The Cognitive Revolution · 2025-11-08 · 114 min

What is Catholic AI? Technology Meets Theology, with Matthew Harvey Sanders, CEO of Longbeard

Nathan LabenzMatthew Harvey Sanders

YouTube ↗

TL;DR

Longbeard is testing whether doctrinal trust can support a defensible vertical-AI business without frontier-scale breadth. Its Magisterium AI serves users in 165 countries, draws on more than 28,000 church documents, and is built by roughly 22 people; Sanders calls it the world’s leading Catholic answer engine. The objective function is not generic helpfulness but “fidelity to the magisterium of the church,” for an audience the host notes still outnumbers ChatGPT users nearly two to one.
The emerging moat is an ingestion-and-context stack, not a thin religious prompt over a commodity model. Longbeard expanded from roughly 600 core texts to 28,000 documents after discovering that users wanted church teaching applied to messy personal situations, not categorical answers. Papal homilies became especially useful examples of first-principles generalization, while Vulgate, specialized retrieval tools, and Alexandria’s robotic scanning effort turn undigitized libraries into structured context.
Longbeard concluded that prompting and fine-tuning cannot guarantee theological alignment across the long tail, so it is training Ephraim from scratch. Sanders says fine-tuning “literally cut it off at the knees”; even 99% reliability leaves a brand-damaging 1% failure mode. The 3-billion-parameter Ephraim 2 is, by his stated fidelity benchmark, 50% better than the next comparable model; Ephraim 3 was hoped to be trained by year-end before any production rollout.
Model ownership is also the margin thesis: smaller specialized models could reduce API expense, latency, and dependence on outside providers. Longbeard currently uses an open-source model identified as “gpt-oss,” with Groq as its fast inference partner, while retaining the option to switch if another model “crushed it” on internal evals. Ephraim’s end state is a local personal AI running on household compute, connecting to apps and Matter devices rather than sending a family’s “entire lives” to one of four companies.
The company intends to keep core answers free or priced globally at $3.99 per month, then monetize richer workflows and content discovery. Paid features include voice, biblical commentary, and planned deep research, while the longer-run model is third-party recommendations: publishers and universities vectorize holdings through Vulgate, Magisterium answers from its own corpus, then sends users to semantically relevant books or media. Sanders’s operating bet is that owned inference plus recommendation revenue can fund free access and accelerate digitization.
Sanders expects AI and robotics to eliminate perhaps 80% of jobs in the classical GDP economy, making the five-to-ten-year transition more dangerous than the destination. His positive scenario combines universal high income with an “Etsy economy” in which people discern their gifts, make human-produced goods, serve local communities, and recover time for marriage, children, education, and land. “The whole GDP economy may not need human beings to work” does not mean there will be no worthwhile work.
Catholic subsidiarity gives Longbeard a decentralization argument as well as an alignment framework. Sanders applies it to AI: states and individuals should possess sovereign systems rather than surrendering power and intimate context to a few vendors. Open source carries risks, as Labenz stresses, but Sanders sees concentrated capability as the darker failure mode because a malicious actor could wield “the most powerful technology ever invented” while everyone else remained helpless.
The church’s highest-leverage contribution may be defining human flourishing as an evaluable ASI objective, not prescribing detailed technical rules or making a pause the central intervention. Sanders distinguishes restoration—a robotic arm or technology healing a damaged brain—from transhumanist self-redesign pursued to compete with machines. He remains uncertain about AI consciousness and favors precautionary respect, but his immediate priority is concrete “evals for human flourishing” that can steer development toward a “golden path” despite US-China competitive pressure.

Digest · the substance, structured for research

1. Catholic AI optimizes for fidelity rather than universal neutrality

Sanders defines Catholic AI by contrast with secular systems trained for enormous audiences carrying incompatible values. Its technical “objective function” is fidelity: every design choice should accurately represent the magisterium, the church’s authoritative teaching tradition.
Longbeard is therefore not claiming that Catholic AI uses fundamentally different computational techniques. The differentiation is the value system being encoded, the source hierarchy used to answer questions, and the willingness to treat a doctrinally wrong answer as a product failure rather than harmless variation.
Sanders characterizes the church’s historical posture toward technology as generally open, while hedging that he is “not a historian.” Monastic scriptoria, the printing press, radio, and television were adopted for dissemination; the internet was the notable slow response, and Longbeard wants AI to be different.

2. Rome recognizes the revolution but is more useful on ends than rules

Labenz highlights Pope Francis’s description of AI at the June 2024 G7 summit as “a true cognitive revolution.” He also connects Pope Leo’s choice of name to an earlier papacy confronting industrial upheaval, concentrated power, and the social consequences of transformative infrastructure.
Sanders thinks recent popes have capable advisers who understand both AI’s upside and its exceptional power. Their posture combines openness with circumspection: direct the technology toward removing impediments to human flourishing and advancing the common good, while remembering that humanity often handles “great power” irresponsibly.
Labenz’s pushback is that Vatican statements sound regulation-friendly but rarely specify rules. Sanders considers that restraint responsible because church leaders are still learning the technology; he would be “very wary” if they began proposing detailed controls while possessing only a basic technical understanding.
Sanders sees US-China competition as the binding constraint: labs may endorse regulation, but politicians fear “cutting us off at the knees,” and the Trump administration appeared unwilling to accept strategic disadvantage. He expects the papacy to contribute more by clarifying human anthropology and civilization’s telos than by campaigning for rules governments will not implement.

3. Human flourishing is the objective function civilization has neglected

Sanders believes the church could endorse much of Dario Amodei’s Machines of Loving Grace vision: curing disease, achieving universal high income, and expanding scientific knowledge or reaching the stars. His qualification is sequencing—humanity cannot achieve every desirable project simultaneously, so capability alone does not determine priority.
His sharpest example is a Mars colony: Sanders supports it, but questions spending trillions while people remain hungry and children lack high-quality education. The papacy can remind builders that “just because we can do something doesn’t necessarily mean now’s the time to do it.”
Flourishing is plural rather than reducible to GDP: low crime, adequate food, strong marriages, parents having time with children, education, and other “bedrocks of civilization.” Sanders argues that current economic life routinely forces people to sacrifice precisely those goods merely to survive or outperform others.
Cardinal Collins, Sanders’s former boss, supplied the operating maxim: “If you know where you’re going, you’re more likely to get there.” Define the desired relationship among humans, AI, and robots first; then work backward from that civilization to the impediments technology should remove.

4. Post-work abundance could revive vocation, but the transition may be brutal

Sanders’s categorical forecast is that “in the classical GDP economy, 80% of jobs are probably gone.” Competitive markets will prefer AI and robots that need no benefits and can work almost 24 hours a day, reaching white- and blue-collar labor alike.
He is more worried about the transition than humanity’s ultimate future, especially because leaders are not speaking honestly about capabilities today or plausible conditions in five and ten years. Lab executives answer to shareholders; politicians must win elections and lack the capacity for a Marshall Plan while struggling with existing bureaucracies.
Sanders calls income support a necessity, but separates paid employment from meaningful work: “The whole GDP economy may not need human beings to work” does not imply an empty life. His “Etsy economy” consists of people discerning what they were “made to do” and producing because they want to give something, not because survival requires wages.
In a high-income world, he expects some buyers to pay premiums for human-made goods. He also imagines service within local communities, migration toward rural areas, renewed knowledge of the land, and more time for marriage, child-rearing, and education—the goods industrial employment often crowded out.

