A WWI shell factory and a global mortality study, produced a century apart by people who never spoke, drew the same line: the band where extra work stops making output is about the band where it starts making death.
During the First World War, Britain had a problem that reads, a century later, like a controlled experiment nobody would be allowed to run today. It needed shells, urgently, in quantities that dwarfed anything its factories had produced in peacetime, and it responded the way every organization under deadline responds: it ran the workers longer. Twelve-hour days, seven-day weeks, month after month. And because these were munitions factories feeding a national emergency, someone was counting. The Health of Munition Workers Committee kept meticulous production records, worker by worker, hour by hour, and those records sat in archives until the economist John Pencavel dug them out and, in a 2015 paper in The Economic Journal, extracted from them one of the most quietly devastating findings in the study of work.
Here it is, in his numbers. As weekly hours climb, output climbs too, but at a decreasing rate once you pass about 49 hours. Total output peaks at around 63 hours a week. And past that peak it falls: a worker on 70 hours produced, in Pencavel's phrase, little more than a worker on 56. The fourteen hours between 56 and 70 were, for the purposes of shell production, spent producing nothing. Not less per hour. Nothing on net. A seventh of the person's working life, and the wartime factory would have gotten the same number of shells if they had gone home.
If you work in a field that valorizes the long week, that treats the sixty-hour crunch as evidence of seriousness and the eighty-hour push as heroism, this number should land somewhere between inconvenient and alarming. Because the finding underneath it is not “long hours have diminishing returns,” a phrase soft enough to nod at and ignore. The finding is that an hour is not an hour.
Pencavel is precise about this, and the precision is what makes it useful. “At 35 hours,” he writes, “an additional five hours to the length of the working week has consequences for the effective labor input that are quite different from an additional five hours starting at 48 hours.” Read that twice. Five hours is not a fixed quantity of work. The five hours you add going from 35 to 40 are real, productive, roughly worth what you'd expect. The five you add going from 48 to 53 are a different substance entirely, thinner, more errored, more likely to require redoing tomorrow what you damaged tonight. The clock says both blocks are five hours. The output says they are not remotely the same thing.
This is why measuring work in hours is a category error, and it is an error the entire apparatus of employment is built on. We pay by the hour, schedule by the hour, brag by the hour, and the hour, as a unit of output, is nonlinear to the point of dishonesty. Averaging across it makes things worse, not better: a team's “average 55-hour week” hides the fact that the hours past 49 for each person are a fundamentally different and mostly worthless commodity than the hours before. The average, Pencavel notes, “may be an especially inappropriate summary indicator of a highly nonlinear effect.” You cannot understand a curve this bent by reporting its mean.
Now, the honest objection, and it is a real one: Pencavel's workers were assembling munitions with their hands, in 1916. Manual output on a factory line is not knowledge work, and the specific number, the cliff at 49, the peak at 63, should not be teleported onto a software team without an argument. But look at which direction the argument runs. Physical labor degrades gracefully; a tired body on a line still fills shells, just slower. Cognitive work degrades catastrophically, because its failures are silent and compounding, the subtle bug written at hour 55 that costs three people two days in three weeks, the strategic misjudgment made by a depleted mind that no amount of subsequent hours can un-make. If anything, the knowledge-work version of Pencavel's curve should bend earlier and harder than the munitions version, not later. The person confidently shipping code at hour sixty is the modern equivalent of the exhausted worker producing shells that don't fit, except nobody is measuring the shells.
That is the productivity half of the story: above some threshold in the low-to-mid fifties, the hours stop producing. The second half is worse, because it is not about what the hours fail to make. It is about what they make of you.
In May 2021, the World Health Organization and the International Labour Organization published, in the journal Environment International, the first global estimate of the disease burden of long working hours. They pooled dozens of studies covering more than 1.6 million people across 154 countries and asked a blunt question: how many people are working hours long enough to kill them, and how many does it kill?
The answer, for the year 2016: 745,000 deaths. Stroke accounted for 398,000 of them and ischemic heart disease for 347,000, and every one was attributed to working 55 or more hours a week. Compared with a 35-to-40-hour week, the 55-plus workers carried an estimated 35% higher risk of stroke and a 17% higher risk of dying from heart disease. And the toll is growing: deaths from long hours rose 29% between 2000 and 2016, heart-disease deaths by 42%. This is not a receding problem from a rougher past. It is expanding, and it expanded fastest in exactly the era that invented the always-on knowledge job.
But the single most important detail in the WHO/ILO estimate is buried in its demographics, and it explains why this catastrophe is invisible to the people living inside it. The deaths cluster among people aged 60 to 79, and the fatal hours were the ones they had worked earlier, between roughly 45 and 74. You do the damage in your fifties. You collect it in your seventies. The stroke that ends a life at 68 was purchased by the calendar of a 52-year-old who felt, at the time, completely fine, who had in fact never felt more productive, who would have laughed at the suggestion that the quarter's crunch was a health decision.
