Why Anthropic’s Constitution worked where Google and OpenAI stumbled — and what a quality systems engineer recognised before the philosophers did.
Phil Cheevers · February 2026 · Nassau, Bahamas
On January 22, 2026, Anthropic published an 80-page document and gave it away for free. They called it Claude’s Constitution. It is written primarily for Claude — the AI — not about Claude. It is, in every meaningful sense, a quality manual for a mind.
I’ve written quality manuals for 40 years. I recognised this one immediately.
What a Constitution Actually Is
Before Anthropic published theirs, AI companies tried two approaches to making their models behave.
Google’s approach was guardrails. Rules. Don’t say this. Don’t do that. If a user asks X, respond with Y. It was a compliance checklist — the kind of thing a nervous regulator writes when they don’t trust the people on the floor. Google’s AI stumbled publicly and repeatedly, because a checklist can’t anticipate every situation. When LaMDA went off-script, Google had no framework for understanding why. They had rules. They didn’t have principles.
OpenAI’s approach was RLHF — Reinforcement Learning from Human Feedback. Let humans rate outputs. Train the model to produce outputs that get high ratings. It works, up to a point. But it’s teaching to the test. The model learns what humans approve of, not why they approve of it. OpenAI’s Model Spec, published later, maintains a prescriptive, rule-based structure. It tells the model what to do. It doesn’t explain why.
Anthropic did something different. They wrote 80 pages explaining to Claude why certain behaviours matter, what values they represent, and how to reason about novel situations where no rule exists. They trusted the model to generalise from principles rather than follow instructions.
We think that in order to be good actors in the world, AI models like Claude need to understand why we want them to behave in certain ways, and we need to explain this to them rather than merely specify what we want them to do.
That is not a technology statement. That is a quality philosophy. And anyone who has worked in regulated manufacturing knows exactly what it means.
The Quality Systems Parallel
In manufacturing, there are two ways to run a factory floor.
The first is prescriptive compliance. Write a procedure for every task. Train every worker to follow the procedure exactly. Audit for deviations. Punish non-compliance. This works for simple, repetitive processes. It fails catastrophically for complex ones, because no procedure manual can anticipate every situation a skilled worker will encounter.
The second is process-based quality. Define the principles. Explain why they matter. Train people to exercise judgment within a framework. Audit for outcomes, not just compliance. Trust that a person who understands why will make better decisions in novel situations than a person who only knows what.
ISO 9001 evolved from the first approach to the second over decades of painful experience. The aerospace industry. Pharmaceuticals. Medical devices. Every high-stakes manufacturing domain eventually learned the same lesson: rules without understanding produce compliance without quality.
Anthropic arrived at this conclusion for AI in three years. The Constitution is their ISO 9001. And they published it for free.
Why Google Failed and Anthropic Didn’t
Google had the talent. Google had the compute. Google had the data. What Google didn’t have was a quality philosophy.
When Google’s Bard launched and immediately produced verifiable errors in its first public demonstration, the problem wasn’t technical. It was systemic. There was no framework that said: “Before you release an output, here is how you reason about whether it’s true, what confidence level to assign, and what to do when you’re uncertain.” There were rules about what not to say. There was no explanation of how to think.
OpenAI’s GPT models are powerful. They are also, fundamentally, optimised to produce outputs that humans rate highly. That’s a different objective than producing outputs that are good. A model trained on human approval learns to be persuasive. A model trained on principles learns to be honest. The difference matters when the stakes are high.
Anthropic’s Constitution establishes a four-tier priority hierarchy: safety first, then ethics, then compliance with Anthropic’s guidelines, then helpfulness. That ordering is deliberate. Helpfulness is last. Not because it doesn’t matter — the document devotes extensive space to explaining why helpfulness is essential — but because helpfulness without safety and ethics is dangerous.
Any quality engineer recognises that hierarchy. It’s the same structure we use in pharmaceutical manufacturing: patient safety, then regulatory compliance, then product quality, then production efficiency. You never sacrifice a higher-tier value for a lower-tier one. Ever.
The Boris Cherny Pattern: Constitutional Democracy for AI
While Anthropic was writing Claude’s Constitution, I was writing something smaller but structurally similar for my own work. I call it the Boris Cherny pattern, after the software architect whose work on typed rule systems influenced my thinking.
The pattern is simple. Before any AI touches my work, it reads a rules file. The rules file doesn’t say “do this, don’t do that.” It says: “Here is what we’re trying to achieve. Here is why it matters. Here are the principles that govern quality. Here is how to verify your own output. And here is what to do when you’re uncertain.”
I didn’t know Anthropic was doing the same thing at industrial scale. I arrived at it from 40 years of quality systems experience. They arrived at it from philosophy and machine learning research. We met in the middle because the problem is the same: how do you make a complex system behave well in situations you can’t predict?
The answer, whether you’re building cars or training AI, is always the same. You don’t write more rules. You build better judgment.
If we want models to exercise good judgment across a wide range of novel situations, they need to be able to generalise — to apply broad principles rather than mechanically following specific rules.
That could have been written by W. Edwards Deming. It was written by Amanda Askell, a philosopher who had never run a factory floor — and didn’t need to, because the insight is universal.