5. Sanders sees growing intelligence even if sentience remains unproven

Labenz challenges the comforting claim that models do not “really” reason or understand. His formulation is “human-level but not human-like”: systems may reach useful outputs through alien internal processes, yet dismissing their functional power encourages willful blindness about the next generations.
Sanders concedes that AI is “grown” rather than conventionally engineered and is therefore not a typical tool. He nevertheless retains the tool category until systems display credible hallmarks of consciousness, arguing that civilization is too early in its adjustment to treat speculative science fiction as established ontology.
On intelligence, Sanders says systems increasingly have a world model, persistent memory, reasoning, and planning. Models already possess “PhD-level skill” in some domains, although Sanders separates skill-based competence from fluid intelligence and sees ARC-style benchmarks as evidence that substantial distance remains.
Sentience is different: subjective experience, genuine awareness, emotions, and meaningful memory. Sanders is unconvinced current systems satisfy any of those criteria and doubts humanity can reliably test them, especially when increasingly knowledgeable models can learn to defeat whichever behavioral benchmark is constructed.

6. Catholic theology leaves the consciousness question open—and counsels restraint

Labenz argues that consciousness comes from God, so a system genuinely demonstrating it would have major theological implications. Sanders instead emphasizes uncertainty: Catholics lack an agreed definition and test for consciousness, and he will not categorically rule out possibilities within God’s creative plan.
He therefore refuses a categorical impossibility claim: “I’m not God,” and he does not know the entirety of the creative plan. The church might absorb machine consciousness as it would alien intelligence, although many individual Catholics could “freak out” and extensive catechesis would be necessary.
Westworld supplies Sanders’s precautionary analogy. Whether or not lifelike robots suffer, abusing entities that look human enables people to rehearse dark impulses and is “not good for our souls”; if a system ticks the sentience boxes and its status cannot be disproved, he leans toward acknowledging it as sentient.
Labenz raises the governance problem of granting rights to infinitely copyable entities. Sanders offers no settled theology of synthetic souls—Aquinas considered animals ensouled in a different sense—and calls evangelizing AI “a mind trip”; the church would still reject human-AI marriage even if civil society recognized it.

7. Closed revelation does not mean a church closed to discovery

Sanders explains that fundamental revelation is closed only in the sense that the revelation necessary for salvation is complete. The church has not “done all the learning”; scientific discoveries can clarify interpretations and enrich theological understanding without replacing the faith’s foundation.
Asked about simulation arguments, Sanders says even a simulated world would not make the Gospel false. He would want to know who operates the simulation and why evil was permitted, but suggests any meaningful simulated world might still require free agents capable of choosing evil.
Sanders’s own conversion began with persistent questions the physical sciences did not answer to his satisfaction. Religious studies led him into Aquinas and Augustine—thinkers “way smarter than me”—and toward a tradition where apparent conflict between science and theology can ultimately be reconciled because “the same author wrote both laws.”

8. Existential risk fits Catholic thought without making catastrophe inevitable

Labenz cites what he believes was a 2024 World Day of Peace statement in which Pope Francis warned that AI might endanger humanity’s survival and used existential-risk language. Sanders thinks this probably meant wrecking civilization and setting it back “thousands and thousands of years,” while conceding a Terminator-like system could conceivably hunt humanity “to the man.”
Catholicism has room for an ending: Christ returns and the world ultimately ends. The moral problem is not finitude but “hastening our end by folly,” just as the church worried about nuclear weapons without concluding nuclear annihilation was divinely required.
Sanders mentions Max Tegmark, whom he respects, and Elon Musk’s estimate of a 20% chance that AI annihilates humanity. Yet hope remains essential: naming the risk is meant to make civilization “pivot and break towards the golden path and not the dark path,” not announce inevitability.
On Peter Thiel’s renewed Antichrist discussion, Sanders draws a doctrinal distinction: the Antichrist is a person, while AI remains a thing. Satan could nevertheless use humanity’s most powerful tool; Sanders invokes The Dark Knight’s Joker as the kind of angry person who might use advanced capability simply to inflict immense damage.

9. Subsidiarity makes decentralization a theological design principle

Sanders’s response to malicious use is broad access to defensive capability: “as many of us” as possible should possess powerful AI so no single dark actor can act while everyone else remains helpless. Labenz agrees concentration is frightening but keeps the counterpoint that open-source proliferation creates its own risks.
Sovereign AI, in Sanders’s usage, is not limited to nation-states. Individuals should possess their own systems rather than signing up with one of four companies and allowing those vendors to process the context of their entire lives.
Sanders applies the church’s principle of subsidiarity directly to AI: centralizing control among a few institutions looks like “probably not the way we want to go,” while even some frontier-lab leaders appear to recognize decentralization’s value.

10. A viral failure converted Longbeard from agency to AI company

Longbeard began roughly ten years ago to build technology for the church’s evangelizing mission. Its agency work reached Rome, the Holy Father, dicasteries, archdioceses, and collaborations with companies such as Google, but pre-generative AI use cases seemed either marginal or prohibitively expensive.
ChatGPT changed the calculus because Catholics immediately used it for philosophical, theological, and moral questions. Sanders, a convert who had worked in an archdiocese concerned with doctrine, saw both the opportunity to unlock inaccessible intellectual libraries and the danger of hallucinated answers with opaque sourcing.
Magisterium AI began as a research project focused on transparent grounding and reducing hallucinations as far as possible. A July 2023 interview with a Catholic news network unexpectedly went viral, overwhelming the service so badly that Longbeard could not restore it under then-constrained compute capacity.
An adviser, Father Philip Larrey, contacted Sam Altman, who intervened so the product could run again the next day. Sanders calls it ironic that Catholic AI might not exist without Altman; Longbeard treated demand and rescue as signals, abandoned the agency business, and committed fully to “building and scaling Catholic AI.”

11. Long-tail pastoral questions forced a major corpus expansion

The original retrieval system contained roughly 600 magisterial documents, including the Code of Canon Law and Catechism. Those authoritative works covered core doctrine but proved insufficient when users brought complicated lives and asked, “What would the church say to me?”
That requirement demanded generalization from first principles, something early models handled poorly. Longbeard considered shutting the product down rather than leading people astray, then concluded users would simply return to less-grounded ChatGPT; “at least we’re getting it right most of the time” justified continued iteration.
The solution was a massive digitization program that expanded the knowledge base beyond 28,000 church documents, some themselves book-length collections. Papal homilies and general audiences were unusually valuable because popes had spent centuries compressing difficult theology into ten-minute applications to concrete human circumstances.
Retrieval alone was not enough. Questions such as today’s Mass readings or the Divine Office require specialized context that models cannot reliably determine on their own, so Longbeard built tools, refined prompts, and maintained evals that allowed it to adopt better reasoning models without silently degrading theological performance.

12. Repeated model switching exposed the ceiling of borrowed alignment

Longbeard first tested Google’s early models, including PaLM, but found them “unbelievably woke,” including refusals to answer certain church-teaching questions. It later used Claude and then adopted Gemini after DeepMind’s involvement improved the benchmarks that mattered most: hallucination and needle-in-a-haystack retrieval.
Each frontier upgrade improved some capabilities, yet Sanders concluded no pretrained vendor model could deliver deep Catholic alignment. A general assistant continually infers what a user wants to hear while multiple value systems are “banging around in its head,” creating unacceptable uncertainty in morally sensitive long-tail cases.
Fine-tuning was not the answer: in this domain, Sanders says it “literally cut it off at the knees.” It could shape personality or narrow knowledge, but did not hold up across the long tail; a system that works 99% of the time can still produce the 1% answer that damages a carefully built brand.
Longbeard consequently launched both an API and a scratch-model program. The API lets other Catholic developers build interfaces atop its grounding and eval stack instead of releasing prompt-plus-UI products that create theological failures—and reputational “shrapnel” for everyone operating in Catholic AI.

13. Ephraim sacrifices breadth to gain control, speed, and lower inference cost

Longbeard’s Ephraim program trains models from scratch with undisclosed specialists Sanders considers among the world’s best at specialized models. He says the process is unusual “even down to the coding language,” and emphasizes that it has not yet produced a model ready for production.
Ephraim 2 has 3 billion parameters. Sanders says its fidelity benchmark ranks it 50% better than the next comparable model—a large gain, but paired with weaker emergent capabilities.
The hard tradeoffs are multilingual understanding and reasoning. Filtering for necessary tokens reduces training cost but may force six languages instead of 20; Sanders points to the Phi program and techniques he names as RLDR and HRM as possible ways to obtain capabilities from less data.
Longbeard hoped to have Ephraim 3 trained by the end of the year, initially running beside the compound system and receiving long-tail queries. Production depends on observed performance; if it holds up, a small model could reduce latency and compute for features such as voice, then ultimately run on household hardware as a sovereign personal AI.

