At 5:16:11 p.m. on November 9, 1965, a backup relay at the Sir Adam Beck No. 2 power plant, on the Ontario side of the Niagara River, tripped one of five lines carrying power to Toronto. The other four took the load, went past their own settings and dropped out in about two and a half seconds. The power reversed and surged south into New York State.
About 12 minutes after that first relay moved, most of lower New England and southern New York was dark. An estimated 30 million people in the United States and Canada lost power.
Every utility checked its equipment. No faulty equipment was found. It took until November 15 to trace the start to Beck.
Nothing there had failed either. The relay had been set in 1963, at about 375 megawatts, when the load on those lines was much lower. That evening the average flow on the first line was 356 megawatts, and normal fluctuation did the rest. Ontario Hydro’s officials said the people running their system “were not aware” of the setting.
A relay or a grid
The obvious culprit was a stale number on one relay. By the time the Federal Power Commission reported to President Johnson on December 6, Ontario Hydro had already acted so the same relays could not trigger another failure.
The Commission didn’t stop there. “The risk of a precise repetition of the power failure is not the main point,” it wrote. The real question was whether “other previously unanticipated malfunctions” could set off the same kind of cascade. It asked every utility to reexamine how its system was designed and run.
Before you fix a problem, ask whether it’s a one-off or a case of a wider problem.
Peter Drucker made that the first step of a decision in The Effective Executive, in 1967, using this very night as an example. The first question, he wrote, is “Is this a generic situation or an exception?” A generic problem “always has to be answered through a rule, a principle.” The exceptional “can only be handled as such and as it comes.” By first explanations, he noted, the blackout had looked “truly exceptional.” He thought otherwise: it was the first of what were “likely to become fairly frequent malfunctions unless generic solutions are found.”
The rule that fit the wrong night
Drucker saw a second mistake that evening.
In New York City, the load on Consolidated Edison’s generators jumped by 600 to 800 megawatts, then past 1,000. The industry’s operating guide said that in a disturbance “a prime consideration is to maintain parallel operation” with neighboring systems, so they can help each other. No standing instruction said when to cut loose. The operator on duty “did not make an immediate clear-cut decision,” the Commission found, and in hindsight timely load shedding “might have avoided a city-wide blackout.”
In Drucker’s reading, the engineers had “applied the right rule for a normal overload,” while their own instruments signaled something extraordinary. That mistake, treating a new event as one more case of an old problem, he called equally common. The reverse, treating something generic as a series of one-offs, was “by far the most common mistake.”
Where the sorting slips
Drucker offered this as a working rule from experience, not a tested finding, and the sorting is hardest when it matters most. His own fourth category, the first sign of a new generic problem, looks like a freak the first time. One event can’t tell you which it is.
Rules cost something too. The Commission warned that cutting ties too early would hurt neighbors in trouble. A rule written for a true one-off makes every later case pay for it. “Truly unique events are rare, however,” Drucker wrote.
The third crushed parcel
Imagine you run a small online shop. A customer emails a photo of a crushed box. You apologize, refund, and move on, which is the right way to handle a one-off.
Two weeks later there’s another, from a different courier and a different town. Then a third. Refunding each one would treat them as three accidents. Side by side, the three orders share one thing: they all went out in the thinner box you switched to last month.
Going back to the old box is a rule. It costs more on every parcel, including the ones that would have arrived fine, so it’s worth deciding once, on purpose, instead of paying one refund at a time.
Rocks and glue
In May 2024, people shared odd answers from Google’s new AI Overviews: how many rocks to eat, and, from a forum joke, glue to keep cheese on pizza. Before those screenshots went viral, wrote Elizabeth Reid, Google’s vice president of Search, “practically no one” had asked the rocks question.
Removing single answers was part of the response: Google said it took action on the small number of overviews that broke its content policies. But, Reid wrote, “we don’t simply ‘fix’ queries one by one.” The company looked across examples for patterns and made “more than a dozen technical improvements,” among them better detection of nonsensical queries and less satire, humor and user-generated content that could mislead.
A model adds a wrinkle: the same prompt can answer differently next time, so one bad output may not be a pattern. Keep a short log of each correction you make to an AI’s work. When the same kind of correction shows up twice, write it into your instructions as a rule. Then rerun a few earlier tasks, several times each, and read what changed. One clean run is a lead, not proof the rule holds.
At Beck, the number on the relay had not changed since 1963. The flow under it had.