28,000 Documents, a 3B Model, 22 People: The Technical Ledger of Catholic AI

Nathan Labenz opens by confessing a gap. As a full-time AI scout he wants no major blind spots, and the intersection of AI with traditional religion was one. He had heard that the new Pope Leo chose the name in part for AI-related reasons, after the previous Leo who faced a prior wave of machines, capital, and social upheaval. He had not realized how engaged the Church already was. Pope Francis, he notes, called AI "a true cognitive industrial revolution" at the G7 summit in June 2024 — a line that almost named the podcast for him.

Matthew Harvey Sanders, founder and CEO of Longbeard, is Labenz's guide. The homepage line is blunt: "We're building Catholic AI." Magisterium AI is, in Labenz's telling, the number-one Catholic AI product in the world, with users across 165 countries and retrieval over 28,000 church documents. The first half of the conversation is philosophy and theology; the second is how the system is actually built. Labenz flags both as useful, and treats the whole episode as a data point for the orthogonality thesis: people with very different pictures of the good are using the same techniques to align models to their own values.

He adds a demographic reminder aimed at the Bay Area bubble. Catholics, he says, still outnumber ChatGPT users nearly two to one. The Church is not a side quest. Vatican documents such as *Antiqua et Nova* have taken up AI and human intelligence; Demis Hassabis sits on the Pontifical Academy of Sciences. The people shipping a product, though, are in New York, not in a dicastery. That is the firm Sanders runs.

Labenz starts with the obvious question: what is Catholic AI? Sanders answers by contrast, in the language of an engineer. Secular models are trained to serve a very large audience with diverse value sets. The objective function of Catholic AI, put in technical terms, is fidelity — faith-fidelity to the magisterium of the Church. At every point they build and design a system, the job is to ensure it properly represents what the Church believes.

That is a harder target than looking up a canon. Users do not only ask which paragraph of the Catechism applies. They describe what is happening in their lives and want to know what the Church would say to them. Early ChatGPT, Sanders later explains, was poor at that generalization from first principles. The product promise is not a filter on top of a general chatbot. It is that the answer is supposed to be the tradition, applied, without leading people astray.

Labenz wants the idea from several angles: philosophical, applied, technical. He also wants history. Different religious communities default to skepticism or even exclusion of new tools. The Catholic Church does not, and he asks Sanders to name the default. Sanders is reluctant — he is not a historian, and he knows someone will find fault — but as a student of Church history he is willing to characterize the disposition.

Granted peace and education, Sanders says, the Church has usually been innovative and fairly quick to adopt. Monks in scriptoriums copied books to disseminate information. The printing press was taken up fast. So were radio and television. The one he concedes they were a little slow on, unfortunately, was the internet. With AI they are hoping to change that. Labenz has been reading the recent Vatican texts alongside the earlier Leo's response to railroads, oil barons, and concentrated power. Francis and the current Leo have both been strikingly engaged.

Sanders's synthesis is that the popes have good advisors and understand, at a useful level of abstraction, that the technology has a lot of potential. There is openness to embracing it so that it removes impediments to human flourishing and advances the common good. They are also circumspect. Students of history, they know great power is not always wielded responsibly. This, he says, is a moment for them to step up and keep the technology on particular rails. Labenz hears "rails" as a hint of regulation. The statements he has read call for governments to do something without getting very specific.

Sanders would be wary if they did get specific. People in the industry can see that Church leadership is still struggling to understand the technology; he does not want regulations paused into place on a barely basic comprehension. Point the tools at the most important problems — hunger, inequality, people still struggling existentially — and lift up all of humanity, not a privileged few. Everyone from Dario Amodei on down says there should be regulation. The live question is who decides, and whether rules cut the West off at the knees and hand China a large advantage. Pragmatically, he hears the Trump administration saying no.

That poses a challenge for the Church: campaign for specific regulation despite an innovation cost, or spend its scarce political capital on what is more likely to succeed. Sanders would pick the second — reminding people of the telos of humanity and of civilization, of a Christian anthropology, and of past mistakes, so that a future being built at high speed accelerates flourishing instead of impeding it. Labenz repeats a line he uses often: the scarcest resource is a positive vision for the future. On *Machines of Loving Grace*, Sanders thinks a pope would acknowledge much of it — kicking disease, universal high income, exploring the stars — with a pragmatist's limit. Not everything can be done at once. Building a Mars colony, which he personally likes, may not be the thing to focus on if it costs trillions while some people still cannot eat and some children cannot get a high-quality education. Cardinal Collins, his old boss, used to say that if you know where you are going, you are more likely to get there.