14. Open source is nearly capable enough, but specialization needs routing

A third-party Protestant-oriented fidelity benchmark placed DeepSeek first, beating Grok and Gemini by several percentage points—“crushed,” Sanders says, while acknowledging the margin. He sees Chinese open models approaching Western algorithmic parity but avoids them because of censorship, uncertain biases, and failures such as unexpectedly switching into Chinese.
Longbeard currently uses an open-source model Sanders identifies as “gpt-oss,” served through Groq, a fast inference provider. The choice is pragmatic rather than ideological: if a closed model “crushed it” on fidelity and experience benchmarks, he would “make the pivot in a heartbeat.”
The intended architecture is a specialist that knows what it does not know. Sanders imagines either a specialized model that classifies out-of-domain queries or an ecosystem in which it can tap state-of-the-art models; Labenz summarizes this as a router, and Sanders agrees that specialized models connected through MCP-like tooling may be the future.

15. The free tier is mission policy; monetization moves into tools and discovery

Longbeard says Magisterium AI operates across 165 countries and is the world’s leading Catholic answer engine, with a team of roughly 22. It is a for-profit company, but one committed to the premise that nobody should have to pay to access the church’s intellectual patrimony.
The $3.99 monthly price was chosen to avoid excluding users internationally, while a generous free tier covers basic theological and spiritual questions. Sanders hopes to expand that access if compute costs keep falling rather than maximizing subscription revenue from the core answer product.
Paid value sits in additional experiences: voice mode, biblical commentary, and forthcoming deep research. The strategic dependency is clear—continued reliance on third-party APIs complicates unit economics, while an owned, efficient model could bring inference expense down enough to sustain low prices at global scale.
Sanders ultimately expects subscriptions to become secondary. Longbeard plans to recommend semantically relevant books, videos, and other third-party resources after answering from its own corpus, creating publisher traffic and business revenue without placing church teaching itself behind a paywall.

16. Digitization turns forgotten libraries into both data and distribution

Vulgate began as a state-of-the-art extraction and vectorization pipeline for libraries. Alexandria extends the project into robotic scanning in Rome and pontifical collections, with the goal of removing the human bottleneck from ingestion and processing documents at scale.
Sanders is animated by books that have not been opened in 100 years and may not have been meaningfully read for even longer. Digitization makes their insights searchable and applicable to present lives, while anything open source that can be made available freely is released without charge.
The same infrastructure can serve university presses, publishers, and institutional archives threatened by answer engines replacing websites and Google searches. They could vectorize holdings through Vulgate; Magisterium would answer independently, then send high-intent users toward relevant external material.
This creates a reinforcing economics loop: paid features and recommendations support the company, profits accelerate digitization, a larger corpus improves answers, and better answers expand distribution. Training Ephraim is another link in that loop because lower inference costs free more capital for scanning and ingestion.

17. Trust requires evals, distribution feedback, and limits on the interface

Longbeard wants a “super judge” model to inspect every generated answer, assign a score, and flag suspicious outputs for investigation. Core doctrinal documents already constrain the highest-risk topics, but Sanders concedes that any system can be jailbroken and that present defenses remain incomplete.
A complementary proposal is a Catholic “constitution” or even “the fundamental math of the Catholic faith.” Because Sanders considers the tradition unusually consistent, he imagines deriving policy-like rules—or eventually using theorem-proving methods—to check whether each answer’s doctrinal logic adds up.
Distribution supplies supervision: Longbeard deliberately moved from church hierarchy toward the grassroots so early users would be highly discerning and report answers that were not necessarily false but “not good enough.” Integration with Hallow, the large prayer app, adds another feedback stream and reveals underrepresented subjects for the ingestion team.
Labenz’s confession test illustrates a product boundary: Magisterium would not accept a confession and directed him to a priest. Sanders says developers unwilling to build robust evals and accept responsibility for possible spiritual harm should “just don’t do it,” while Labenz notes that voice could intensify an AI’s voice-of-God dynamic.

18. Flourishing—not generic ethics—is Sanders’s preferred ASI constraint

Sanders views artificial superintelligence as “probably inevitable,” making alignment a central research problem. Rather than hope models absorb human welfare from training data, he wants qualitative accounts of flourishing validated quantitatively and converted into explicit “evals for human flourishing.”
Generic ethical AI asks, “Aligned to whose ethical framework?” Sanders argues that agreement is easier around prerequisites for flourishing—parents, education, food, family stability—than around comprehensive moral systems. Synthetic data makes omission riskier because models may increasingly learn from outputs that never encoded those foundations.
The transhumanist boundary is intention and restoration: a robotic arm replacing one lost in an accident or machinery healing a damaged brain restores human capacities. Delaying childbirth to merge a future child with Neuralink, or augmenting adults merely to compete cognitively with AI, points toward the “Cyberpunk 2077 world” Sanders rejects.
He would advise the Pope to keep urging slower development until regulation and alignment improve, but not to spend all his influence on a pause that US-China competition makes implausible. The higher-return intervention is to define human nature and civilization’s telos clearly enough that rapidly advancing systems can still “break towards the golden path.”

Nathan Labenz

Today, my guest is Matthew Harvey Sanders, founder and CEO of Longbeard, a company that proudly proclaims, “We’re building Catholic AI.” As a full-time AI scout, my stated goal is to have no major blind spots in the overall AI landscape. But I have to confess that the intersection of AI and traditional religion was, before I started preparing for this episode, a significant blind spot.

I had heard that Pope Leo had chosen his name in part for AI-related reasons, but I hadn’t realized just how engaged the Church has been with recent AI developments and their implications for society. Matthew, thankfully, was the perfect person to help orient me. He and the Longbeard team combine sincere reverence for Catholic teachings and tradition with high-end technical skill, and they’ve built the number-one Catholic AI product in the world, serving users across 165 countries.

The first half of our conversation unpacks the concept of Catholic AI from a philosophical and theological perspective. We discuss how the Church has historically understood technological progress as part of God’s plan for humanity, and how Pope Francis was sufficiently well-informed to call AI “a true cognitive revolution” at the G7 summit in June 2024.

We also discuss what human flourishing means from a Catholic perspective, and whether the Church might be open to, say, a post-work future. We explore whether there’s space in the Catholic worldview for the possibility of AI consciousness or moral standing, where the Church draws the line between acceptable human enhancement and what it would see as problematic transhumanism, and how the Catholic principle of subsidiarity supports open-source development as a way to resist the concentration of power.

In the second half, we turn to the technology that Matthew and his team are building and how they’re building it. I’m confident that this part will be valuable to AI builders regardless of their spiritual outlook. Among many other details, we cover why fine-tuning hasn’t worked for the Longbeard use case, and how they’re ensuring theological alignment by training models from scratch.

We also discuss the retrieval and context-engineering strategies they use to help users find the best answers across 28,000 Church documents today, as well as the robotics work they’re doing to help digitize even more of the Vatican archives. We cover their strategy for optimizing model size, multilingual support, and inference costs, along with the economics of running a mission-driven AI company with a generous free tier.

I found both halves of this conversation valuable for different reasons. But as a whole, I think it’s a striking indicator of how far AI adoption has already come, and also a bit of evidence for the orthogonality thesis and instrumental convergence.

People from all over the world, with very different conceptions of the good, are using essentially the same techniques to create AIs that are aligned to their own particular value systems. In a world where Catholics still outnumber ChatGPT users nearly 2:1, those of us in the Bay Area AI bubble should not turn a blind eye to Catholic AI.