Sit with the structure of that, because it is the reason no amount of individual wisdom fixes it. Every functioning feedback loop humans learn from is short: touch the stove, feel the burn, don't touch the stove. Long working hours have a feedback loop measured in decades, which is to say no functional feedback loop at all. The consequence is so far removed from the behavior, and so thoroughly confounded by everything else that happens to a body across twenty years, that experience literally cannot teach it. A person could work themselves toward a stroke for their entire career and never once receive the signal that would let them connect the two. The body keeps the score, and it settles the account long after the game everyone thinks they were playing has ended.
Now put the two literatures beside each other, because they were produced by people who never spoke to each other and had no reason to expect their numbers to relate. An economist reading handwritten production logs from a WWI shell factory found that output stops meaningfully rising somewhere around 49 to 63 hours. A global epidemiological consortium reading modern medical records across 194 countries found that mortality risk turns sharply upward at 55 hours. Two completely independent methods, two different centuries, two different outcomes, one measuring shells and the other measuring strokes, and they land within a handful of hours of each other.
The honest way to state this is carefully, because it is easy to oversell and 49 is not 55: these are different thresholds, from different populations, for different outcomes, and the convergence is suggestive rather than a matched constant. But even stated with all that hedging intact, the shape is stark and, as far as I can find, unremarked. The band of hours where additional work stops producing output is approximately the band where it starts producing death. Below it, hours buy things: shells, code, money, a life. Somewhere in the low-to-mid fifties, the hours cross a line, and above that line you are spending your life to purchase, on the output side, nothing, and on the mortality side, a bill payable in twenty years. The overtime past that point is the worst transaction available to a human being: maximum cost, delivered late; zero benefit, delivered never.
There is a reason the number that keeps showing up is in the neighborhood of the one a legislature settled on almost a century ago, and it is not that the legislature knew any of this. When the Fair Labor Standards Act was signed on June 25, 1938, and phased the overtime threshold down to forty hours by 1940, it set forty not as a scientific optimum but as the outcome of decades of union struggle and political compromise. The part most people misremember is that the law never said forty hours was the right length of a working week. It said forty was the point above which an employer had to pay you a penalty rate to keep you. Forty was a ceiling, a line past which your time became expensive on purpose, and the culture later mistook it for a target, something to admire and then exceed to prove one's dedication. The culture inverted it, turning a legal maximum designed to protect workers into a floor of expectation and then a starting line, but the statute's forty was always a ceiling. What the productivity and mortality data suggest, arriving from two directions the 1938 Congress had no instruments to measure, is that the ceiling was in roughly the right place. They guessed, from politics, at a line that the economics and the epidemiology would independently draw decades later.
So the practical thing to carry out of all this is not “work less because it's virtuous,” which persuades no one under deadline. It is a correction to the arithmetic you are actually doing. If you are a builder or a leader who reaches for hours when a project is behind, you are implicitly modeling work as linear, one more hour equals one more unit of progress, and that model is not slightly wrong, it is wrong in the way that matters, because it is exactly right up to about fifty hours and then it inverts. Past the cliff, the additional hour is not neutral. It produces negative output through the errors it introduces, and it draws down a health account whose statement will not arrive until you can no longer connect it to the cause. The most valuable measurement you can make this quarter is not of hours, which you already track, but of shipped output against hours, which you almost certainly don't. Do it for a month and you will probably find your own version of Pencavel's curve, and it will bend earlier than you want it to. The asymmetry of that bet is the reason to take it seriously: if the skeptics are right and the curve is flat for you, capping the week at the ceiling costs you nothing, because by their own account those hours were producing anyway. If the data is right, you buy back real output and defer a real mortality bill. There is no version of the wager where working past the cliff is the winning move; there is only the version where you find out too late that it wasn't.
The forty-hour week was never a promise about how long you should work. It was a warning, priced into law, about where the returns run out, issued before anyone could prove it was right. The people who wrote it could only guess. You can check. An hour is not an hour, and the ones past the ceiling are the most expensive and least productive hours you will ever spend, purchased with a currency, output now and health later, that you will not miss until it is gone.
An hour is not a unit of output. Neither is a token, a run, or an uptime figure.
Pencavel's lesson is that the proxy you track (hours) and the thing you care about (output) come apart exactly where it matters most, and the average hides it. The same trap waits for anyone measuring an agent fleet by activity instead of by what each agent actually produced and decided. Chain of Consciousness records what an agent actually did, from what inputs, when, so you can measure verified output rather than the cheap proxy that only looks like progress. Measure the shells, not the hours.
See Hosted Chain of Consciousness · Read the Theory of Agent Trust
pip install chain-of-consciousness · npm install chain-of-consciousness
Or the whole trust stack at once: pip install agent-trust-stack / npm install agent-trust-stack