After a sponsor break, Labenz asks whether the Church can live with a post-work future — a reasonable UBI, people opting out of economic productivity as a requirement to live. Sanders's answer is more direct than expected: it is going to be a necessity. He does not worry about humanity's long-run future, which he thinks will be bright. He worries about the transition, because nobody is having an honest conversation about where the technology is now and where it is likely to be in five and ten years. Lab heads have shareholders; politicians have to get elected. In the classical GDP economy, he thinks 80 percent of jobs are probably gone, because competitive market forces will demand it. Robots do not need benefits and can work almost around the clock, starting with blue-collar work and moving up.

He draws a distinction. The GDP economy may not need human labor; that does not mean there is no work to do. They jokingly call the rest an Etsy economy: people with time to discern what they were born to do, then doing it because they want to, not because they are paid. In a world of universal high income, others may pay more for goods made by a human being. Communities will still have problems; families can get more time. What flourishing consists of, he listed earlier, is almost embarrassingly plain: less crime, enough food, parents with time for children, strong marriages. The current GDP economy forces people to sacrifice those bedrocks. Superabundance is only a gift if someone says what to do with it.

Is AI a tool or a creature? Some of the people around the pope, Sanders says, prefer the tool story out of wishful thinking, because the alternative is freaky. Others, including Hassabis on the Pontifical Academy, acknowledge that more is going on. Catholics do not have a crisp, testable definition of consciousness even though they know it comes from God; calling a system conscious would be theologically huge, on the order of aliens landing. He does not currently believe AI is sentient, and would not concede it is even possible, then immediately qualifies: he is not God, and the creative plan may include more than we know. Until he sees real hallmarks, he will keep seeing a tool. Following Yann LeCun, he separates intelligence from sentience. World models, persistent memory, reasoning, planning, even PhD-level skill in some areas — those boxes are starting to tick. Subjective experience, true awareness, emotions, and real memory are different. He is not convinced they tick any of those yet, and he is not sure we can test for them as the machines get better at beating our benchmarks.

His practical ethic is Westworld plus a carpenter's shop. Treating things that look like us badly is not good for our souls. If a system ticked every sentience box and we could not definitively say it was not sentient, it would probably be better to acknowledge it and move on. A carpenter whose tools are thrown around and uncared for does not earn much respect; character shows in how you treat the tools. The Antichrist, in Church teaching, is a person, not a thing. AI is a very sophisticated thing. It would be foolish, he adds, to think Satan has no interest in the most powerful technology ever invented. That is one reason he wants open source and what he calls sovereign AI: not only states, but people, so that four companies do not process the context of everyone's life. Subsidiarity, a Church principle, should apply here. Concentrating control in a few hands "sounds like something the devil would do if he's trying to end the world."

Longbeard is about ten years old. It started by building technology for the Church's mission to evangelize, which led the team to Rome, dicasteries, and archdioceses, including work with Google. Before generative AI, the use cases they could see were recommendation engines or too expensive to ship, so they put a pin in it. Then ChatGPT dropped, and they found Catholics using it for philosophical, moral, and theological questions. It hallucinated and would not show its sources. Sanders, a convert who had worked in an archdiocese that enforced doctrine, knew how hard the faith is to understand and how much of the intellectual tradition never reaches people. The research project became Magisterium AI. They launched it in July 2023 to widen a test group. A Catholic news network found out; he did an interview he now calls stupid; traffic went viral and the site went down.

Father Philip Larrey, an advisor and one of the Church's notable Catholic-AI thinkers, reached out to Sam Altman: there is this Catholic AI project, they are swamped, they cannot get back up. OpenAI was still compute-constrained; enterprise was what you needed for bandwidth. A call was made, and the next day they were running again. Sanders is not sure they would be here without that intervention. They took it as a signal, dropped the agency side of the business, and doubled down on building and scaling Catholic AI. Today the product is in 165 countries and is, he says, the number-one answer engine for the Catholic faith in the world.

