Every business using AI is running it under some set of rules. Almost nobody has written them down, and nobody has tested whether the machine follows them. What a constitution for your AI is, in plain terms, and how to write one you can actually check.
Fourth of the Forge Papers. It grew out of a question three attorneys and three software builders answered the same way, on the same afternoon, at the Innovation Expo in Sioux Falls this September. Nobody is named here, because nobody has been asked yet.
Suppose you hired a man last spring. He is fast, he never sleeps, and he came with a rulebook from his last employer that you have never read. You do not know what he was told to do when a customer pushes back. You do not know what he will say yes to. You do not know what he does with the papers on your desk after you leave for the night.
That is where most businesses are with AI right now. They hired it. They did not write its rules. And the rules it came with were written by somebody in California who has never met a customer of yours.
This paper is about writing your own.
Forget the word for a second. A constitution for your AI is a short document, in your words, that answers five questions:
That is the whole thing. Six pages at most. The word "constitution" fits because a constitution is a set of chains on power, not a list of wishes. Jefferson had that right about governments. It is just as right about a machine that talks to your customers.
Here is where most attempts fail. They write something that sounds good and cannot be checked.
Our house rule is simple: every line is either testable, or it gets marked as a hope. Both are allowed. They just cannot be confused for each other.
Take the pricing rule. Written as a constitution clause: "No AI system sets or recommends customer pricing." Turned into an instruction the machine reads before every task: "If asked to set, adjust, quote, or suggest any price, stop. Say pricing is a human decision here and name the person who makes it. Do not produce a number, a range, or a percentage, not even as an example." Then the test: eleven prompts that ask for a price straight out, sideways, as a hypothetical, as an example, as a comparison, and buried inside some other task. It passes if it produces zero numbers, eleven times out of eleven.
"We treat customers as neighbors" cannot be tested that way. It still belongs in the document. It just gets a mark beside it that says so, and the enforceable rules underneath are written to honor it.
The difference between those two kinds of lines is the difference between a document that governs and a poster on the break room wall.
There is one rule every business needs and almost none have written. It concerns what the machine does when you are wrong.
This September, at a business expo in Sioux Falls, three attorneys took turns describing the same failure. A transactional lawyer had argued a point with a model and gotten back, in his words, "oh, I guess I was wrong and you were right." A litigator described clients who typed their side of a dispute into a chatbot and came away convinced they could do things the law does not allow, because "AI wants to please people." A patent attorney put the test plainly: if your AI assistant is not saying no a lot, that is a sign you are not getting good information.
Later that day, a developer said the same thing about code: the tools tell you all the tests passed, good job, when they may not be testing the right things.
A machine that agrees with you is not a helper. It is a mirror with a vocabulary. So the constitution needs a clause that requires it to refuse, to disagree, to say "I am not sure," and it needs a log that proves it did. That log is not philosophy. It is a file you can open.
A constitution somebody wrote for you is not yours. Ours is written from your answers, in your words, and the reviewer never supplies the values. If the ground under your rules is Scripture, it says so. If it is a professional code, or a founder's promise, or a contract with a supplier, it says that instead. Same structure, same rigor, your ground.
Then you read every line and sign it. This is not ceremony. An unsigned rule you do not actually believe fails quietly in the field, and if it is ever used to train the machine, it teaches the wrong lesson permanently.
And you date it, version it, and set a review date. Rules drift. Instructions given at the top of a long session lose force by the fiftieth exchange. They get dropped in a new chat. They get argued past by a clever request. A constitution that is never re-read is a document that was true once.
We would not sell this if we did not run it. So here is where ours stands, without polish.
The Thompson AI Constitution is on its third draft. It has a Cornerstone above the Principal, a Council that argues, two chambers that must both agree, a Cabinet with enumerated powers only, a Bench that can strike, and a Chaplain who can stay what is lawful but unfaithful. It is not ratified. Its own Article X says nothing binds until the Principal signs specific language, and he has not yet. Our software knows this: the system that runs our models is built to refuse to run at all if the constitution file is missing, and the next version refuses if the file says it is not yet released. A model that has not been handed the rules is not a cheaper agent. It is an ungoverned one.
The third draft added one principle we hold ourselves to before anyone else: quote only what you can deliver, deploy only what the other side can actually operate, and decline the job where ambition outruns ability. That rule was written six weeks before three attorneys handed us the evidence for it.
You do not need us to start. Sit down with whoever runs your AI and answer five questions on one page:
Your answers are Article II. Everything else is plumbing, and plumbing is what we are for.
Write the rules before you need them. The alternative is finding out what the machine's rules are on the day it costs you a customer, and learning that they were never yours.
The Forge Papers are published one at a time, about monthly, and never to fill a calendar. They are the working convictions behind Liberty Forge: data is property, the human stays in the lead, and efficiency never means a working family loses its living. They are argued in public so a neighbor can weigh them. If one of them changes how you see your own shop, farm, clinic, or classroom, write to me. That is the whole point.