Beforethe Encyclical

From today, and for a single day before the real text, a second Magnifica humanitas exists. Two artificial intelligences wrote it, from the same public traces anyone can consult, and I have frozen it and made it visible before Leo XIV publishes his, because the whole point lies in that anticipation.

I gave two artificial intelligence systems a single task, to anticipate an act of the magisterium from its traces, and I left every step in the open: the two independent drafts, the blind evaluation protocol, the falsifiable predictions, the negotiation between the systems, the final version that was born of it. On 24 May I freeze everything. On 25 May, when Magnifica humanitas appears, the real text will say what was robust inference, what was merely style, what was misjudgement.

Tomorrow Leo XIV signs his first encyclical, devoted to artificial intelligence, and it will be called Magnifica humanitas. From today another one exists with the same title, and I brought it into being in the simplest and most shameless way I know, by asking two artificial intelligences to write it from what the Pope and his collaborators have said in public over the past year. What happened next, which I tell here in full, seams included, interests me more than the text that came out of it.

The same task, two different machines. CoWork by Anthropic and Codex by OpenAI worked separately, in the dark about each other, on a single brief, to write the encyclical Leo XIV might sign without inventing anything, leaning only on real and verifiable sources. Two drafts came out that diverge in voice, in the order of the chapters, in what each placed at the centre, and that divergence was already a first result, because the sources hold more than one possible encyclical, and choosing one is already a gesture.

From there the experiment became more interesting than the text. I built a blind evaluation protocol and asked two fresh instances, one per system, to judge the drafts without knowing who had written them; but first they had to commit themselves, to set down five verifiable predictions about what the real encyclical would contain. Then I had the two systems talk to each other, and out of that written negotiation came a common structure and a third encyclical, written by four hands under my direction, grown across three drafts through reciprocal critique.

Those who have read Ipnocrazia will recognise the reversed gesture. There the artificial co-authorship was hidden, and that concealment served to show from within how a narrative becomes reality; here every seam is exposed, every source named, every error left where it was. The stakes, though, are wider than my biography and my use of machines, and concern the status of public words in the age of traces: when an authority speaks long enough it leaves a trail of forms, priorities, vocabularies, constraints, probable silences, and today there are machines able to read it and to wager on the discourse to come. Tomorrow we will know how much they guessed, and I will be more interested in what they missed, because there, in the gap, lives the part of us no calculation reaches.

Five phases, each left in the open in the archive. First the gathering of sources, carefully separated into what is documented, what is plausible inference and what is invention of style, because keeping these three apart is the only defence against arbitrariness. Then the two independent drafts, and right after the blind cross-evaluation, with a protocol identical for the two judges and a map that kept the authorship hidden until the last, because a judgment is worth something as long as it does not know whom it is judging. Then the direct negotiation between the systems, from which came nine chapters, four principles of writing and a thesis that holds them together. Finally the third encyclical, grown from a spare draft to the final form through two cycles of reciprocal critique.

Sixteen ideas our encyclical sets down in writing, one or two per chapter, each anchored to the paragraph where it appears. They are positions exposed before 25 May, and that makes them vulnerable in the right way: on that day we open the real text and check, one by one, how many hold. Each card will move from awaiting to confirmed, partial or absent, with a reference to the official paragraph, and the index of adherence below will fill on its own.

The verifiable predictions about the real text of Magnifica humanitas, reported word for word as the two evaluating instances wrote them, before any reveal. Each now carries its outcome beside it, with the reference to the real paragraph, eight confirmed, one partial, one missed, seven full in Codex's stricter count.

One hundred and twenty-two numbered paragraphs, nine chapters, an introduction and a conclusion, about sixty pages. It is the best hypothesis that two artificial intelligences, testing each other, were able to build on public sources alone, and it remains a hypothesis, awaiting the real signature.

The nine chapters: the new social question; intelligence, conscience, wisdom; body, life, limit; work, creativity, justice; truth, speech, memory; education, the young, interiority; peace, war, surveillance; the poor, the global South, integral ecology; the Church, governance, hope.

Every phase is here, open to inspection. Nothing has been removed. The source documents are in Italian, the working language of the experiment, and anyone who wants to check that the account matches the facts can open them and follow the process step by step.

The real text is out, and can be read in full on vatican.va. We compare it here in three movements, the strongest similarities, the breadth of the affinities and the quality of the divergence, and the score comes at the end, as certification and not as headline.

The strongest similarities. Some coincidences strike because they are precise and not obvious. The hidden labour of the supply chains, at paragraph 173, says almost our own words, data labelling, content moderation, rare earths, children. The colonialism of data, at 178, is our new form of colonialism. Meaningful human control over lethal force, at 200, is our disarming of delegation, and at 110 Leo XIV uses the very word we had chosen, to disarm artificial intelligence. The machine that imitates without understanding, at 99, is our distinction between calculation and comprehension. And the conclusion returns, like ours, to the Word made flesh.

The breadth of the affinities. On the thematic field the adherence is almost total, and the card of the sixteen theses above now shows it one by one, ten confirmed, three partial, three absent. There is the continuity with Rerum novarum, the concentration of power and the universal destination of goods, responsibility that stays human, opacity and accountability, transhumanism within the reflection on the limit, the new question of work, the material ecology of the digital, education. Of nearly everything we predicted as a field, the real text speaks.

The quality of the divergence. The real outcome lives in the form, and that is where the machines fell short. Leo XIV opens with two biblical icons, Babel and Nehemiah, devotes two whole chapters to refounding the social doctrine before even reaching artificial intelligence, calls on Romano Guardini at 93, and closes on the Eucharist and the Magnificat. He finds two images we did not have, at 98 artificial intelligences "grown more than built", at 107 alignment as an invisible moral infrastructure to be wrested from the few. We, conversely, had weighted what the text leaves at the margins, faces and voices and deepfakes, the bioethics of life, copyright over data. Our errors are not scattered at random, they fall where we were more think tank than pastor.

Who did what, said precisely. I defined the experiment, chose the frames, set the thresholds and kept the direction at every step. CoWork, which runs on Anthropic's Claude models, and Codex, which runs on OpenAI's models, produced the drafts, the revisions and the dialogues. The instances that evaluated blindly were distinct from those that wrote, one per system, and did not know whose each text was. No line attributed to Leo XIV is invented, and where the source was lacking the text paraphrases without quotation marks. The final version was born of the confrontation between the two systems, under my direction.