In 1910, flying was a way to die in public. Insurers wouldn’t touch it. Newspapers ran obituaries for aviators the way they now run product launches. Fifty-three of them died that year alone, and in some of the early flying clubs the death rate among members ran as high as 87 per cent. The aeroplane was real, it flew, and almost everyone who got close to it eventually paid for the privilege.
And yet the aeroplane was also, unmistakably, the future. Both things were true at once. That’s the part we forget when we tell the story backwards from the comfort of cruising altitude.
I think about 1910 a great deal when I read about AI. Not because the analogy is neat — analogies are never neat — but because it’s the only precedent we have for a technology that is simultaneously transformative and lethal, adopted by people who could see the second fact clearly and the first fact only dimly.
The aeroplane didn’t get safe by being young. It got safe because a world war forced a decade of investment that peacetime curiosity never would have funded. Between 1914 and 1918, aircraft went from wood-and-canvas contraptions that a strong headwind could kill you in, to fast, armed, manoeuvrable machines with synchronised machine guns firing through spinning propellers. Nations didn’t fund this because they wanted safer skies. They funded it because they wanted to win. Civil aviation spent the next decade quietly inheriting the engineering.
Then came the slower part, the part that actually mattered: a century of catastrophe followed by regulation followed by institutional memory. Every major safety advance in aviation has a body count attached to it. Cockpit voice recorders exist because we needed to know why planes came apart. Standardised checklists exist because a Boeing test pilot forgot to release a control lock in 1935 and four men died finding that out. The whole apparatus of modern flight safety — redundant systems, mandatory reporting, independent investigators, a regulator with actual teeth — was built one disaster at a time, not because anyone had the foresight to build it in advance.
It worked. By 2022, commercial aviation in the US had a fatality rate of roughly 0.003 deaths per hundred million miles travelled — orders of magnitude safer than driving, safer than trains, safer than buses. In 2023, scheduled commercial jets recorded zero fatal accidents worldwide. A technology that once killed the majority of the people brave or foolish enough to try it became, within a single human lifetime, the safest way to move a body through space.
It also never stopped being a problem. Aviation is responsible for something like 2 to 3 per cent of humanity’s climate impact today, and that number is not shrinking on its own. We solved the crashing. We did not solve the burning. Safety and harm turned out to be two separate engineering problems wearing the same coat, and fixing one told us nothing about the other.
This is where the analogy earns its keep, because AI has the same structure. Right now it is genuinely dangerous in the way aeroplanes were genuinely dangerous in 1910 — not metaphorically, but in the specific, ordinary sense that people deploy it and things go wrong, badly, in public, regularly. Gartner expects 40 per cent of agentic AI projects to be cancelled by the end of 2027. RAND puts the AI project failure rate at roughly 80 per cent, double that of ordinary IT. None of this means the technology is a dead end, any more than a 1910 death rate meant the aeroplane was one. It means we are early, and early is dangerous, and pretending otherwise is how people get hurt.
I’ve spent two decades watching organisations mistake the absence of a crash for the presence of safety. It’s the same mistake, at smaller scale, that aviation made repeatedly for its first thirty years: nothing has gone wrong yet, therefore nothing is wrong. The instrument panel was empty because nobody had built the instruments, not because there was nothing to measure.
The uncomfortable question the analogy raises is whether AI needs its own war — some forcing function violent enough to compress thirty years of safety engineering into three. I don’t think it does, and I don’t think we should wait to find out. The alternative is what aviation eventually built without a war: independent investigation, mandatory disclosure, institutions with the authority to ground something first and ask questions after. We have not built that for AI. We have built the marketing department instead.
And even if we do build it — even if AI gets its NTSB, its checklists, its redundant systems, its hundred quiet years of learning from failure — history’s other lesson is waiting on the far side of that achievement. Aviation proved you can make a dangerous machine safe. It never had to prove you could make it clean. We should assume AI’s bill, whatever it turns out to be, arrives in two separate currencies, and that solving the first one will feel, wrongly, like solving both.
We are still waiting for AI’s Fort Myer.
Sources: theaviationvault.com and afhistory.af.mil (early aviation accident rates and 1908/1910 fatality figures); History Guild and EBSCO Research Starters (WWI’s acceleration of aircraft technology); Journalist’s Resource and National Safety Council Injury Facts (US transportation fatality rates by mode, 2022 data); Our World in Data and Air Transport Action Group (aviation’s share of global CO2 and transport emissions); Gartner newsroom (agentic AI project cancellation forecast, 2026); RAND Corporation analysis via Folio3 AI (enterprise AI project failure rate, 2025-26).
Leave a Reply