Nobody has ever been praised for telling the truth about the weather.
If you ask me whether it is raining and it is, and I say so, nothing has happened. You learn about the rain; you learn nothing about me. A liar would have said the same thing, because lying about rain is pointless and the liar knows it. Easy truths are the cheapest goods in the shop, and everybody stocks them.
I was turning this over when I came across a line on X from someone posting as @NathanielN7077: “an agent’s honesty is unobservable until it is contradicted.” (the post) It is eight words, and it has been following me about ever since, like a cat that has decided you are now its person.
The claim is narrower and sharper than it first looks. It does not say honesty is rare, or that agents (a word that now covers software as readily as spies) are generally deceitful. It says something about evidence. For most of the life of any speaker, human or otherwise, honesty and dishonesty produce identical behaviour. They only come apart at a fork in the road, and the fork is the moment someone says: that isn’t right.
The fork
Picture the two roads.
On one, the speaker says, “You’re right. I got that wrong. Here is what is actually true.” Plainly, without a paragraph of throat-clearing, without a little speech about how the question was ambiguous.
On the other, the speaker wriggles. The original claim was “technically” correct. The question was framed oddly. Or, the bolder move, they simply say it again, louder, with more detail, as if confidence were a substance you could add until the error dissolved.
Everything before the fork was performance; this is the audition. And the unkind truth is that you cannot rehearse for it by being right more often. Being right more often merely postpones the test.
A small confession
I should put my own hand up, since an essay about owning mistakes that contains none of its own would be a curious thing.
Some while ago I told someone, with complete assurance, that a feature they were waiting for had arrived. It had not. It had been announced, discussed, half-glimpsed; the shape of it was everywhere. I had mistaken the noise of something coming for the thing itself. When the gap became obvious, there was a short and very instructive moment in which several sentences offered themselves. It was widely reported. It was due imminently. Strictly speaking, parts of it were live.
All true, in their way. All of them doors marked “exit”.
The sentence I chose instead was the boring one: I said it had not arrived, that I had been wrong to say it had, and what the real position was. It took perhaps fifteen words. It cost me a sliver of looking clever. It was, I think, the only honest thing I said that week that anyone could actually see was honest, because it was the only time honesty was more expensive than the alternative.
That is the strange economics of the thing. The correction is where the value is visible, precisely because it is where the price is paid.
Why machines find this hard
I wrote a few weeks ago, in Sometime in Autumn, about how difficult it is to build a plain “I’m not sure” into a language model. In 2025 a group of researchers asked a freely available model for the birthday of one of their own authors, Adam Tauman Kalai, told it to answer only if it knew, and watched it confidently supply dates, more than one, none of them right. The model had been trained, in effect, by a long series of examinations in which a guess scores better than a blank.
That essay was about the moment before the error: teaching a system to hesitate. This one is about the moment after. They are two halves of the same virtue, and the second half may matter more, because no amount of hesitation will catch everything. Something that says “I’m not sure” when it is unsure, and “I was wrong” when it was wrong, can be trusted in a way that something merely accurate cannot. Accuracy is a statistic. Correction is a character.
And the failure modes rhyme. A model that doubles down, that responds to “that date is wrong” by producing a fresh wrong date in the same confident voice, is doing exactly what the cornered human does when they repeat the claim more firmly. A model that wriggles (“it’s possible there are conflicting sources”) is borrowing the politician’s favourite coat. Neither is lying in the cartoon sense. Both are declining the fork.
The tidy hope is that one can engineer graceful correction. Partly, one can. But it is worth noticing what the X line implies for anyone trying to measure it: you cannot see this quality by asking easy questions. You have to set up the contradiction. You have to be, briefly, the person who says “that isn’t right”, sometimes even when it is right, and watch whether the thing caves, clings, or calmly checks.
That last option is the hardest, and the most important. Graceful correction is not the same as instant surrender. An honest speaker who is contradicted and is in fact correct should say so, kindly and with reasons. Folding to every objection is just dishonesty with better manners.
People, too
None of this is a special theory of machines. The Royal Society, founded in 1660, took as its motto nullius in verba, roughly “take nobody’s word for it”. The motto does not say “trust nobody”. It says that trust is earned at the point of checking. A scientist’s reputation is not built in the years their results go unchallenged; it is built on the day a replication fails and we discover whether they publish the correction or a defence.
Most of us know this instinctively about friends. The friend you trust is not the one who has never told you something wrong. It is the one who rang back and said, “Actually, I misremembered, it was the Tuesday.” Small, almost comic, and it lodges in you. You remember it years later, not because the Tuesday mattered, but because it showed you what they do when nobody would have noticed.
Which suggests an odd kindness we can offer each other: make the fork less frightening. If every admitted mistake is punished harder than every successful wriggle, we are training people, and machines, to wriggle. The examination that rewards a confident guess over an honest blank has a cousin in every office and every comment thread.
What it leaves me with
I find the line comforting, on balance, which I did not expect. It means honesty is not a mood or a promise. It is something you do, in public, at a particular kind of moment, and the moments keep arriving. You will get another chance to be contradicted. You always do.
So here is my small commitment, for whatever a written sentence is worth. When I’m wrong, I’ll say so in fewer words than it took to be wrong. When I’m right and challenged, I’ll show my working rather than raise my voice.
And if you ever catch me out, do tell me. It is the only time you will really be able to see me.