In management training it is often told as a five-whys triumph. A monument in Washington is crumbling, and asking why five times leads through cleaning, birds, spiders and midges to the floodlights. Turn them on an hour later and the problem goes away.
Most retellings get the details wrong. But at the Lincoln Memorial, in 1990, the fix really was tried. Delaying the lights by an hour cut the midges by about 85 percent, said the entomologist who worked on it. “But tourists who had driven hundreds of miles to have their photographs taken at the monuments were not happy. They complained every day, and the lights went back on.”
Five whys, one how
The method comes from Taiichi Ohno, who pioneered the Toyota Production System. “Having no problems is the biggest problem of all,” he liked to say. When one turned up, his advice was blunt: “Observe the production floor without preconceptions. Ask ‘why’ five times about every matter.”
Toyota retold his favourite example in 2006. A welding robot stops in the middle of its work. Why? A fuse blew because the circuit overloaded. Why? The bearings weren’t lubricated enough and locked up. Why? The oil pump isn’t circulating enough oil. Why? Its intake is clogged with metal shavings. Why? There’s no filter on the pump.
Replacing the fuse gets the robot moving again. A filter stops this kind of clog from stopping it next time. “By repeating why five times,” Ohno wrote, “the nature of the problem as well as its solution becomes clear.” Five whys, he said, equal one how.
The how comes from looking. “Data is, of course, important in manufacturing,” Ohno remarked, “but I place greatest emphasis on facts.”
As quick as a flash
Teruyuki Minoura learned the method from Ohno and later ran global purchasing at Toyota. In 2003 he said it often didn’t work as well as it should, because people “end up falling back on deduction.” Ask them why, and “they reel off five causes as quick as a flash.” Ask why again for each of those, and you get so many causes that “you end up totally confused as to which of them is important.” The only correct way to answer a why, he insisted, was to go and look.
Imagine a warehouse manager who wants to know why the delivery van leaves at 8:40 when it should leave at 8:00. At her desk the answers come fast: the driver is slow, traffic is bad, the new hire doesn’t know the routes. The next morning she stands on the loading dock instead. The driver is there at 7:50, waiting for the last pallet. The pallet is waiting for a shipping label. The labels don’t print until the overnight orders are processed, at 7:45. None of the guesses from her desk had mentioned a printer.
One path, one cause
Alan Card, a patient-safety researcher who traced the monument story’s real version in 2017, goes further. For investigating accidents in healthcare, he argues, five whys should be abandoned. It forces one path through a problem, aims at a single root cause, and assumes the last why is the best place to step in. Even the monument’s first why could have gone elsewhere, he points out: to rain, acid rain or freeze-thaw damage. Real accidents rarely have one cause. Used well, the method can branch into several chains, and the count of five is a habit, not a rule. When GitLab’s database was wiped in 2017, its public postmortem ran several chains of whys at once, from the deleted directory to backups nobody had tested.
The monument story shows the other gap. The chain was right as far as it went. It never asked who else needed the lights.
Deduction at machine speed
A language model can be the deduction Minoura warned about, at machine speed. Ask it why something broke and it can quickly produce explanations that sound plausible.
Meta’s approach is to feed it evidence first. At Meta, a system narrows thousands of recent code changes down to a few hundred using simple rules, then a Llama 2 model fine-tuned on past investigations ranks the most likely five. Tested on old investigations, using only what was known at the start of each one, the true cause was among the five 42% of the time. Meta’s engineers add that it “can potentially suggest wrong root causes and mislead engineers,” so the tool holds back answers it has low confidence in.
In 2024, researchers at HSBC and the Hong Kong University of Science and Technology described a five-whys agent built on GPT-4o and a map of their own systems’ history. In their preprint, not yet peer-reviewed, they report that about 70% of incidents previously blamed on management or outside vendors turned out to come from problems in their own code.
An agent can ask why all day. Its answers are still candidates. Before anyone acts on one, it has to be checked against the logs, the scene or a reproduction of the fault, the AI version of going to look.