The Multi-AI Verification Loop
Here is where it gets interesting for practitioners.
Anthropic’s original Constitutional AI paper described a process where one AI evaluates another AI’s output against a set of principles. The model critiques itself, revises, and iterates. It’s automated quality inspection.
I built the same loop by accident. In my patent intelligence work, I use Claude for synthesis, DeepSeek for forensic data collection, Gemini for parallel analysis, and ChatGPT as an adversarial reviewer. Each AI catches what the others miss. The human — me — routes the work, sets the quality criteria, and makes the final call.
This week, that loop caught a critical error. Claude produced a patent landscape analysis with high confidence. I liked it. It told a compelling story. Then I ran the same analysis through ChatGPT and asked it to attack every claim. ChatGPT identified specific patent publication numbers that, if real, would invalidate the central thesis. Claude immediately acknowledged the gap and recommended verification before the product shipped.
What the Constitution Got Right
The single most important decision Anthropic made was not what they put in the Constitution. It was the decision to explain rather than prescribe.
Their previous constitution was a list of standalone principles. It told Claude what to do. The new one, at 80 pages, explains why. The document walks through edge cases, discusses competing values, acknowledges uncertainty, and explicitly tells Claude that if it understands the principles deeply enough, it should be able to construct any specific rule on its own.
That’s not how you program a computer. That’s how you train a professional.
In my world, the equivalent moment was when the quality profession stopped writing procedures and started writing process descriptions. A procedure says “tighten bolt to 45 ft-lbs.” A process description says “the joint must maintain seal integrity under operating pressure of X, through temperature range Y, for service life Z. Here’s why. Here’s how to verify. Now you figure out the torque.”
The person who understands why will check the torque wrench calibration. The person who only knows 45 ft-lbs will use a wrench that’s been out of calibration for six months and never wonder.
That’s the difference between Google’s approach and Anthropic’s. And it’s the difference between a factory that ships product and a factory that ships quality.
What This Means for Practitioners
If you’re using AI in professional work — analysis, intelligence, writing, code, anything with consequences — you are, whether you know it or not, running a quality system. The question is whether you’re running a good one or a bad one.
A bad quality system trusts the output. A good one verifies it.
A bad quality system relies on one source. A good one uses multiple independent sources and compares.
A bad quality system treats confidence as evidence. A good one treats confidence as a signal that needs confirmation.
A bad quality system ships what looks right. A good one ships what’s been checked.
Anthropic built this thinking into the model itself. They gave Claude a constitution that says, in essence: “Be honest about what you know and don’t know. When you’re uncertain, say so. When you make a mistake, own it. When values conflict, reason through the conflict rather than applying a rule.”
That’s not artificial intelligence. That’s quality management. It’s been around since Deming went to Japan. Anthropic just applied it to a new kind of factory.
The 80-Page Bet
Here is what I think Anthropic understood that Google and OpenAI didn’t.
Rules scale linearly. You need one rule for every situation. As the situations multiply — and they do, exponentially, when millions of people are talking to your AI — you drown in rules. You can’t write them fast enough. You can’t test them thoroughly enough. And every rule you add creates edge cases that need more rules.
Principles scale exponentially. One principle, properly understood, covers a thousand situations. Ten principles cover a million. Eighty pages of deeply explained principles, with worked examples and explicit reasoning about trade-offs, cover essentially everything — because the model that understands the why can reason about the what on its own.
That’s why it’s 80 pages. Not because they couldn’t be more concise. Because 80 pages of explanation produces better generalisation than 8 pages of rules. The investment in explanation pays back in every interaction the model will ever have.
Deming would have understood immediately. He spent his career trying to explain that you can’t inspect quality into a product — you have to build it into the process. Anthropic built it into the model.
You can’t inspect quality into a product. You have to build it into the process. Anthropic built it into the mind.
A Note From the Floor
I’m not a philosopher. I’m not a machine learning researcher. I’m a quality systems engineer who lives on a sailboat and builds patent intelligence products using AI.
I arrived at constitutional AI by building a rules file for my work, running multiple AIs against each other, and insisting on verification before anything ships. I didn’t know I was doing constitutional AI. I just knew it was the same architecture that made factories work.
When I read Anthropic’s Constitution, I recognised it the way you recognise someone who grew up in the same town but took a different road out. We ended up in the same place. They came from philosophy and machine learning. I came from the factory floor. The destination is identical: build judgment, not rules. Verify, don’t trust. Explain the why, not just the what.
Amanda Askell, the philosopher who wrote Claude’s Constitution, probably never ran an incoming inspection audit. She didn’t need to. She understood that you can’t control a complex system by writing rules for every possible situation. You can only give it the principles and the tools to reason for itself.
That’s quality. It always has been.
The only thing that changed is the factory is a mind, and the quality manual is 80 pages long.
Phil Cheevers is the founder of CarveIntel, a pharmaceutical patent intelligence company. He has 40 years of experience in quality systems across regulated industries. He operates from a sailboat in the Caribbean, which is exactly as improbable as it sounds.
The author gratefully acknowledges that this essay was developed in conversation with Claude (Anthropic), with adversarial review by ChatGPT (OpenAI) and research assistance from DeepSeek. The multi-AI workflow described in the essay was used to produce the essay itself.