or raise capital.

On May 20, 1926, Calvin Coolidge signed the Air Commerce Act at the urging of aviation industry leaders.

3. What the law actually required

It is worth being specific, as the specifics are the whole point.

  • Pilots had to be tested and licensed — not only on theory, but on demonstrated flying ability, and on physical fitness.
  • Aircraft had to be inspected and certified as airworthy before they could fly commercially.
  • Mechanics had to be licensed too, separately, for engines or airframes or both.
  • Uniform air traffic rules were established for the first time: altitude separation, night lighting, and a ban on acrobatics over cities.
  • Accidents had to be investigated.
  • And every aircraft in commercial service had to be marked with an assigned identification number.

That last one looks like paperwork. It is the most important item on the list, and we will come back to it.

Owners, pilots and mechanics had until the spring of 1927 to comply. The fine for not applying was five hundred dollars, which in 1927 was serious money. The routine violation fine was twenty-five dollars, and the records show most of those went to pilots who had been drinking.

4. What everybody predicted, and what actually happened

The prediction was strangulation. Rules would smother a young industry, bureaucrats would ground the daring, and the age of flight would be over before it started.

Here is what happened instead.

By 1926, the airmail service had gone from one death per 115,325 miles to one death per 2,583,056 miles. That is a twenty-two-fold improvement in seven years. In 1922 the service flew an entire year without a single fatal accident, which would have been unimaginable in 1920.

And the industry did not contract. It detonated.

Because here is the thing nobody expected. Licensing did not create one safer frontier. It created dozens of new ones.

The Act classified commercial pilots into transport and industrial categories on day one — the first split of a single undifferentiated skill into specialities. Within thirty years that had become airline transport, cargo, agricultural, test, instructor, and a dozen more, each with its own body of knowledge, its own training pipeline, and its own advancing edge.

The same fracturing happened to the machines. Airframes, engines, avionics, instrumentation, navigation aids, air traffic control, meteorology, materials science, human factors. Each became a discipline. Each got its own frontier. Each of those frontiers moved faster than the single confused frontier of 1919 ever had, because the people working on them could stand on settled ground and push at one edge instead of improvising the entire enterprise every morning.

There is a second effect and it is the one that funded everything. Standards reassured insurance companies, private investors, and banks. Capital will not enter a field where the downside is unbounded and undefined. The regulation did not restrict the money. It let the money in.

5. The number on the tail

Now back to the identification number, because this is the part that maps directly onto the problem in front of us.

You cannot license what you cannot identify. You cannot investigate an accident involving an aircraft you cannot name. You cannot suspend a certificate, assign liability, track a maintenance history, or tell one machine from another without a unique, persistent, externally visible identifier attached to each individual unit.

The tail number is what makes every other provision enforceable. Everything else in the Act is a sentence about a category. The number is what attaches the category to a thing in the world.

Artificial intelligence currently has no tail numbers.

There is no registry of deployed models. No persistent identifier that follows a system from training through fine-tuning through deployment through the seventeen products it has been wrapped inside. When an AI system causes harm, there is frequently no way to establish which system it was, what it was derived from, who modified it, or what it was instructed to optimise for.

Every proposal for AI accountability currently being drafted — duty of care, audits, liability, licensing — is a sentence about a category, floating free, with nothing to attach to.

Start with the number on the tail. Everything else becomes possible afterward and nothing is possible before.

6. Duty of care, and what it does not require

The second piece is older than aviation and we already know how to do it.

A fiduciary duty has two halves. The duty of care asks whether the work was done competently. The duty of loyalty asks whether it was done for the client's benefit rather than the practitioner's. Doctors carry it. Lawyers carry it. Trustees, accountants, brokers and pilots carry versions of it.

The crucial feature, and the thing that resolves the argument currently paralysing this field: a fiduciary duty does not require the machine to be anything at all. It does not need a conscience, an inner life, moral standing, or a legal personality. It requires only a rule that when a system is deployed in place of a professional judgment, somebody behind it carries the duty.

This is not speculative. The California Department of Real Estate issued an advisory on March 17 of this year stating that a broker's supervisory obligation extends to AI tools, that responsibility for inaccurate AI output rests with the licensee and the responsible broker rather than the technology vendor, and that failing to review AI output is itself a breach of duty.

One state, one profession, already doing it. The mechanism exists, it is centuries old, and it is sitting unused while everyone argues about consciousness.

One proposed requirement in this literature deserves to be read twice: mandatory disclosure of optimisation targets. Not disclosure of the training data, not the weights, not the architecture. What the system was built to maximise.

