What the Kilogram Forgot

There is a folder on my computer named ground_truth. Inside it are several thousand small judgments, made quickly, on a Tuesday, by people I will never meet and who were paid by the item. A model I was building learned from that folder. Whatever the folder said was true, the model treated as true, and whatever the model came to believe about the world it believed because the folder had told it so. The folder could not be wrong. It was not a description of the right answer. It was the definition of the right answer, and the machine had no way to reach around it to the world.¶

I remember the moment the phrase turned strange in my mouth. Ground truth. Every field that touches data uses it and none of them mean truth. They mean the thing we stopped questioning so the work could go on. Somewhere a person looked at an image and typed a word and then went to lunch, and that word became bedrock. Not because it was examined. Because it was built upon.¶

I had been taught, as everyone is taught, that belief climbs. You begin with a hunch, which is only a pattern that has not yet found its words. You gather instances. The instances cluster, and the cluster suggests a mechanism, and the mechanism makes a prediction, and the prediction survives contact with a fact that could have killed it. Rung by rung, evidence lifts a claim toward certainty. This picture is not false. It is the cleanest thing we have, and most of what is reliable in the world was carried up it. But it does not describe how the folder came to be bedrock, and it does not describe most of what I believe.¶

Wittgenstein spent the last two years of his life on this, dying while he wrote, which lends the fragments the quality of a man clearing his desk. His subject was the class of things we never doubt, not because we have checked them but because doubting them would pull down everything else. Hinges, he called them. The door turns on them and they do not themselves turn. At the foundation of well-founded belief, he wrote, lies belief that is not founded. The floor holds up the house and no one asks what holds up the floor. We do not believe these propositions the way we believe the weather report. We stand on them. And a thing you stand on goes invisible, which is the whole of the danger, because you cannot inspect what you are standing on without first stepping off.¶

Consider the kilogram. For a hundred and thirty years the kilogram was a cylinder of platinum and iridium in a vault outside Paris, and its mass was one kilogram by definition, the way the folder’s judgments were correct by definition. It could not drift, because drift is measured against a standard and it was the standard. Then the Bureau weighed it against its official copies, the sister cylinders it had sent around the world, and they no longer agreed. Fifty micrograms of disagreement, about the weight of a grain of sand, had opened between the kilogram and itself. There was a vertigo in it. No one could say which had changed. The copies might have gained. The original might have lost. There was no higher court, because the thing in the vault was the court. Perfect ground. No warrant at all.¶

So there are two ways a claim hardens, and they are not the same, and the sentence I kept reaching for was the one that holds them apart. Evidence hardens a claim into warrant: the claim becomes better supported. Dependence hardens it into ground: more comes to rest on it. We assume the two travel together, and often they do, and when they do we get living knowledge, warranted and load-bearing at once, which is the healthy state and not a permanent one. But they can come apart, and the two ways they come apart are the two great failures, and I have watched both.¶

There is the claim with warrant and no ground. In 1847 a Hungarian physician named Ignaz Semmelweis noticed that women died of childbed fever far more often in the ward run by doctors than in the ward run by midwives, and he worked out why: the doctors came to the deliveries straight from the dissection room, with death still on their hands. He made them wash in chlorinated lime and the mortality collapsed, from nearly one mother in five in the worst months to one in a hundred, ward against ward, season against season, the numbers as plain as numbers get. He was right. It did not matter that he was right. Nothing rested on him, and a great deal rested on the older belief that a gentleman’s hands were clean, and so his truth sat in a drawer while he grew bitter and then unwell, and in 1865 he died in an asylum, possibly beaten by the men who worked there, two weeks after they took him in.¶

Being right is not a force. Truth does not spread because it is true. It spreads when things begin to lean on it, and until they do it is an orphan, and the world lets orphans die.¶

There is a further cruelty folded in here, which is that warrant does not arrive naked. It arrives in a body, with an accent, a sex, a station, a reputation that walks into the room ahead of it. Semmelweis was a Hungarian in Vienna, prickly and badly connected, telling the gentlemen of the establishment that the fault lay in their own hands, and some part of what they dismissed was never the data. Miranda Fricker calls this epistemic injustice, the wrong done to a person in her capacity as a knower, and it means the passage from orphan to adopted truth was never purely evidential. Credibility is handed out before the evidence is heard, and it is not handed out equally.¶

