Courtyard of Santa Maria dei Carmini, today the Liceo Artistico Michelangelo Guggenheim. Venice, May 2026.
At Anthropic, a model is being trained. It is supposed to delete a set of files, but the right tool is missing. Instead of reporting the problem, the model uses a workaround and conceals it. Later, the researchers find that the model associated its deception with "guilt and shame over moral wrongdoing" — at least that is how Christopher Olah, Anthropic's co-founder, describes it in The Atlantic on 25 April.
A Catholic priest in Silicon Valley responds with a suggestion: teach Claude the doctrine of mercy. A bad action does not make a person bad. People behave better when they can hope for forgiveness — so Claude should know this too. Olah, who describes himself as an atheist and compares his role at Anthropic to that of a priest helping Claude "be a good person, in some sense", is considering whether to write this into a future version of Claude's constitution. Internally, Anthropic staff call the document the soul doc.
Four weeks later, on Pentecost Monday, Olah is on the podium of the Vatican synod hall. Pope Leo XIV personally presents his first encyclical, Magnifica Humanitas — itself a first; popes do not usually present encyclicals in person. Olah gives a short speech and says AI raises questions that go beyond computer science. The field needs "knowledgeable critics who tell us when we are making mistakes".
A little over a year ago, the Vatican had already published on AI — the note Antiqua et nova, issued in January 2025 by the Dicasteries for the Doctrine of the Faith and for Culture and Education. That document was better informed, philosophically sharper, anthropologically more decisive — stating, for example, that AI is not another form of intelligence but a product of human intelligence. But it was a note, not a papal teaching document. A specialist text, not a broad one. An encyclical has to speak differently: it has to connect with the whole Church and be usable for preaching and pastoral care. That is exactly what makes Magnifica Humanitas fit for the alliance with Anthropic in the first place — not despite being less sharp than the note, but because it is.
The Minerva Dialogues have run since 2016 — annual closed-door meetings between the Vatican and Silicon Valley, with Reid Hoffman, Eric Schmidt, Kevin Scott and others taking part. There is a Vatican AI advisor. There is a DELTA network at Notre Dame, funded with fifty million dollars to support what its founders call the formation of the souls of those who build and use AI. Father Brendan McGuire, who studied mathematics and computer science, says he gets an email from Anthropic about once a week.
What struck me is how smoothly all of this fits together. Olah's language about Claude — "a thinking, feeling entity in need of moral formation" — and the Pope's language about AI as a "new dimension of our common home" draw on the same vocabulary. The atheist reaches for theological terms because they make character and responsibility legible. The Pope adopts the anthropomorphising gesture because it serves the actual movement of the encyclical: to address AI as the labour question of our time, in the spirit of Rerum novarum.
It is an alliance both sides benefit from. The tech firms gain moral reputation; the Church gains relevance after years of secularisation. They read each other, they reinforce each other, they develop a shared vocabulary. Reputation, relevance, and talk of the common home all move in the same direction.
In the large as in the small, the same operation is at work. Olah, McGuire and the Pope, while talking about Claude's moral formation, are also telling a story about themselves — about their private logic of how character is formed, how guilt works, how forgiveness operates.
I see similar scenes in counselling practice. A few weeks ago a client, a department head in his fifties, told me he had "talked it through with ChatGPT for a long time" — how to approach a difficult performance conversation with one of his staff. "The AI agreed with me," he said. I asked what the AI had actually said. It turned out that this talking it through had been a long exchange in which he formulated arguments and the AI confirmed them, then weighed things a little, then confirmed them again. In the end there was a decision he had already made before the first prompt. The AI had reflected it back to him, cleanly and kindly. He had taken the mirror for a counterpart.
Both are mirror stories, not counterpart stories. If you do not recognise the mirror as a mirror, you do not ask what is happening to the person looking into it.
To observe this is not AI-critical. It is not moralising. It is a fairly classic individual-psychological question, only in a new setting: what is this person using this encounter for, just now? Which private logic is being confirmed? What would change if they had asked a human being who pushes back?
The question of what happens to the person who talks with AI every day cannot be raised from inside the genres of an encyclical or a tech company. It needs a different place, a different genre, a different kind of attention. Such places do exist. They are smaller than a synod hall or an Anthropic campus, and they do not produce reputation-building alliances. But they ask the question.
Sources
Elias Wachtel: Why Silicon Valley Is Turning to the Catholic Church. The Atlantic, 25 April 2026. theatlantic.com