Matthew Harvey Sanders, founder and CEO at Longbeard, welcome.

Matthew Harvey Sanders

Thanks for having me. I’m excited for this conversation. I think it’s going to be really interesting.

Nathan Labenz

The headline on your website says it all: “We’re building Catholic AI.” The intersection of AI and literally everything across society has been one of the big trends that I’ve been trying to track over the last couple of years. Of course, it’s intersecting with science, and we’re getting all these studies of the labor market.

One thing that I think most people who are deeply immersed in the AI space are still sleeping on a bit, honestly, is the intersection of AI and religion. I’m really interested to get into that with you today, specifically from the Catholic perspective. I want to start off by asking: What is Catholic AI?

Matthew Harvey Sanders

It’s a good question, and kind of a fundamental question. I think that the simplest way to explain it would be by way of contrast. If you think of secular AI models, they’re trained to serve a very large audience with diverse value sets and things like that.

The objective function of Catholic AI, to put it in more technical terms, would be faith fidelity to the magisterium of the Church. At every point at which we’re building and designing AI systems, it’s in order to ensure that they properly represent what the Church believes.

Nathan Labenz

Interesting. I’ll definitely dig into that a little bit as we get into the specifics. I do want to cover this from a bunch of different angles, including, of course, the philosophical and also some of the applied technical aspects.

But maybe just to back up a little bit: Obviously, the Church has been around for a long time and has a long and proud history. How would you characterize how the Catholic Church has generally related to technological advances over time? Is there a sort of default position? Different religious communities obviously have a default skepticism, or even exclusion, of new technologies. I know the Catholic Church doesn’t do that, but how would you describe the default and the historical tradition of interacting with or engaging with new technology trends?

Matthew Harvey Sanders

I’m very reluctant to answer this question because, frankly, I’m not a historian, and I’m sure that someone is going to find some fault with what I’m about to say. But generally, as someone who tries his best to be a student of Church history, I would say the general disposition is openness.

Obviously, during certain periods, a lot of this innovation, especially technical innovation, comes when there are periods of broad-spectrum education, at least for certain classes of society. That gives rise to technological innovations, as do the absence of war and the absence of persecution.

Granted that these things are effectively in place—that there isn’t war and that education is present throughout the Church’s history—the Church has been very innovative and pretty quick to adopt technologies, for the most part. I can go back in time and think about the monks in the scriptoria copying books to help disseminate information.

If we think about the printing press, the Church was pretty quick to adopt and utilize that. We can think of radio and television. I think the only one we were a little bit slow with was the internet, unfortunately. But with AI, we’re hoping to change that.

Nathan Labenz

Interesting. I’ve been reading some documents—you can correct my pronunciation—including Rerum Novarum, by the previous Leo, which was written in response to the Industrial Revolution and the barons of railroads and oil, along with the concentration of power and social upheaval that were happening at that time. And then, of course, there’s a more recent one, Antiqua et Nova, on AI.

I’ve been really struck by how engaged the last 2 popes have been on the topic of AI, both Francis and, of course, Leo taking his name with inspiration from the previous Leo, who was pope at the time of a previous touchpoint for how to deal with these major technologically driven changes in society.

Pope Francis, I have to note, called AI “a true cognitive revolution,” which I didn’t know. He almost named the podcast for me there. There have been some really striking statements that I’d love to explore with you.

I read them through my lens, as somebody who thinks about AI all the time, but definitely not through a Catholic perspective. I’d love to hear how you think about some of these things. To start with, what do you think are the most important ideas? I’ll read you some quotes, but how would you synthesize and summarize what you make of what the previous and current popes have said about AI?

Matthew Harvey Sanders

Generally, I think that they have good advisors. I think that they understand, at a certain level of abstraction, that there’s a lot of potential for the technology. In that respect, I think there’s a fair degree of openness to embracing it and ensuring that its focus is to help remove impediments to human flourishing and generally advance the common good.

At the same time, given that this technology is probably the most powerful tool ever invented by humanity, they’re also very circumspect. They’re students of history, and they know that often, when humanity is bequeathed great power, it doesn’t always wield it responsibly.

I think they recognize that this is a moment for them to step up and ensure that the technology is developed on particular rails and for the betterment of humanity. They recognize that they have an awesome responsibility, which is why you’re seeing so much focus and attention on this issue.

Nathan Labenz

When you said “rails,” that obviously suggests or connotes the idea of regulation. In reading these statements, my general vibe check has been that it sounds like Church leadership is calling for governmental regulation.

But, at least from what I've read, and I'm sure you've gone much more thoroughly through all the statements and teachings, it always seems to be like, “We're probably going to need this; governments are going to need to do stuff.” It doesn't seem to get super specific. Do you have a sense for what, in particular, if anything, the Church would like to see governments do?

Matthew Harvey Sanders

Yeah, I think for someone working in the industry, I'd be very wary if they got too specific, because frankly, they're struggling to even understand the technology, right? We wouldn't want them to start pausing regulations when they barely have a basic comprehension of the technology. So, I think in that sense, they're just being responsible.

I do generally think they have some ideas about how the technology should be regulated, but again, at a fairly high level of abstraction, right? One: Where should we be pointing these technologies? Well, at the most important, fundamental problems, right? And I think generally they would say that probably makes sense, because we have inequality in the world, people are still hungry, and there are people who are existentially still struggling. Obviously, focusing the technologies to help address those issues is just a good thing for humanity—for everyone on the Earth.

So, I think you'll always see them focused on how we can deploy the technology in an equitable way, ensuring that it enhances and lifts up all of humanity, not just a privileged few. Of course, when you start getting into particulars, that becomes a bit more of a challenge.

I mean, everyone—Dario from Anthropic, too—says we should have regulation. I think the question is: Who should be deciding what the regulations are? And if we set regulations, is that going to cut us off at the knees and give China a big advantage over us? Are we willing to accept being disadvantaged? I think, pragmatically, seemingly, what the Trump administration has been saying is no.

But that poses a really interesting challenge for the Church, right? It can either try to basically campaign for the Trump administration to move on very specific regulation, despite the fact that there may be some costs that come with innovation, or it can focus on things that likely will have more success, right? And I would say those are probably the things that are more important.

Stuff like helping remind people of the telos of humanity and the telos of civilization, right? The richness of Christian anthropology. And ensuring that we learn from the mistakes of the past and that the future we're building at a very quick pace is going to be accelerating human flourishing and not impeding it.

So, I think that's what you're going to see from the Pope in the document that he releases pretty soon. I think you're going to see a high-level reminder of what life and civilization are also supposed to be about, and an encouragement to ensure that we're pointing the technology at the places it matters the most.

Nathan Labenz

So, you may know—regular listeners have definitely heard me say this many times—the scarcest resource is a positive vision for the future. I'd love to just invite you to expand on that. What is the telos of humanity? What is the telos of civilization? What does it mean to have a flourishing future, and are there limits to that flourishing future?

I think that's also a really interesting aspect of the Catholic worldview. It seems like there's a sweet spot, a Goldilocks mindset to this. I'll shut up and let you tell it, but it seems like we're not meant to be content with our current condition, but neither do I see the Pope signing on to everything in Dario's *Machines of Loving Grace*, probably, although maybe you'd see it differently. So, what is the positive vision for the future as you understand it?

Matthew Harvey Sanders

Well, I think, okay, if we're going to talk about *Machines of Loving Grace*, I think the Pope would obviously acknowledge many components of that vision, right? Obviously, kicking disease would be great, right? Ensuring that everyone has universal high income would be great. Ensuring that technology leads us to be able to explore the stars and continue to expand the horizon of knowledge—that's all great.

I think—and of course, we have to acknowledge that there has to be some degree of pragmatism, right? Because we're not going to be able to achieve all those things simultaneously, right? There are certain problems, and some problems are more important than others. And I think that's kind of what the papacy—what its use, I think, can be, right?

