What the AI Critique of the Encyclical Got Wrong

The most shared critique of Pope Leo's encyclical on artificial intelligence was written by artificial intelligence. The lesson almost everyone drew from that is the wrong one.

Within days of Magnifica Humanitas, a critique of it went viral in Catholic circles. Thirty points, sharp and confident, taking the Holy Father to task on just war, on Babel, on human rights, on the Church and slavery. It was also, by its own author's open account, written by a chatbot, disclosed openly, never disguised. Hand the machine the encyclical, ask it to find the weaknesses, publish what comes back.

Then someone sat down and actually read it. Point by point, a careful reader showed that the critique was riddled with the particular mistakes these systems make. It misread which paragraph came first. It dropped the context around a quotation. It put words in the Pope's mouth that the text never used. It attributed a phrase to the wrong document.

The lesson most people took from this was simple. See what happens when you use AI on something serious.

That is the wrong lesson. The failure on display had a name long before anyone trained a language model. And the difference between that viral mess and careful work comes down to two ways of working, not to human versus machine. I spent that same week practicing one of them on the same document, hard, for days. What came out of that, I would hand to a hostile reader. The machine was the same. The method was not.

The oracle

Most of us, when we reach for one of these tools, use it as an oracle. You type a sentence or two into a box, and you take what it gives you. A question in, an answer out. It is the natural way to use a chat window, and almost everyone works this way, including people far smarter than I am.

On one of my own podcast panels this spring, Fr. Philip Larrey, a philosopher who has advised the Vatican on artificial intelligence, heard a description of how serious researchers actually prompt these systems and said, with complete honesty, "I've never seen a fifty-page prompt before. Prompts are usually like a sentence or two, right?" That is the honest starting line nearly all of us share, and no sign of a man behind the times. I shared it too, longer than I'd like to admit.

The trouble is that a one-sentence prompt does not get you the machine's careful reading of a text. It gets you the machine's reflexes: its best guess at what an answer should sound like, drawn from everything it has ever absorbed, lightly bent toward whatever you pasted in. There is no step where anyone checks anything against anything. And the tell, every time, is that the person hands over the judgment along with the labor. The viral critique did not merely borrow the chatbot's points. It borrowed its posture: the confidence, the faintly annoyed tone, the verdict. That is the abdication. Using a machine was fine. Nothing happened after the machine spoke.

The chatbot was still useful in places. It correctly flagged a redundant sentence, a Latin typo, a broken link in the Vatican's own text. Keep that in mind. It turns out to be the whole game.

The shame is on the wrong thing

There is a reflex, in Catholic circles and academic ones alike, to treat using AI to research or to write as something faintly shameful. A shortcut. A way of not doing the work.

I want to put the shame where it belongs. It does not belong on augmenting your judgment. It belongs on abdicating it.

And here the encyclical is more help than the people quoting it against AI seem to notice. Leo's fear is about what happens when a tool does the part of the work that was meant to form us. He warns against information that extinguishes the desire to ask questions. He names the danger of an education in which "an incessant flow of information replaces the essential exercise of research, reflection and discernment." And of the machine itself he says the line everyone now repeats: that its learning is "a form of statistical adaptation based on data and feedback," effective, but one that "does not imply inner growth."

Read those together and you have a criterion. The danger is letting the tool spare you the slow interior work of reflection and discernment by which a person is actually formed. Which means the real line runs between those who stay inside that interior work and those who quietly hand it off, whether or not they use AI. The whole question is whether a person, carrying judgment the machine does not have, stays in the loop.

That is what I mean by method. Not a clever prompt. A person who refuses to leave.

Prompting versus orchestrating

There is a name builders use for the difference, and I'll borrow it rather than pretend I coined it. Most use of these tools is prompting: a single request, the whole job handed over in one breath. Serious use is closer to orchestrating. You start from the goal and walk the work through its real stages, the way any careful writer already does, only now out loud and with a collaborator who never tires.

An old piece of management wisdom hints at why the one-shot version fails. You can tell someone what to do, or how to do it, but you cannot really tell them both. "Critique this encyclical" is pure what, and it smuggles in a hundred hidden hows. Which sources count. What standard of proof. When to read a phrase as a hedge rather than a claim. When to refuse a verdict for lack of evidence. You cannot pack all of that into one instruction, so either you leave it out and inherit the machine's defaults, or you spell it out and you have done the work yourself.

Orchestrating escapes the bind by refusing to ask for everything at once. You break the one impossible request into a sequence of small, ordinary ones. Gather the sources. Now hold them against each other. Now show me where the actual thesis is. Now find where the argument breaks. Each step is small enough that you can see what the machine did, and correct it, before the next one begins. The judgment lives in the spaces between the steps.