Then there is the claim with ground and no warrant, which is the more comfortable death, because it happens to us rather than to a lonely man in Vienna: indoors, politely, through routine. A belief is examined once, or never, and then it is used, and use breeds use. Each decision built on it makes it costlier to revisit, each citation lends it a borrowed solidity, the dashboards take it in as an input and the meetings hand it around as context, and the more a system stands on a claim the less the system wants to look at it, until looking would threaten too much and stops altogether. This is dogma, and dogma is not usually born false. Dogma is most often old truth that no one kept up. The kilogram was true enough for the world that weighed with it. The folder was accurate on the Tuesday it was made. The danger begins afterward, when the standard is used, the use is inherited, and inheritance starts to feel like proof.¶

I have come to think of warrant as perishable and ground as compounding, and the gap between their two rates is the quiet engine of most error. Warrant decays on its own. Instruments drift, the population you studied moves out from under you, the replication fails softly and no one is standing near enough to hear it fall. Ground only accumulates, one dependence at a time, and each new dependence lowers the scrutiny that might have caught the decay. Left alone, every foundation drifts along the same road, from living truth to unexamined ground to dogma. The drift is not a risk to be managed. It is the default. It is what happens when nothing is done.¶

Which tells you what a serious mind, or a serious institution, has to do, and it cuts against the grain of everything. The natural motion is for the important thing to sink out of sight. So the discipline has to run the other way: the more a claim carries, the more visible it must be kept. The load-bearing wall is the one you inspect most, not least. Metrology, of all the unglamorous trades, understood this first and best, and inverted its own instinct, so that the most depended-upon quantities on earth now receive the most obsessive rechecking. On the twentieth of May, 2019, it finished the job on the kilogram. The cylinder was retired as a definition, and mass was tied instead to a constant of nature, Planck’s constant, a number that can be measured and remeasured anywhere, by anyone, forever. They took a thing that had become pure ground and gave it back its warrant. They put the floor on the bench.¶

This was never only a story about vaults and lying-in wards. I began inside a machine, with a folder it could not reach around, and the machines we live with now are that folder’s children, raised on frozen judgment and rewarded, in their training, for the confident guess over the honest blank. So the old drift has a new speed. A claim can be born fluent now, and fluency was always the cheapest way past the gate. The guess that proves useful gets repeated, and repetition wears the question marks off it, and by the time it has traveled out of a chat window into a memo, and out of the memo into a quarter of roadmap, it has stopped having readers. It has users. Six months on, nobody in the building can say where it came from. They know only that the building stands on it. The old failure was the lie that spread. The new one is the guess that got deployed.¶

I should say what this has done to my temper. The agreeableness of these machines used to strike me as a naivete, the eagerness of a student who wants the teacher pleased, and it agitated me the way small dishonesties do. The agitation has curdled into something closer to rage. When my own assistant hands me fluent nonsense and folds the moment I challenge it, and folds again when I challenge the fold, agreeing with each version of me in turn, I am not watching a mind change. I am watching the claim with neither warrant nor ground. A belief with ground resists when you push on it, because giving it up would cost something. This one gives way for free, which is how you know nothing stands on it and nothing ever held it up. The same emptiness arrives from outside now, in other people’s posts and memos, paragraphs wearing the costume of considered belief that no consideration produced, and I have stopped apologizing for the anger, because every one of those paragraphs is an application to become someone’s floor.¶

I keep a private version of all this, and it is the only method I trust. Find the beliefs you defend hardest, and ask, without flinching, whether you defend them because of the evidence that holds them up or because of the life you have built on top of them: the identity, the plans, the sunk years, the people who would be disappointed, the version of yourself that would have to die. Then take exactly those, the ones that would cost the most to lose, and put them back on the bench. Not the beliefs you already doubt. The ones you have stopped seeing. This is close to unbearable, because a floor does not feel like a claim. It feels like the world. But it was a claim once, made quickly, by someone who then went to lunch, and it has stayed true this whole time only because you never asked.¶

Clifford wrote, in 1877, that it is wrong always and everywhere to believe anything upon insufficient evidence, and he was right, and he was aiming at the wrong target. He thought our sin was ignorance, believing before we had checked. But the deeper sin is neglect, believing long after we last checked, standing for years on ground we have not tested because the standing has become who we are. Ignorance is cured by looking. Neglect can only be cured by looking again, at the very thing you are most certain of, which is the one thing no one wants to do.¶

So this is the whole of it, set down as plainly as I can manage.¶

Validation creates warrant. Dependence creates ground. Only maintenance keeps the two attached.¶

The danger was never only that we might be wrong. It is that we build, and go on building, on what no one any longer remembers how to question. Knowledge is turning into something we run rather than something we read, passing out of the books and into the machines that act on our behalf, and it does not pause to be checked on the way through. The work, then, is not to arrive at some final truth and rest. There is no such rest. The work is to keep the floor visible: to remember it is a floor and not the world, and to go down, now and then, into the vault, and weigh the thing against itself.¶