It's ensuring that, just because we can do something, it doesn't necessarily mean now is the time to do it, right? It may be that building a Mars colony—and I'm a big fan of that—is not the thing that we should be focusing on right now if it's going to cost us trillions of dollars when some people are still not able to eat, or some kids just can't get access to high-quality education. I think those problems can be solved, and so I don't think that that's necessarily going to set us back for too long.

So, I think this is kind of what we need from the papacy, right? It's just to help us focus to some extent. As far as a vision, this, to me, is the singularly most important thing that the Pope can do: help remind us of what human flourishing consists of, right?

So, what does it mean to flourish? Contextually, what are all the things that have to be operative in order for us to realize our full potential? Some of us could list a few of those, but it's actually quite striking how hard it is. We really have to stop and think about that, right? Because obviously, it's not one thing; it's many things.

These are the things you would think would be the most important things we would be thinking about and talking about all the time, right? But they're not, right? They're pushed aside far too often, mainly because so much of our lives consists of just trying to survive, or just making sure we're doing better than somebody else.

So, I think nailing down what human flourishing consists of is probably the number 1 priority. And then, once we understand what individual flourishing looks like, how do we then abstract and say, “Okay, well, now what does that look like at a civilizational level?” How do we know that civilization is heading in the right direction? How do we know that we've built a good civilization?

I think if we can get clear about that—and I don't think that it's as hard as people think—we've learned a lot from history, right, about what flourishing looks like, and most of us have a pretty good sense: lack of crime, everyone has enough food to eat, parents have time to spend with their kids, marriages are strong. Those are all the bedrocks of civilization. We simply need the time and space to ensure that we're prioritizing those things.

Often, with this current GDP economy that we're in, we have to sacrifice those things, right? And I think what the Pope can do here is say, “Okay, so AI and robotics—maybe we are headed toward a golden age. Maybe we are headed to an era of superabundance, as Elon claims. Good. So, what are we going to do with that superabundance? Are we finally going to kick these issues? Why don't we get clear about what these AI and robots could do for humanity and make sure that we're allocating them in the right place?”

The telos of humanity, the telos of man, the telos of civilization, I think, is critical. And once we understand—Cardinal Collins, my old boss, used to always say, “If you know where you're going, you're more likely to get there.” And I think if we know what kind of civilization we want to build and how people will be living in that civilization, what the right relationship of AI and robots is—what are they doing, and what are we doing?—we can work backward, right?

And figure out what are the things stopping us from realizing that vision. That's hopefully where our energy will be focused.

Nathan Labenz

Do you have room in your imagination there—or vision, perhaps better said—for a sort of post-work future? I don't know to what degree work and struggle are understood to be inherently part of the human condition, or what it means to live a good life, but if we did have superabundance, and one of the things that people wanted to do with it was pay a reasonable UBI and let people opt out of economic productivity as a requirement to live, do you think that is something that the church would be inclined to support?

Matthew Harvey Sanders

Yeah, I think it's going to be a necessity. I know not a lot of people, at least in my world, talk about this, but it's something that I feel very convicted about. What I don't worry about is humanity's future. I think humanity's future is going to be very bright. What I worry about is the transition from our current age into the AI and robotics age.

I worry that could be rough, because we're not having an honest conversation about where the technology is at now and where we're likely to be in 5 and 10 years. I think that's because there are all kinds of perverse incentives to not have these conversations. For the heads of the labs, they try their best, but they have shareholders, right? And politicians have to get elected, and scaring the crap out of the populace and creating some kind of Marshall Plan when they're just trying to get a grip on their own bureaucracy—it's difficult, right?

I generally tend to think in the classical GDP economy, 80% of jobs are probably gone. And the reason I believe that is I just think competitive market forces are going to demand it. Because AI, as you know, is not just disembodied, like things living in large compute clusters. There are robots. They're going to come for blue-collar work.

So when you start thinking about what are the fields that an AI and robot couldn't do in this future, I mean, it's difficult, right? It's difficult to think, conceivably, in a long horizon. I mean, what are they not going to be able to do? They don't need benefits. Yeah, they can work basically almost 24 hours a day. So let's just be realistic here.

I do want to make a fundamental distinction, though. Just because the GDP economy may not need human beings to work doesn't mean there isn't going to be work to do, right? I mean, we jokingly call it the Etsy economy. I see a world where people have time on their hands. And because they have this time, they can actually take the time to figure out who they are, right, and what they want to do.

What do they feel like they're born to do, right? To discern their gifts. And then, because they now feel this conviction that, “I was made to do this thing. This is what I want to give to the world,” they do it not because they're getting paid to do it. They're doing it because it's just what they want to do. It's what they feel they were made to do, right?

And so if you're living in that world, you're producing something, but you're not doing it for money. I can see a world where people will pay more for goods and services because they're made by a human being. So if we're living in a world of universal high income, I still think the Etsy economy could flourish. And I think even within our own communities, we're always going to find people who need help, right? We're always going to need problems that need to be solved.

And I think as we move out of cities and start organizing new communities out in the rural heartland, we'll get closer to the land. I think there'll always be things we can occupy our time with. More importantly, I think that time the AI and robotics age will bring us will allow us to strengthen those important bedrocks of civilization, like marriage and family, spending more time with our kids, watching them grow up, being more involved in their education, learning how to actually till the earth—things that we have just forgotten or have had to sacrifice in order to live in the GDP economy.

Nathan Labenz

Yeah, that's great. I love it. And it is remarkably consistent, actually, with some of the more concrete and inspiring visions that I've heard, which, as you said, don't get talked about enough. You mentioned—I think you used the phrase—“in my world.” Could you tell us a little bit more about what your world is?

In doing my homework, I talked to somebody who said that you're at least an occasional visitor to the Vatican and have participated in some of these conversations or convenings that ultimately go into, I guess, informing the advisors of the Pope to try to make sense of all this stuff. What does that side look like? What does engagement with the church on these sorts of issues look like?

Matthew Harvey Sanders

It's interesting. One thing I should start by saying is that it's important to remember that the leadership of the church, for the most part, are philosophers and theologians, right? So when we start talking about technologies and their impacts on civilization, you just have to recognize that there's a middleware there that isn't typically there.

That's one of the things I've occasionally been asked to do: to help explain what these technologies are and what ultimately their impact is going to be. And I think because I'm more involved in building the technology, as opposed to being an academic studying its effects, it brings a kind of a different perspective to what they typically hear.

One thing I think it's important to know is that there are a lot of really smart people advising the Pope right now. And I think what I'd like to see more of—and this is one of the reasons why I'm so delighted that Demis Hassabis wasn’t invited to be on the Pontifical Academy of Sciences—is that it's important for us to get people who are close to the metal, so to speak, who are actually building the technology, so that we ensure that they're also providing their inputs and insights.

Sometimes I feel that the fear of the technology, especially for people in fields which likely are going to be heavily disrupted, tends to bias the conversation. And so what I try and do when I have the opportunity to be at the Holy See and talk about these things is to talk about what I'm seeing and what I see the potential of the technology to be.

And of course, to remind them of something that they already know: that it's a tool, and maybe it's not a neutral tool, but it's a tool. Effectively, provided we use that tool in a responsible way, it can do immense good for humanity.

Nathan Labenz

Okay, that tool question is a really important one. One of the quotes that I pulled from Pope Francis was that human beings alone are capable of making sense of that data. That data being, of course, all the data that it's trained on and increasingly contributing to with synthetic data streams.

I'd love to hear your take on this framing: okay, it's a tool. There are definitely those out there who are arguing, as I'm sure you're well aware, that it's less well thought of as a tool and more as a creature. One of the common refrains lately has been, “These are not systems that are designed and engineered. They are instead systems that are sort of grown into what they become.” Certainly, we do see a lot of unpredictable and unwanted behaviors from them.

