How a letter became a book, and a book became a method
For seven centuries, silver in England has carried a small stamped mark — a leopard’s head, at first — and the whole institution turns on one fact: the mark was applied at the hall, not in the workshop. The person who makes the thing does not certify the thing. I will come back to the hallmark properly in a few weeks, when we reach evidence. I open with it today for a smaller reason. When I decided to write this book, the first thing I had to decide was what I was not: not the hall, not the assayer, and above all not a commentator on the Pope.
You know the shelf I mean. Principles for Trustworthy AI. Responsible AI Frameworks. Ethics Guidelines, Charters, Manifestos. I have read most of them, some with real admiration. And I have noticed that they share a fate: they are agreed with. Universally, warmly, and without consequence. Everyone is for transparency. Nobody has ever been found in breach of a manifesto.
So the question that organised my summer was not “what does the letter say?” — it says it very well by itself — but: what would it take for a text of principles to become something an engineer can be held to? Not inspired by. Held to. Checked against, on a Tuesday, by somebody who was not in the room when the text was written.
That is a different kind of question, and it forces a different kind of reading.
Let me tell you how it found me, because the origin explains the method.
A year earlier I had published a book on how AI agents are orchestrated — many specialised agents composed, like an orchestra, into something greater than any soloist. It was a book about machinery, and it ended honestly on a question the machinery could not answer: if artificial agents collaborate this well with each other, what role remains for us? I left the question standing, because architectures, protocols and benchmarks have nothing to say about it.
Then the letter arrived, and a few days after reading it I realised two things at once. The first was something the earlier book had been too cheerful to dwell on: an agent is an effective tool, and effectiveness has no preferences. The same properties that make software that acts so useful — it scales, it never tires, it addresses each person individually, it is unfailingly polite — make it, unchanged, the most effective instrument of exploitation ever built. The agent that helps a customer and the agent that flatters him while taking his measure can be the same code with a different instruction. The difference is not in the engineering. It lies in the rules a system is held to, and in whether a person answers for it.
The second thing I realised was that the letter already contained those rules. Not as slogans — as behaviours. Concrete things that rose from its pages, that could be named, required and checked. This book is the second panel of a diptych: the first showed the instruments; this one asks who the music is for, and what it must never cost.
Here is the rule I set myself on page one and kept to the last: the Pope’s voice is never bent.
It sounds like piety. It is a method. If I had written a book about the encyclical — what I think of it, where I agree, what it means for the industry — I would have added one opinion to a market that has no shortage of them, and the Pope, who does not need my opinion, would have been entitled to ignore it. So I did something humbler and much harder. I read the letter the way a builder reads a specification: closely, sceptically, paragraph by paragraph — there are 245 of them, numbered precisely so that they can be quoted, challenged and checked — and of each paragraph I asked one question. Does this say something that must be true of a system that acts?
Most paragraphs do not; they are theology, or history, or pastoral care, and I left them where they are. But nineteen times the answer was yes, and each time I did three things. I recorded the paragraph numbers. I re-expressed what the paragraph requires in design language — dropping the vocabulary of faith, which I flag wherever it appears, and keeping the structure of the claim. And I gave the result a code and a name, so that it could be handled.
Let me show you one, because a model without an example is a sermon.
Paragraph 105 says that what must be guaranteed is “the possibility of identifying who must account for decisions, justify them, monitor them, and, when necessary, challenge them and remedy any harm caused”. Five verbs: account, justify, monitor, challenge, remedy. Paragraph 103 names the evil on the other side — selection that no one answers for. Paragraphs 102 to 105 and 199 together become P3 — Human accountability, identifiable end to end: for every consequential output, an identifiable human answers. Four lines from text to principle, and every line carries its number. If I have bent anything, the number is there to catch me.
That is the move. The question is never “who authorised you to interpret the source?” It is “is the derivation rigorous and verifiable?” — and that question is meant to be put back to me, by anyone, with an obligation on my side to answer.
The nineteen come in two kinds, and the first letter of the code tells you which.
Seven are foundational — F-principles — and they are borrowed from a tradition much older than computing: dignity independent of performance; the common good as the measure of a system’s purpose; the universal destination of digital goods; subsidiarity; solidarity, including with the invisible; justice as a design precondition; integral human development as the decisive test. They tell you why.
Twelve are operative — P-principles — and they are the letter’s direct reply to this technology: non-equivalence between agent and person; the non-neutrality of design; accountability end to end; non-delegation of irreversible judgment; traceability and contestability; discussable normativity; the person strengthened, not replaced; embedded limits and the legitimacy of non-use; truth and non-manipulation; no exploitation of fragility; justice across the supply chain; care for the common home. They tell you where.
I will give each of them its own space in two weeks. For today, notice only the shape: a constitution and its articles. A constitution does not regulate traffic, but every traffic rule can be tested against it.
From there the book took the only shape it could. Part One explains the letter, one theme per chapter, always in the same three steps: what it says, with the paragraph numbers; what it means in your life — the loan, the shift, the chatbot; what you can do. It ends by folding the nineteen into three questions you can carry into any meeting — who answers? does it strengthen the person? where does it stop? — and then hands you a door: the Tester, which asks those questions of a real agent and answers with evidence. Part Two is for people who carry professional responsibility: how the nineteen principles were forged into twenty-four criteria, how an examination works, who signs, and why the numbers that matter are being calibrated on real systems rather than invented at a desk. Part Three sets the result beside the laws of Europe, the frameworks of America, and the rules being written in China and Saudi Arabia.
And then it does one more thing, which I am prouder of than anything else in it. The last pages apply the standard to the book’s own making. Twenty-four criteria, read against the process that produced the text you are holding. Seven are met. Two, no record can tell. Three are not met — and the page says which, and why.
A book about the encyclical would have been one more opinion. A book from the encyclical is a tool. The difference is not modesty; it is the whole method, and it is what makes the result checkable.
And checkable is the point. Nothing I will show you this autumn asks you to share a faith. It asks you to check the references. The letter’s own introduction ends with two words that contain everything that follows: remaining human. I have tried to turn those two words into something an engineer can be held to. Whether I succeeded is, from here on, your call — paragraph by paragraph, and I will be listening.
— Maurizio
When the Machine Is Not Enough — a season of six letters on building AI agents that respect the people they touch. The book is Neemia: Guarding Human Dignity in the Age of AI Agents, on Amazon. The Tester reads your agent in ten minutes, free: try.neemia.org.
If you want to take part — as a builder, an assessor, a critic, a pilot participant — reply to this email. I read everything, and I answer.
Maurizio Grassi
When the Machine Is Not Enough arrives by email every second Monday.