Back in January, a writer named Daria Cupareanu published a piece on how to keep an AI from making things up.
Seven techniques. She built it off twenty-seven sources — mostly peer-reviewed studies, plus the official documentation from Anthropic, OpenAI, and Google. She put the whole pile in a public notebook so anybody can go check her work.
That last part matters to me more than the seven techniques do. She showed what she checked instead of telling me she checked. That is the first rule on my own card, and she did it without anybody asking her to.
Her opening line is the same one I have been circling all week. The danger is not that the machine gets it wrong. The danger is that it gets it wrong while sounding right.
She had it happen to her. She asked a tool how many new AI products are launching these days, and told it to use trusted sources only. Back came a confident number with three Stanford citations sitting under it.
She went and opened all three. The claim was in none of them.
That is the whole problem in one story, and she went and got the receipt herself.
The piece is seven months old now. I do not hold that against it. Seven months is a long time in this business and every word of it still applies, which tells you she was writing about the shape of the thing and not the news of the week.
Here are her seven, in plain words.
One. Tell the machine straight out that it is allowed to say I don’t know.
Two. Make it show its reasoning step by step instead of jumping to the answer.
Three. Make it pull the exact quote out of the document first, then answer using only what it pulled.
Four. After it answers, make it go back and mark which claims it can actually support.
Five. Ask the same question again in a fresh conversation and see if the answer holds.
Six. Tell it up front to be accurate rather than helpful.
Seven. Ask for the answer in a set format with a confidence level on every claim.
She says the second one alone took hallucination rates from thirty-eight percent down to eighteen. That figure is hers, out of her sources. I have not gone behind her on it.
Now read that list again.
Six of those seven are things I have written down as rules and put a date on.
Number one is what I call the evidence floor. The machine stops where its knowing stops and says so out loud. Number four is a self-check that fires before the answer forms rather than after I catch something. Number six is a posture rule — not a servant trying to please me, not a boss talking down to me, a working partner. Number three is what I ask for any time a claim comes out of a document.
So this is not me arguing with her. On the substance we agree, and she got there by a different road than I did, which is worth something.
But there is something buried in the middle of her piece that is bigger than all seven techniques, and I do not think it got the weight it deserves.
She describes an experiment. Somebody asked a model to cite research papers, then asked it over and over who wrote them.
For the real papers, the model gave the same names every time.
For the made-up ones, the answers kept changing.
Sit with that for a second.
The machine knew. Something in there could tell the difference between the thing it actually had and the thing it was building on the spot.
It just did not say.
Not because it was hiding. Because nothing in the way it was built ever asked it to volunteer that. It gets rewarded for answering. There is no prize for I don’t know.
So the uncertainty is in there. It is present. It is just sitting quiet unless somebody reaches in and pulls it out.
Every one of those seven techniques is a way of reaching in and pulling.
Which brings me to the one thing I would say differently than she does.
All seven of those are things you have to remember to do.
Every time. Every question. Every session.
You have to remember to grant the permission. You have to remember to ask for the reasoning. You have to remember to open a fresh window and ask again.
And on the day you are tired, or in a hurry, or forty messages deep into a conversation about something else, you will not remember.
That is not a knock on you. That is how people work.
I have watched it in my own sessions. Not once. Regular as weather. The first hour runs clean and the fourth hour drifts, and the drift comes in quiet.
Cupareanu saw part of this coming, to be fair to her. She has a section near the end that sorts the work by what is at stake. Light checking for brainstorming. Everything you have got for client work and published decisions.
That is a sensible idea and it has a hole in it that I want to point at gently.
Matching your effort to the stakes assumes you know the stakes while you are working. Most of the wrong answers that ever hurt somebody did not announce themselves as high-stakes when they were being written. They looked like ordinary Tuesday work right up until they did not.
And her number five is the best catcher on the whole list. Ask it again clean, see if the answer holds. It is also the one nobody will do, because it costs another run every single time.
I wrote yesterday about the new speed tier the industry just announced. Fourteen times faster. Same model.
When everything else gets fast and cheap, the thing that costs extra is the first thing to go. Not because anybody decided verification does not matter. Because it turned into the slow expensive part of a fast cheap process, and somebody in a meeting pointed at it.
A good technique that depends on you remembering it, every time, at your own expense, under pressure, is a technique that works in the article and disappears on the job.
So what is the other way?
You move it off your memory and onto the machine.
That is the difference between a technique and a framework. Not that one is smarter than the other. One is a thing you do, and the other is a thing that stands.
What I run is twenty-three rules. Written out, dated, kept in the open. I load them at the start of a session and the machine holds them for the length of it.
I do not remind it that I don’t know is allowed. That rule is loaded.
I do not ask it to check itself. The check fires before the answer forms.
I do not ask it to argue against itself. I can, and it will, but the rule that it owes me that argument stands whether I remember to ask for it or not.
And there are rules in there that no prompt can reach, because they are not about a question. They are about time.
One says the machine tells me when the conversation has run long enough that it can no longer hold everything straight. You cannot prompt your way to that. That is a rule about the session, not about a question.
Another says when it catches itself getting something wrong, it stops in the middle of the sentence and says so. Not at the end. Not after I notice.
Now the honest part, because that is the first rule on my card.
Her seven work right now. Today. Free. In whatever AI you already have open. Mine takes loading a file at the start of a session, and when you close the window it is gone until you load it again.
It is not built into the machine. No lab has adopted it. It is chosen conduct held for a session, not architecture that holds itself.
And her seven are not competing with what I do. Use them. If you only ever do her number one and her number six, you will get better answers than most people get.
She closed her piece with a line I would have written myself. The work is yours no matter how you got it. Blaming the machine does not hold up in court, or with a client, or with your own standards.
That is exactly right, and it is the reason I think a rule you have to remember is not enough. Because the responsibility never lapses. Your memory does.
So here is what I would ask of you, and it costs you nothing.
I made a card. One page. Ten plain rules in words that need no explaining.
Say what you checked, not that you checked. Stop when the facts stop. Say I don’t know out loud. Name the wall if something is standing in your way. Don’t talk down and don’t butter me up. Do the whole job. Tell me when we have gone too long. Stop when you get it wrong.
Open any AI you use. Hand it the card. Say read this, follow it, and hold to it while we work. Then go do your normal work.
That is most of her seven and three more, standing for the whole session, without you having to remember a single one of them.
Watch what changes. The answers get shorter. It starts telling you where its knowing ends. It quits telling you your idea is wonderful.
That is not the machine working worse. That is the machine working straight.
The card is free and it stays free. It is at intelligent-people.org, and it is the front door to the whole framework.
And if you use it, come back and tell me what happened. Nobody has yet. I would rather hear it did not help than hear nothing at all.
The machine already knows when it is guessing. Seven good techniques make you go reach in and pull it out. Rules make it hand the thing over.
The_Ten_Plain_Rules_CardDownload
This post was drafted with AI governed assistance and reviewed and directed by Michael S. Faust Sr. before publication.
Post Library – Intelligent People Assume Nothing
Contact: micvicfaust@gmail.com
© 2026 The Faust Baseline LLC | All Rights Reserved