There is a nuance here that matters for Catholics especially. With a capable person, tell them what, not how is good counsel. You trust a colleague with the how because they bring their own judgment to the how. With these machines, that logic inverts. The model has no practical wisdom of its own (no phronesis, to use the older word) and no stake in whether the work is any good. So you cannot hand it the how. You keep it. The machine is an instrument whose judgment you have to supply, unlike a junior colleague you delegate to, because it has none worth trusting.

In practice this is only a handful of stages, run almost every time: build a narrow, trusted library to work from instead of the open internet; compress it down; find the thesis; draft; then verify every claim against its actual source. And running underneath all of it, one habit I would recommend to anyone, with or without a tool in hand: keep a deliberately hostile reader, and turn it loose on your own work before anyone else can.

Mine has a name. I call it Kontra. It is an AI agent I have set up for one purpose: to be merciless. I point it at my own drafts and arguments and tell it to tear them apart: as a skeptic would, as a secular journalist would, as a tired pastor would. On one of those same panels, my guest Letty Garcia mentioned, warmly, that there is one AI tool she will not use, because "it's mean to you." I laughed, because that is exactly why I use it. The meanness most people sensibly avoid is, for the single job of pressure-testing your own work, the entire point. Kontra has gutted more of my drafts than I would like to admit, including an early version of this essay. That is the method working.

What the method can and cannot do

Let me be honest about the limits, because the temptation is to claim too much.

The method cannot give you knowledge you do not have. The careful reader who dismantled the viral critique caught, among much else, that John Paul II really did call human dignity "infinite" in a 1980 Angelus address, in Italian. No verification rule surfaces a fact you never suspected you needed. You cannot check an attribution you do not know is wrong. The rebuttal worked because a man who knows the tradition deeply sat down and read with care. The method was downstream of the knowledge.

Hold that, and then notice something almost funny. Read that rebuttal as a procedure instead of an argument, and it is a nearly perfect description of how you would orchestrate the task well. Where the chatbot claimed the document "ignored" an objection, the reader checked the next paragraph and found the answer sitting right there. Where the chatbot lifted a phrase, the reader traced it to what it actually referred to. Where the chatbot invented a word the Pope never used, the reader quoted the real one. Where the chatbot delivered a verdict, the reader withheld one and showed the evidence instead. Verify in sequence. Trace every quotation to its source. Honor the text's own hedges. Tell a typo apart from a misreading. Render no verdict; surface the findings and let a human judge.

The man wrote the recipe without meaning to. But the recipe alone leaves out the one thing that made it work. He could write it only because he already possessed what no prompt supplies: the formed judgment of someone who has lived inside this tradition for years. He is the part of my method a machine cannot be.

Which is the real lesson, and a less flattering one than the viral take. The difference between the slop and the careful work was never human versus machine. It was whether a person who knew what he was doing stayed in the loop. The method amplifies that person. It cannot replace him.

The hard question

I would be running the very con I am describing if I stopped there, because there is a serious objection I have not answered, and it comes from the encyclical itself.

Grant everything I have said. Grant that I stayed engaged, checked the sources, supplied the judgment. The encyclical's deepest worry is what happens to me, whatever the accuracy of my output. Leo's fear is that we let the tool do the interior work, and the interior faculty, the one that grows only by being used, quietly wastes away. And here is the trap. The better my method, the more of the labor I move off myself. I feel intensely engaged the whole time. But engagement is not formation, and feeling stretched is not the same as being changed. The most disciplined version of this work might be the most seductive form of the very thing the encyclical warns against, precisely because it feels like virtue.

I take that seriously. I take it seriously enough that I have made the argument myself, at length, in another essay. I do not think there is a clean answer, and I distrust anyone who hands me one. What I will say is this: the safeguard was never the tool, and it was never the method either. It is whether you remain the kind of person who is still doing the forming work somewhere: still reading the hard thing slowly, still sitting with what you do not understand, still letting yourself be wrong before you let yourself be helped. Used one way, this work keeps you in that posture. Used another, it lets you flee it while feeling busy. The tool will not decide which. You will.

That is uncomfortable, and I think it is meant to be. We are not called to resolve this tension cheaply, to pronounce AI good, or pronounce it bad, and be done. We are called to stay in it. That is where discernment comes from in the first place.

The mirror

Hand the same machine to two people and you can read the difference in what comes out. One produces, in the rebuttal author's own phrase, slop passed off as analysis. The other produces something a person checked and stands behind. The tool is a mirror. The line is older than I am; Shannon Vallor and Marshall McLuhan got there first. What it reflects is the person holding it. The machine did not choose. The person did, at every step, or at none.

So the answer the encyclical points toward is to stay human and engaged, and to use the thing well, with neither fear nor surrender. That is hard, and slow, and impossible to fake, because it asks for the same interior discernment, the same slow forming work, we were told to practice all along.

It was never the machine. It was always the method. And the method, in the end, is a person who refused to stop thinking.