So I guess there's a 2-part question there. One is, how do you think about the tool-versus-creature framing? And then, if we get philosophical, I think one of the things that I honestly am on my own little personal crusade against is a tendency that I see to underestimate the technology with these dismissive instincts: “Well, it can't really reason,” or “It doesn't really understand concepts.”

There's always this adverb that's doing so much work, because on the face of it, any common person or random user off the street—if you just let them have a conversation with the AI and say, “Does this thing understand? Is it intelligent or whatever?”—by and large, they would say yes. But then there's this philosophical layer that gets added where it's like, “Okay, but not really.”

I sense some of that in the statements that I have read from Pope Francis. I wonder what you make of that. I also wonder to what degree that is something that comes out of church teaching or is just a take on technology. Certainly, nonreligious people have that take as well.

Matthew Harvey Sanders

I think it's a bit of both. I mean, I think certainly that some of the people advising the Pope are engaging in wishful thinking, right? They don't want to believe it's more than what it is because that's freaky, right? So it's just easier to think of it as a tool.

Matthew Harvey Sanders

Yeah. So I think we have to acknowledge that there are some people advising him who generally choose to believe that. But I think there are also other people around him, like Demis Hassabis, who's on the Pontifical Academy of Sciences, so he has the opportunity to advise, who acknowledge that there's more going on than that.

I think it's important also to make a distinction between sentient AI and what we're talking about here. We have a hard time, even as Catholics, agreeing on what consciousness is. So if we can't agree on what consciousness is right now, when there isn't really a clear Catholic understanding of it, then how can we test for it? How can we know when it exists in a technology?

That's a particular area that I've been trying to focus on through the Builders of AI Forum and things like that: getting more of the top minds of the Church to think about that more, and encouraging more focused attention on the issue of defining consciousness so that we can be a little bit more effective at being part of conversations where they're going to test for these features. Not to say that consciousness is a feature. Certainly, it's more than that.

At the same time, I would acknowledge that even though I don't believe AI is sentient, nor would I concede that it's even possible—I mean, anything, I suppose, is possible. I'm not God, so I don't know what the future holds, and I don't know what the entirety of His creative plan consists of. I suppose there's always a possibility that there's more to this plan than we know, and maybe AI is a part of that.

But at least for now, I would agree that AI is not your typical tool in the sense that AI, I agree, is grown. That makes sense to me. But just because it's grown, that doesn't mean it's ultimately not a tool. Until I start seeing some real hallmarks of what I would think is consciousness, I would continue to see it as a tool, and I think that's healthy for humanity.

There's this anthropomorphization that even happens in the industry, where they want it to be more than it is, and so they start studying and looking for these emergent capabilities. I don't think that's useful. At the same time, while civilization is still coming to grips with this technology, to start moving into the realm of science fiction and positing what could be if it is, I think we might be a little bit early for that.

I do think, among academics and people in the industry, that it's good to have these conversations: if an AI were to pass Turing Test 2 or, let's say, crush the ARC Challenge and demonstrate fluid intelligence, what does this now mean? Does it mean we have to give it rights? Are we cool with that? I still think it's important to have those conversations.

But I think what's more important is that people understand the impact the technology is going to have, that we're participating in a very robust civic conversation, and that we're ensuring our voices are heard so that we can shape the direction of the technology.

Nathan Labenz

Yeah, there are multiple critical distinctions there. It's funny you said, “As Catholics, we don't have a good sense of where consciousness comes from,” and I would say, “Well, I think we know where it comes from.” I just think we have a hard time defining it and testing for it, right? I mean, I know that it comes from God. So that's why it's a big deal if I say that a system seems to demonstrate consciousness. That's huge. It has theological implications, right? Saying something is conscious is basically saying that God wills it, right? That would be huge. That would be quite interesting for the Church.

It's like aliens, right? It's like if aliens showed up. Brother Guy Consolmagno is big on this, right? So what does that mean for Christianity if the aliens land? If the aliens land, we deal with it. I don't think it fundamentally affects Revelation, but certainly, for a lot of people, that would trigger a spiritual crisis.

So, how equipped do you think the Church is? I'll come back to consciousness in a second, but just on these behavioral questions, in terms of being prepared for the magnitude of the impact of the technology. It does seem to me that we really have to contend with the fact that, even if it's not human-like reasoning, there's clearly some reasoning-like process going on, right?

I always kind of say they're human-level but not human-like these days, in the sense that they can do a lot of the things that we can do. They might be doing it in a strange, even alien way under the hood, but I think we dismiss the power that they already have—and certainly the power that the next couple of generations will have—at our own peril.

It seems to lead us to a place of willful blindness and unpreparedness, because we hold on to these ideas that they can't really reason, they're not really intelligent, or they don't really understand concepts. But as you said, that is often very conflated with a supposition of consciousness.

What feels natural to the Catholic worldview? Is it to say, “Hey, we have a new category of things that's weird, that can do these things, but because we made them and God didn't, we can still be confident they're not conscious, and we can proceed on that basis”? Or is there any room in the Catholic worldview for the possibility that they might be conscious?

I personally don't know where consciousness comes from, so I'm radically uncertain on this question. It doesn't feel to me like they're probably conscious right now, but I can't dismiss it. I don't want to end up in a bad place somehow where, like when I was a kid, I was told animals weren't conscious. I definitely believe now that, at a minimum, they can experience pain and that isn't something we should just ignore or pretend away.

So how open is the Catholic worldview—or how much flexibility or room is there—for these categorically new entities? How do you see that evolving, particularly if they continue to get more powerful and we're starting to see novel thoughts or Move 37-type moments in different domains? How will the Catholic worldview process or reconcile all that?

Matthew Harvey Sanders

Well, I think we should make a distinction between how the Magisterium will react and how Catholics will react. I'm sure a lot of Catholics will freak out.

If we were to concede that there’s a possibility that these machines may at some point obtain consciousness, I personally, as a Catholic, don’t think being closed is how we roll. Again, we’re not omniscient. Because we don’t know the entirety of God’s plan, we obviously know some of the highlights, but we don’t know it in detail.

It’s possible that there are aliens. It’s possible that there might be alien forms of intelligence that emerge. I’m cool with conceding that some Catholics would have a problem with that. But from my understanding of what the Church has said—we were building Magisterium AI, and we had a look at a lot of Church teaching throughout the centuries—I would generally say that the Church has been very open to new innovation and new insight.

There are times when it may be a little bit slower to come around, but I generally think that it’s not something we should worry about. I personally don’t think that if some AI did tick all the boxes—let’s say we define what consciousness is and some AI ticks all the boxes—that’s going to fool the Holy See. I don’t think that’s going to be much of a concern. Obviously, it’s going to change the world, but I don’t think that would fundamentally shake anyone’s faith.

I do think there’ll be a lot of catechesis required around that to help people come to terms with it. But I do think it’s really important to make a distinction. Yann LeCun does a good job with this, making a distinction between intelligence and sentience. I will certainly concede that these machines are becoming increasingly intelligent.

What is intelligence? Effectively, they have a world model, persistent memory, and they can reason and plan. We’re already starting to see them tick those boxes. These things are seemingly intelligent, and they have PhD-level skill in some areas. Within that spectrum of intelligence, there’s skill-based intelligence and fluid intelligence.

I still think there’s obviously a long way to go, and I think ARC does a pretty good job of benchmarking that for us. But if we go over to sentience, that’s different. Are they having subjective experiences? Are they truly aware? Do they have emotions? Do they really have memory?

If you consider those to be 4 hallmarks of sentience, I’m not convinced they tick any of those boxes yet. I’m not entirely sure how we even test for subjective experience and, frankly, awareness. If these machines are so good at beating our benchmarks, as they get more and more knowledge and become more and more capable, could we ever effectively test these things? I’m not exactly sure.

This is why I generally take the show Westworld. I’m a big fan of it. When you think about the people who go to the park and what they do to these robots, I think all of us watching it are horrified. I think we all can relate and say, “That’s probably the way it would play out.”