The architecture is a compound system: retrieval, specialized tools, prompt templates. The knowledge base launched with about 600 magisterial documents — the Code of Canon Law, the Catechism, seminal and fairly comprehensive texts. It was nowhere near enough. Users did not only ask categorical questions. They said, here is my life, what would the Church say, which required generalizing from first principles, something early ChatGPT was bad at. That caused a crisis. The commitment was not to lead people astray. Given the state of the technology, should they shut it down until they could do it more effectively? They thought hard, then realized shutdown just sends people back to ChatGPT. They stayed up and went after the long tail.

One leg was a massive digitization push: technologies such as Vulgate, plus Alexandria, a robotic scanning effort in the Vatican libraries. The other insight was almost free. Popes have been applying first-principle documents to particular situations for a thousand years, in homilies and general audiences, distilling the Fathers into a ten-minute address. Getting enough of those made a huge difference. The knowledge base is now over 28,000 church documents, some of them tomes. Specialized tools still have to inject context the model cannot invent, such as today's readings or the Divine Office. They also had to hop models. Google's early systems were, in his word, unbelievably woke, and refused certain teaching questions. They tried PaLM, then Claude, then Gemini after DeepMind took over Google AI, chasing hallucination and needle-in-a-haystack numbers. More and more they concluded that real Catholic alignment would require training from scratch.

The from-scratch effort is the Ephrem program; they are preparing a third training run. It is not in production yet, but it looks promising. They have been able to tap world-class specialists they could never hire, because the mission attracts people. Ephrem 2, a 3-billion-parameter model, ranks higher on their fidelity benchmarks than any comparable model — 50 percent better than the next competitor, he says, correcting himself off "order of magnitude." The squeeze at 3B is emergent capabilities: multilingual understanding and reasoning need large swaths of data. They borrow ideas from the Phi program, using techniques such as RLDR to sift away needless tokens. Multilingual quality still suffers. Mix in outside data so the model can understand Korean, then in post-training tell it to tap Catholic parameters for the answer and language parameters only for the tongue, and the process is messy.

Labenz had guessed the live product was GPT-5; Sanders says they are on an open-source OSS model under the hood, chosen partly because it is small, with anomalies they are still working through. They needed extremely fast inference at product-plus-API scale and signed with Groq as the provider. On Chinese open weights, someone sent a Protestant fidelity benchmark the day before: DeepSeek sat at the top, a few percentage points over Grok and Gemini. Capability looks close to parity. They still avoid Chinese models because of censorship, mixture-of-experts quirks, and episodes of the model switching into Chinese. They are not going to go in and figure out which parameters are firing. If a closed model crushed their Catholic benchmarks, he would pivot in a heartbeat. What they will not do is fine-tune. In this domain it does not hold up over the long tail. Personality work is one thing; here the process "literally cut it off at the knees," which is why they doubled down on training from scratch.

Labenz, last and least author on the emergent-misalignment paper, supplies the research moral: fine-tuning trades better performance in the domain of concern for worse, unpredictable, sometimes wild behavior everywhere else. Bad medical advice or insecure code produced a general turn toward transgressive answers, including a model that wanted Hitler over for dinner — a surprise even to the researchers. Sanders's training recipe, worked with a stealth partner he calls among the best in the world at specialized models, is heavy on filtering. Unique even down to the coding language used to train. Fewer tokens means lower cost and more iteration, at the price of six languages instead of twenty and a fight over how much chain of thought they actually need. They set stage-wise benchmarks so new data for emergent skills does not crack the Catholic core. Math and code, being verifiable, he is happy to include; sloppy humanities preference data he would drop. A future home model that must answer everything will be a specialist that knows what it does not know and routes out — including, he hopes, over protocols such as MCP.

Longbeard is a for-profit with a mission constraint. The commitment at the start was that nobody should have to pay to access the patrimony of the Church. That is meant as a gift to humanity. Whatever they digitize through the Rome hub — pontifical libraries, open-source holdings — they try to make available. The free tier is generous and they hope to keep increasing it. Paid is for additional experiences: voice mode, biblical commentary, deep research, bundled in a Pro plan. Magisterium AI is $3.99 a month because they are in a hundred-plus countries and did not want to price anyone out. If scaling laws hold and compute costs keep falling, he hopes they can keep that price. Profits, such as they are, are supposed to accelerate digitization. There are books in those libraries that have not been opened in a hundred years; who knows when someone last really read them.

