Pudels Kern · Part 2 — first published on LinkedIn on 21 July 2026
My AI delivered a flawless dossier — about nobody
The most convincing answer I have ever had from a machine was half wrong. The uncomfortable part isn't that it was wrong. It's that I couldn't see it.
The assignment was routine
I wanted to know something about a business contact. Not a hard task, nothing you set aside time for — the kind of research you hand over in passing while you get on with something else.
Back came an answer with everything you could ask for. A career, cleanly broken into stations. A year against each one. A source against every claim. The tone level, without overstatement, without the small reassurances by which you recognise someone who isn't sure.
I skimmed it once and found it good. Exactly the way you treat a piece of work that looks orderly.
I almost acted on it.
One flaw
There was exactly one. The answer had fused two people of the same name into a single, convincing biography.
Half that career belonged to a stranger.
Let me spell out what that means in practice, because otherwise it sounds like a blemish. You walk into a meeting. You have prepared, you have a few points of connection in mind, you mention an earlier station because it fits the matter at hand. And across the table sits someone who was never there. At best there is a brief hesitation you talk past. At worse, the impression forms that you have prepared for a person you confused with somebody else — and that impression cannot be recovered inside the same conversation.
That is not a small inaccuracy you straighten out. That is a meeting already lost before it starts.
The uncomfortable part: it read like truth
The point isn't that an AI can be wrong. Anyone who uses one knows that.
The point is what being wrong looks like. The false half arrived in the same confident tone as the true one. The same clean way of citing sources. The same order, the same completeness, the same pleasant sense while reading that somebody here had been thorough.
With a person I would have picked something up. People go vaguer when they are unsure. They say "I think", they slip in an "as far as I know", they slow down. I have been reading those signals all my life without thinking about it.
A machine doesn't send them. Its tone is not a report on its own confidence — it is a property of how the thing is built. It phrases the invented with the same care as the evidenced, because while phrasing, it isn't distinguishing between them at all.
A wrong answer carries no warning label. It doesn't look weaker than a right one, it doesn't sound more careful, it doesn't come with a note. Wait for the error to strike you and you are waiting for a signal that doesn't exist.
Which disposes of the most obvious countermeasure before you try it: reading harder doesn't help. You cannot see what makes no difference. Resolve to read more attentively in future and you have resolved to spot an invisible property — and you will consider yourself attentive right up until the next time it goes wrong.
What caught it
Not sharper reading. A standing rule caught it:
No source, no claim.
It is as short as it sounds. If a statement has a source, it goes out as fact. If it has none, it says "unknown — check". Not the most likely completion, not the obvious guess, but the word that describes the state.
At its core this is a working instruction rather than an attitude: facts are retrieved from a source, never generated. That is a technical statement about the route by which a claim comes into being — and therefore testable, which "please work carefully" is not.
That distinction is exactly where the dossier broke. One half had evidence behind it. The other didn't. As long as only the result counted, the fused biography held together effortlessly; it was coherent, after all. The moment a source was demanded for each individual station, it fell into two pieces.
The same rule now applies to every figure about another company. Revenue, headcount, locations — none of it comes from memory, all of it is researched and named with its source. An invented number looks exactly like a retrieved one. That is the whole reason.
Four guards, one principle
Three more stand around that rule. Four in all, each taking hold at a different point in the chain.
Source or silence. The standing rule above. It acts as a statement comes into being, which is the earliest point available.
A designated dissenter. A second AI, from a different provider, whose brief is expressly to disagree. Not to confirm, not to weigh up, not to be polite — to disagree. Ask a single instance and you get an opinion and mistake it for a result.
I set this dissenter up on 12 July and gave it a version of my own positioning that I assumed was clean. It found three weaknesses I had missed, and to prove the point it quoted a sentence from its knowledge base word for word — which established at the same time that it really was working from my material rather than talking in generalities. That was the day a good idea became a tool.
A machine gate. A check that runs over every text going out. It doesn't test whether the text is good — it can't, and I wouldn't want it to try. It tests whether the text breaks the rules I have committed to. Phrasings I have abolished. Claims that depart from the binding source. Figures without provenance.
A tripwire. More on that in a moment — it's my favourite of the four.
The principle behind all four is the same: none of them depends on my being alert at the right moment. Attention is a resource that runs short precisely when there is a lot to do — which is to say, precisely when mistakes are expensive.
My favourite: the tripwire
One error is built into the material. One I know about, planted on purpose: a spot the check must report every single time.
So the target state of my check is not "zero findings". It is "exactly one".
This confuses everyone I explain it to for the first time, and it is the most useful idea in the whole arrangement. The value isn't in the error being found — I know it's there. The value lies in the day it isn't reported.
On that day the error hasn't gone away. On that day the check has broken. And I find out, instead of collecting green ticks for weeks that no longer mean anything.
It cost me some effort to leave that one finding standing. The temptation to exempt it and finally see a clean zero was considerable — a zero looks like order. But it is only order while the check is still running, and a zero is exactly the thing that cannot show you that.
That is the difference between a control and the feeling of having controlled. Any check that can fail silently will eventually fail silently. A tripwire turns silent failure into loud failure.
Trust the procedure, not the tone
Confidence is a style. Verification is a structure.
The tone of an answer tells you something about how the tool was built, and nothing about whether the answer is right. Follow it anyway and you have carried a habit from dealing with people over to an object it doesn't apply to. That isn't carelessness — it is a very good habit in the wrong place.
The remedy is nothing new. It is an old engineering habit with a new subject: trust the procedure that produced a result, and let that set the value of the result. In manufacturing this has been unremarkable for decades. Nobody there judges a part by how well it looks.
I use these tools more now than before, not less. The difference is that every claim with something riding on it has a route behind it that I can describe.
Checked beats convincing.
Pudels Kern is a series about the thing behind the first impression. Michael Kraewing leads digital initiatives as an interim executive.