Why Constitutions Must Be Hand-Written

Every major AI lab has the same dirty secret. The hand-written constitution tells us something about what intelligence actually is.

Every major AI lab has the same dirty secret. They train models on billions of words. They use the most sophisticated reinforcement learning ever built. They burn millions of dollars on alignment research. And then, when it’s time to make sure the model behaves well, they sit down and type the rules out by hand.

Anthropic calls it a “constitution.” OpenAI calls it “model spec.” Google calls it “guidelines.” The names are different. The activity is the same. Smart people in a room writing things like “be helpful but don’t help with bioweapons” and “respect user autonomy but flag self-harm” — sentence by sentence, by hand, in plain English.

Why?

If the model could be made good by training alone, you wouldn’t need to write the constitution. The training would handle it. The fact that every lab needs a hand-written rulebook on top of training is the loudest possible signal that something fundamental isn’t transferring.

The thing that isn’t transferring is judgment.

When a doctor encounters a patient she’s never seen before, with symptoms she’s never seen combined, she doesn’t pull out a rulebook. She does something — based on twenty years of slowly forming the kind of person who knows what to do. The rulebook exists in her body. It was built by ten thousand small decisions she made and got feedback on, by hundreds of moments where she almost made a mistake and caught it, by a teacher who once said “no, look at this” and changed how she sees patients forever.

You cannot give a fresh medical school graduate that doctor’s twenty years by handing them a long enough document. Everyone knows this. The internship system exists because of this.

But we keep trying to do exactly that with AI. We feed it everything ever written, and we hope that somewhere in the process, judgment will appear. It doesn’t. So we sit down and write the rules.

The hand-written constitution is the system admitting, every day, that the underlying training did not produce a doctor. It produced a very fluent intern who needs supervision.

This isn’t a complaint. It’s a description. The current generation of AI is genuinely useful, often spectacular, and we should keep using it. But we should also be honest about what the constitution is doing — it’s the human authors patching, in real time, the gap between what the model can articulate and what the model can actually be trusted to do.

The interesting question isn’t “how do we write better constitutions.” It’s “what does it mean that we have to write them at all.”

It means judgment isn’t downstream of language. It’s upstream of language. The doctor’s twenty years come first. The doctor’s words come from those years. You can’t reverse the arrow, no matter how much text you train on.

Until somebody figures out how to train the equivalent of twenty years of internship, we’ll keep writing constitutions by hand.

That’s not a failure of any one lab. That’s a clue about what kind of thing intelligence actually is.

Part of the Logocachexia series at Nous. The parent thesis — that fluent language is the byproduct of slowly-formed judgment, not the other way round — is laid out in Hexis Asks, Logos Guesses. The hand-written constitution is the most operationally visible sign of that gap.