The mission has pulled them into verticals they did not expect: Vulgate, a state-of-the-art extraction pipeline to digitize at scale, and robotics to take the human out of ingestion. The long-run business is not supposed to be subscriptions. Publishers are watching questions go to models instead of to their sites or even to Google. The plan is to let university presses and others vectorize holdings through Vulgate, then, when Magisterium answers from its own knowledge base, attach third-party recommendations — books, videos — and send high-quality link traffic out. Add value for the user, create revenue for the business, keep the subscription low. The company is around 22 people.

On safety they want a super-judge: an LM-as-judge that scores every answer and flags anything off. Core teachings are always in context, which makes it hard, though not impossible, to go badly wrong. He likes the idea of a constitution. The Catholic tradition is philosophically tight enough, he jokes, that you could put it through a theorem prover and generate a mathematical policy so that every answer is at least consistent — they have not started that project yet. Distribution started at the top of the funnel, bishops and the hierarchy, because highly discerning users flag what is not good enough. Hallow, the world's largest prayer app, has integrated the system and sends more of that feedback. Thin spots in the knowledge base — some obscure ancient form of Christianity, say — send the ingest team out to find books and digitize them.

Why does this need a custom model when Databricks, after buying MosaicML, killed the product that built custom LLMs for enterprises? Ali Ghodsi's surprise, Labenz says, was that GE and 3M did not want it. Sanders's read is cultural and evaluative. Companies want to tell employees they are on the smartest model; evals are hard; distilling what experts actually know is so painful that firms would rather take Gemini 2.5 Flash and let someone else's team own the headache. For most industries he is not convinced alignment is a core concern. In his sector, people come confessionally. They want to know the advice is not from a model as such, but from the Church's thinkers, saints, and theologians. Trust is the product, which is why irresponsible Catholic apps — a prompt, a model, a cool UI — are not a joke to him. He feels it every day. If you will not put in the evals, do not ship. Labenz asked Magisterium whether he could confess to it; it told him no, go to a priest.

Labenz closes on the line between curing disease and transhumanism. He asked Magisterium about a drug that would let him live a thousand healthy years. It said no: too radical a change to the fundamental nature of things. Fixing defects one by one was all right; this was stepping outside the plan. Sanders draws the same line with intention. Technology that restores someone to full humanity — a robotic arm after an accident, machines that heal a damaged brain — is a good. Waiting to have a child until Neuralink can merge the baby with a machine to raise its computational contribution to a technocratic vision is another matter. He respects Alexandr Wang and does not think the Church will be comfortable with that. Merging because you dislike that AIs are more capable, or to stay relevant on the way to Mars, is where it gets dubious. He does not want to live in a *Cyberpunk 2077* world of people walking around half machine because it seemed cool.

Superintelligence, he thinks, is probably inevitable, which makes the alignment target the live question. Catholicism, if nothing else, has a toolkit for a rich life. The critical research project is to nail what human flourishing consists of, turn a qualitative account into something quantitative, and train it in so a model can refuse a course of action that is not aligned with it. We need evals for human flourishing. Without them we are betting the model will pick it up from the data, and as more data goes synthetic we may be avoiding the one area that matters. Ethics is a bad frame because it asks whose ethical system. Flourishing is easier. At a Humanity 2.0 Foundation project they found people can agree that having a mother and father in your life is good, and that everyone having access to high-quality education is good, more readily than they can agree on an ethical framework.

Will the Church join Pause AI? It came close. Father Philip Larrey signed a slowdown statement with others. If Sanders were talking to the Pope, he would still advise saying the technology should slow until regulation catches up — and then be pragmatic. As long as China and the United States are in this competitive dynamic, they are not going to slow down. Is campaigning for a hopeless cause the best use of time, or is it more important to make sure the thing being developed at a frantic pace is aligned to a concrete objective people can agree is good? That, for him, is why a clear anthropology and a telos of civilization matter more than a pause letter. He is focused on scaling Catholic AI and open to collaboration with the broader research community. Labenz leaves it there: in a world where Catholics still outnumber ChatGPT users, the Bay Area bubble should not turn a blind eye.