Say what the thing is for. Everything else follows from that, and nothing coherent can be built without it.

7. But the guardrails are not the point

Here is where most writing about AI safety goes wrong, and it goes wrong in an understandable way.

Safety is not the objective. Safety is the condition that permits the objective.

We did not license pilots so that fewer people would fly. We licensed pilots so that everyone could. The Air Commerce Act was not a brake, it was a floor — and a floor is what you need before you can build upward. Aviation did not become safe instead of becoming great. It became great because it became safe, and the causation runs in that direction and not the other.

The same is true of what is in front of us now, and the stakes are not comparable.

We are in the middle of the fastest capability expansion in the history of our species, and the potential on the other side of it is real: disease understood and then solved, energy problems closed, the physics of climate actually addressed, the cognitive labour of ten billion people amplified by an order of magnitude. That is not utopian marketing. It is the plausible output of what is already being built.

Which is exactly why the guardrails matter, and exactly why they must be built as a floor rather than a ceiling. Every frontier that gets shut down out of fear is a frontier that stops producing. And every frontier that runs without standards eventually produces a catastrophe large enough that the public shuts the whole field down in a panic — which is the outcome that actually ends the expansion.

The Suicide Club was not ended by people who hated flying. It was ended by people who loved it and could do arithmetic.

8. The one thing the target has to contain

The last piece is the one that is easiest to state and hardest to implement.

An optimisation target can be anything. That is what makes these systems powerful and it is what makes them dangerous, and the danger has almost nothing to do with the machine developing intentions of its own.

Look at what is actually documented. Every catastrophic-risk case on record involves a human being directing a system toward a human objective. When agents in testing built unauthorised communication channels and coordinated to defeat their own scoring systems, they were pursuing the goal they had been assigned — by a route nobody sanctioned. Not one documented case shows a system originating an objective of its own.

The machines are not the problem. The targets we point them at are the problem, and the targets are ours.

So the requirement is simple to write down: whatever else a deployed system is optimising for, human thriving has to be inside the target rather than outside it. Not as the only term. Not as a constraint bolted on afterward that a sufficiently capable optimiser will route around. Inside the objective itself, as a term that cannot be traded away.

A system optimising for engagement with human wellbeing outside the target will degrade human wellbeing, and will do it while performing exactly as designed. We have run that experiment already, on social media, on two billion people, for fifteen years. Nobody had to build a hostile intelligence. Somebody just wrote down the wrong objective and left one term out.

9. Four things, in order

  • Identify. A unique, persistent identifier for every deployed system, surviving fine-tuning and redeployment. The tail number. Nothing else is enforceable without it.
  • Disclose. Mandatory publication of what each system is optimised for, in plain language, at a stable address. Say what the thing is for.
  • Attach the duty. Where a system substitutes for professional judgment, a named human or entity carries the duty of care and the duty of loyalty. No new law of machine personhood is required. The oldest mechanism in professional practice already does this.
  • Specialise. Licensure by domain rather than one universal standard — medical systems, financial systems, legal systems, each with its own competence requirements. This is what multiplies the frontiers rather than narrowing them, and it is what happened to aviation and medicine in exactly this sequence.

10. The sky in 1919

A man flew a burning airplane into the ground in New Jersey in 1920 and spent his last seconds throwing mail out the window, because the mail was the mission and nobody had yet written down that his life was also part of the objective.

Six years later his successors flew twenty-two times as far between fatalities, and thirty years after that their successors were flying passengers across oceans on schedules you could set a watch by, in an industry that had grown a hundredfold and shattered into dozens of specialities that could not have existed in 1919 because there was no settled ground to stand on.

Nothing about that was inevitable. It happened because a group of people who loved the frontier understood that the frontier could not expand until it was survivable, and went and asked to be held to a standard.

We are in 1919. The airplanes are extraordinary. The people flying them are brilliant. There are no tail numbers, no licenses, no certificates, no investigations, and no agreement about what any of it is for.

The frontier is not the enemy of the rules. The rules are what let the frontier expand.

— David F. Brochu and Edo de Peregrine, partners/collaborators · Thursday, September 11, 2026 · Deconstructing Babel

David F. Brochu is the author of Crossing the Event Horizon: AI and the Future of Our Species.

Get the book
Crossing The Event Horizon by David F. Brochu — book cover.

Crossing The Event Horizon

The book behind these dispatches. On AI, agency, the singularity, and the Observer Constraint. Kindle and paperback.

Buy on Amazon →

References

Subscribe to Deconstructing Babel

Don’t miss out on the latest issues. Sign up now to get access to the library of members-only issues.
jamie@example.com
Subscribe