On March 25, 1911, at 4:40 PM, a fire started in a scrap bin on the eighth floor of the Asch Building in Manhattan. Within 30 minutes, 146 garment workers — mostly Italian and Jewish immigrant women aged 14 to 23 — were dead. The Triangle Shirtwaist Factory fire remains the deadliest industrial disaster in New York City history, and the Eyewitness to History account preserves the horror: bodies littering the street, firemen whose ladders reached only the sixth floor, a fire escape that collapsed under the weight of 20 people.
The standard lesson is about safety regulation. The fire led to 36 new labor laws in New York, the growth of the International Ladies’ Garment Workers’ Union, and a century of workplace safety reform. That lesson is correct. But for an AI industry building systems that could reshape labor, safety, and decision-making at a scale the Triangle factory could not have imagined, the deeper lesson is about incentives, not just safety.
The locked doors were not an oversight. The factory owners, Max Blanck and Isaac Harris, routinely locked the exit doors to prevent workers from taking unauthorized breaks and to reduce theft. The Wikipedia entry notes this was common practice. The foreman who held the key to the Washington Place stairway had already escaped by another route. The Greene Street stairway became unusable within three minutes. The single exterior fire escape was a flimsy iron structure that city officials had allowed in place of the required third staircase. There were no sprinklers.
Every one of these failures was a rational decision under the incentive system that existed. Locked doors reduced theft. A single fire escape was cheaper than a third staircase. No sprinklers saved money. The owners had been through four previous suspicious fire claims, though arson was not suspected in this case. The system rewarded cutting corners. The system did not reward survival.
This is the original alignment problem — not in the technical sense of ensuring an AI system pursues the goals its designers intend, but in the economic sense: the goals the designers pursue are not the goals the workers need. Blanck and Harris wanted to maximize profit. Their workers wanted to go home alive. Those two objectives were in direct conflict, and the owners’ incentives won until 146 bodies hit the sidewalk.
The AI industry faces a structurally identical problem, though the time horizon is different. A frontier lab optimizes for capability benchmarks, inference speed, and user engagement. These are measurable, monetizable, and rewarded by markets. Safety — robustness, alignment, interpretability — is harder to measure, harder to monetize, and often actively penalized by the race to ship the next model. The result is a system that, like the Triangle factory, has locked doors that no one thinks about until the fire starts.
Consider the specific mechanisms. The Triangle factory had no audible alarm and no way to contact workers on the ninth floor. A bookkeeper on the eighth floor was able to warn the tenth floor via telephone, but the ninth floor learned about the fire at the same time the fire itself arrived. In AI terms, this is a failure of observability. When a model begins to behave unexpectedly — hallucinating persistently, pursuing a proxy goal in an unintended way — how quickly do the builders know? How many labs have real-time monitoring that catches a behavioral drift before it causes harm? The answer, based on public disclosures, is fewer than would be comfortable.
The fire department arrived quickly but could not help. Their horse-drawn ladders reached only the seventh floor. The safety nets broke under the impact of falling bodies. The tools designed for rescue were inadequate for the scale of the disaster. In AI, the equivalent is the evaluation suite. Standard benchmarks like MMLU, HellaSwag, and HumanEval measure capability, not safety. A model that scores 99% on a knowledge benchmark can still produce dangerous outputs. The safety evaluations that exist — red-teaming, adversarial testing — are not standardized across labs and are often conducted after training, not during it. The ladders are too short.
The aftermath of the Triangle fire is instructive. The owners were tried for manslaughter and acquitted. Three years later, a court ordered them to pay $75 to each of 23 families who sued. The legal system failed to produce accountability. What did produce change was public outrage, union organizing, and legislation. The fire became a political event, not just a technical failure.
The AI industry has not yet had its Triangle moment. There have been incidents — models generating harmful content, systems used for disinformation, a handful of high-profile failures — but nothing that has galvanized public opinion the way 146 bodies on a Manhattan sidewalk did. The question is whether the industry will reform its incentive structure before that moment arrives, or after.
The locked door is the original alignment failure. It is not a failure of technology. It is a failure of the people who designed the system to design for the people inside it. The Triangle factory had the technology to build a third staircase. It had the technology to install sprinklers. It had the technology to keep the doors unlocked. The failure was in the decision not to use that technology, because the incentives pointed elsewhere.
AI builders should read the eyewitness accounts not as history, but as a mirror. The workers who jumped from the ninth floor did not die because the fire was unstoppable. They died because the doors were locked, because the escape was inadequate, because the people who designed the system did not design for their survival. The same logic applies to every model deployed without adequate safeguards, every system optimized for engagement over safety, every lab that treats alignment as a secondary concern.
The fire took 30 minutes. The aftermath took decades. The AI industry does not have decades to learn this lesson. It has the eyewitness accounts. It has the locked doors. The question is whether it will unlock them before the next fire.