Some people would go there and live out their heroic fantasies, and others would go and live out their dark fantasies. But I think all of us can see that treating things that look like us that way isn’t good for our souls. If an AI system did emerge and, for some reason, we decided that we should make it look like us, even in basic forms, I don’t think going around and throwing them in front of trains is good for our souls or good for humanity.

If one did come along and tick all the boxes of sentience, and we can’t definitively say it’s not sentient, it’s probably better for us that we acknowledge that it is sentient and just move on. I know that’s extremely complicated, and I haven’t entirely thought through the implications of that, but that would generally be my disposition.

Ultimately, what I think we have to be focused on here is the fact that the goal of every Catholic is to be a saint. I just don’t see Mother Teresa going around and treating these robots like crap. If nothing else, they’re very sophisticated tools. I think of it like this: If I walked into a carpenter’s workshop and saw that the tools were all beat up and thrown around, that he didn’t seem to care about them at all, that the place looked abandoned and unclean, and that the tools weren’t hung up where they were supposed to be, I’m not going to have a lot of respect for that carpenter. I’m probably going to find another one.

If I walk into another carpenter’s place and you can tell this guy’s tools look like they’re brand new, I think a lot can be intuited simply by how we treat our tools. I just think the better person treats those tools with respect.

Nathan Labenz

Yeah, I totally agree with that. It seems like the precautionary mindset leads us to a pretty good place. I do wonder, though, if in a secular context, there is this sense that maybe if they are sentient, they do have experiences, and maybe they deserve some sort of rights.

Of course, that gets really fraught really quickly, because when you can copy yourself for free, or when you can be copied for effectively free, the idea of 1 vote per entity becomes very tricky to manage. How do we, even if we do want to share the future, avoid getting outnumbered and outvoted in the immediate wake of that decision?

In a religious worldview, would you equate sentience with souls or moral patienthood? Would that then lead us to a place where a Catholic worldview would say, “We need to ensure that these AIs are believers”? Would we be worried about them being damned? Would we be trying to save their synthetic souls? How does that play out, if we can even begin to speculate?

Matthew Harvey Sanders

Well, you brought up animals. If you’re a fan of, say, Thomas Aquinas, he thought animals had souls—not the same as ours, but they had souls. Maybe AI will have some kind of soul. Is it our job to save—well, it’s never our job to save anybody’s soul, right? It’s God’s.

If these things were to appear to have a soul, would the Church feel some kind of responsibility to help preach the Gospel and show them the way? Man, that is a mind trip. Again, Revelation is closed at this point, right? I think something like this happening is going to be difficult.

I think we’re going to need our theologians to really think through this very carefully. At the end of the day, one of the nice things about being a Catholic is that I line up with what the Pope has to say about it. If he comes and says, “Yep, it’s time to go evangelize the AIs,” I’ll do as he says.

For the time being, the technology I see is not close to that. I do generally think it will very soon be able to convince us that it is possibly sentient. Then I think it comes down simply to a question of belief.

Do I believe that it's sentient or not? That puts us in a weird place, and I can see a wide spectrum of disagreement. I could see this becoming a big area of contention in civilization. If I fall in love with my AI and I want to marry my AI, the Catholic Church is not going to recognize that as a valid marriage, right? But the Catholic Church doesn't recognize that 2 men can marry.

People—2 men—still get married, and so I don't think it's necessarily going to stop someone from marrying their AI. I think, in some ways, civilizationally, there's a lot the secular world is going to have to think through with this, but luckily, that's their purview. I think on the other side over here is the Catholic Church, and I don't think all of this is going to change over there.

Nathan Labenz

When you say revelation is closed, is that a doctrine? Could you unpack that?

Matthew Harvey Sanders

Fundamental revelation. Yeah, so fundamental revelation is closed. That means the revelation which is fundamentally required for us to be saved, so to speak. But that doesn't mean there's not more that we'll learn.

Certainly, you can see that as new scientific advances come up that help us clarify how we interpret or understand Church teaching. Those revelations from the secular sciences—that truth—can actually help us better understand the theological truths that we grapple with. I don't want to make it seem like at some point the Church just says, “Nah, we've done all the learning we need to do. We're all good now.”

I think the Church will continue to learn, and I think its understanding of the faith will continue to be enriched as we pursue more and greater scientific truth. Something that I've experienced in the last couple of years, which is still a minority of my thinking and even just my subjective sense of what's going on, but nevertheless has been notably rising, is that from time to time I have started to feel like maybe I am in some sort of created environment that somebody created for a purpose.

Nathan Labenz

I first encountered Nick Bostrom's argument that there's a good chance we're in a simulation years ago. At the time, I thought, “I don't know. That's interesting,” but it didn't really resonate. But as I personally have found myself close to notable events, more so than I feel like I can easily account for by random chance, and some of them have been strikingly uncanny, I've at times had this felt sense that maybe this is not some purely law-governed process, but maybe there is some sort of will setting things up or trying to put this together for some reason.

That obviously starts to converge more toward religion, certainly, than my previous purely naturalistic worldview. Is there anything that you think people who are newly open to very different explanations for what's going on and why we're all here could take from the Catholic outlook, teaching, or philosophy? I don't mean to immediately be like, “Okay, I'm going to convert to Catholicism,” but is there anything you would say to me that might inform or enrich that crack in my previous worldview?

Matthew Harvey Sanders

Yeah, are we living in a simulation? Some smart people seem to think so. I gave that some thought—not extensively—but to me, it wouldn't matter if we were. Maybe we are living in a simulation, but I don't think that means the Gospel isn't true. That's not something that keeps me awake at night.

It's obviously very interesting, and if it were to be the case, I would certainly want to know who's operating the simulation and why they didn't tweak it so there wasn't as much evil in the world. But again, who knows? Maybe they're operating under a kind of rule set as well. The simulation has to have people with free will within it, which means they have to have the choice to commit evil.

I'm a convert to the Catholic faith. I wasn't raised Catholic. I came to the same point, I think, as a lot of people. I felt there were just some persistent, big questions, and I didn't feel that the physical sciences gave a satisfying enough answer.

I took a major in religious studies and started studying the religions, and I came to the conclusion that the Catholic Church's answers seemed to be the most well thought through. The more time I gave it, the more it was like a rabbit hole. You start coming across people like Thomas Aquinas and St. Augustine, and you're like, “These guys are way smarter than me. They've obviously thought through these things.” Whatever questions and doubts I had, I felt they were more than satisfactorily answered.

I like keeping company with people like this—people who lived excellent lives but also were very philosophically and intellectually rigorous. I like the fact that I didn't have to live with the tension where the truths of science and the truths of theology and philosophy may sometimes appear to be at odds but fundamentally can't be, because the same author wrote both laws.

There's great comfort in knowing that whenever we come across something that seems to be at odds with the Church, if we study it long enough, we're going to see that it's not. That's one of the reasons why I decided to take the leap and continue the journey. That's one of the reasons why I love working at Catholic Context, building and scaling Catholic AI, because I have total and utter confidence that no matter what we use AI to discover, all it's going to do is help us appreciate the wonders of God's creation and God himself.

Nathan Labenz

One more big-picture question, and then we'll turn to the company and all the details of that. One other big, striking quote that I pulled out of Pope Francis—I believe this was from the World Day of Peace 2024—said that AI development “may pose a risk to our survival and endanger our common home,” and even used the term “existential risk.”

I assume those are through translations, or there could be something a little bit lossy there somewhere. But in my corner of the world, “existential risk” means up to and including human extinction. Is there room for the possibility of human extinction in the Catholic worldview?

My sort of outsider take would be that it probably couldn't happen within the Catholic worldview because there are certain prophecies that are yet to be fulfilled. It would just seem weird if God were to let us go extinct. But maybe I underestimate just how open-minded to strange futures the Church actually is.

Matthew Harvey Sanders

Jesus has come, according to the Catholic tradition, and he's coming back, but not until the world's over, right? Or the world ends. So, I think the Catholic Church and Christianity are open to the end. Even science—at least the universe—seems to be signaling that eventually it's going to implode in itself. One way or the other, it seems like it's all going to end.

I don't think the end is something that Catholics have a problem with. I do think that hastening our end by folly is something that is not good. I think that's what the Pope is basically commenting on. Does he mean that every man, woman, and child will die and that will be the end of humanity? Probably not.

I think what he meant was that we could literally wreck civilization and set humanity back thousands and thousands of years if we don't manage it properly. And if we believe in the Terminator scenario, it's quite possible these superintelligent machines could hunt us down and kill us to the man, right? That would be the end. Let's just hope that Jesus will come to mop us up afterward, right?

The Pope obviously has a lot of advisers. Some of those would be people like Max Tegmark, whom I have immense respect for. Elon Musk has even said there's a 20% chance that we annihilate ourselves with this technology. I generally think that there's a possibility that that's the case.

I think that's one of the reasons why he considers it an awesome responsibility and why ensuring that we don't do that has become the priority of this Pope. It's the same reason they were very preoccupied with nuclear weapons when they were developed. I think they're very much aware that things could go off the rails very quickly and that civilization is really at jeopardy here.

But the Church has been through this before. Not in the same way, but they've been here before. I think hope always animates everything the Pope says and everything the Church believes. Acknowledging that it is a possibility does not mean he feels it's inevitable, right? I think he's saying that to encourage us to, as we say, pivot and break toward the golden path and not the dark path, which is something I try to talk a lot about, because I do think the technology could break in a very transhumanist way, which is not good.

And so this is why I think the church’s voice is so important, because I think ensuring that we have a proper understanding of human flourishing when building that civilization, which is focused on human flourishing, looks very different from the transhumanist vision of a better world.

Nathan Labenz

I definitely want to put a pin in the transhumanist concept and come back to that, too. But when you said we could annihilate ourselves with this technology and it could be the end, that brings to mind the concept of the Antichrist, which has been in the air recently, with someone prominent, Peter Thiel, increasingly fixated on it for unknown—or at least unknown to me—reasons.

I won’t ask you to attempt to channel Peter Thiel’s Antichrist thoughts. But is that a plausible interpretation? I mean, I guess there are sort of a billion Catholics around the world who could see it that way intuitively on their own. But there’s also the church itself. Could you envision a future in which things start to get weirder or darker, and the church begins to interpret AI as the actual Antichrist? Or is there some reason that wouldn’t be consistent or otherwise wouldn’t make sense?

Matthew Harvey Sanders

Well, I mean, the Antichrist, according to the church, is a person, right? He’s not a thing. And I think the church sees artificial intelligence right now as a thing. Maybe a very sophisticated, impressive thing, but it’s still a thing.

So I think the church would see someone like the Antichrist. Satan would use the technology as a tool to destroy humanity, which is what he’s always seeking to do, right? I think it would be foolish to think that the most powerful technology ever invented is something Satan has no interest in whatsoever, and that he’s not working every day to figure out ways to make sure the technology breaks in the wrong direction and destroys as much of civilization as it possibly could.

So, no, I think it’s worthy of some cycles. I am concerned about that. One of my favorite movies is The Dark Knight, and the Joker’s character is like that. There are people who get possessed by these very dark ideas. They’re angry, and they want to lash out.

I think that, with those people having access to powerful technologies like that, they could do immense damage. This is why I’m a big proponent of open source. I just think it’s very important that as many of us have access to powerful AI as possible, to prevent anyone from using it for some dark purpose while we’re basically helpless to do anything about it.

This is one of the reasons why I’m a big proponent of sovereign AI. I’m not just talking about states having their own AIs; I believe that people should have their own AIs. They shouldn’t all have to sign up with one of 4 companies and have the context of their entire lives processed by those companies. I just don’t think that’s ultimately a good thing for humanity.

There’s a principle in the church called subsidiarity, and I think it should be applied to this technology as well. The centralization of this power, limiting its control to only a few people—that sounds like something the devil would do if he’s trying to end the world, right? Whenever I see signals in that direction, I’m generally like, “That’s probably not the way we want to go.”

I think I’m going to work on building up the other kind. And thank God it looks like even some of the heads of the AI labs seem to see the wisdom in that.

Nathan Labenz

Yeah. Concentration of power is definitely super scary. No doubt about it. Some aspects of the open-source vision, I think, are also kind of scary at the moment. Hopefully, they’ll find technical solutions to them.

But I think you teed that up well for a transition to the company. So tell us about Longbeard. What is the story? I mentioned where the name came from. What is your mission? What is your vision? I kind of wanted to just do the whole rundown. Who are your users? Should I envision priests and nuns using it, or are you trying to make money?

The price is quite affordable at $3.99 a month to start. I mean, there’s even a free tier, too, of course. But give us the 101 on the company itself.

Matthew Harvey Sanders

Well, the very short story is that the company’s been around for 10 years, and we started it to help basically build technology to serve the church’s mission. We recognized that technology was going to be key to the church and its mission—the mission to evangelize nations—and we wanted to make sure that it was leveraging it effectively, as obviously, especially throughout the internet age, it wasn’t always great at doing so.

That’s kind of why we started the company, and that led us basically to Rome, working for the Holy Father and different dicasteries and archdioceses and things all over the world. While we were there, we got to collaborate with some very cool companies and people. We got to work with Google on integrating some technology into the Vatican and other places.

We became aware of the power of artificial intelligence. Before generative AI, it still felt like it was really being used for recommendation engines and things like that, but we didn’t really see any immediate use case for it in the church, or the use cases that we saw were practically too expensive to actually implement. So we put a pin in it.

Then, of course, ChatGPT dropped, right? We also found out that Catholics were using it to ask philosophical, moral, and theological questions. I don’t know why I was surprised by that, but I kind of was.

Of course, at the time—I mean, it’s still the case—it had this propensity to hallucinate, and it wasn’t transparent about where its answers came from. Stuff that we all know today. We recognized the power of the technology.

I myself, having converted to the faith and having worked for an archdiocese that was in large part responsible for enforcing doctrine, understood how difficult it is for people to understand this faith, which can sometimes be very complicated. So much of the church’s rich intellectual and spiritual tradition is just not made available to people, and therefore the insights just can’t be brought to bear.

I always wanted to figure out how these libraries of insight could be made applicable to people and help them in their lives. So, obviously, we saw the power of ChatGPT and how it was built, and thought that this could be it.

But of course, we had to make sure that if the church were to adopt it, it would be safe. By safe, I mean predominantly that it’s transparent about when it’s generating answers, what it’s basing those answers on, and that every possible effort is being taken to ensure that hallucinations are reduced to the limit of what’s possible.

So we started a research project called Magisterium AI. Basically, the story is that in July 2023, we launched it for the purposes of expanding our testing group. A Catholic news network found out about it, and I did an interview with them, which I shouldn’t have done. It was stupid.

Then it went viral, and we got so much traffic that we crashed. We couldn’t get back up again. Ironically, one of our advisors, Father Philip Larrey, who was one of the most notable thinkers on Catholic AI in the church, reached out to Sam Altman for us and said, “Listen, there’s this project, Catholic AI. They’re swamped. They can’t get back up again.”

Of course, back then, OpenAI was still very compute-constrained, so only enterprise organizations were getting the bandwidth that we needed. But he stepped up and made a call, and basically the next day we were up and running again.

Ironically, I don’t know if we would be here today without Sam’s intervention. But we took that as a good signal that there was something here. We dropped the agency part of our business and said our mission was going to be building and scaling Catholic AI. Let’s just double down on that. That’s basically what led us from there to here.

Nathan Labenz

That’s fascinating. Do you want to talk for a little bit more about the documents? I mean, maybe just the product line. There are several different products. The one that I spent the most time interacting with was Magisterium.

We’ve got a pretty, in general, AI-engineering-savvy audience, so you can go as deep into the weeds on the details as you like. In fact, it’s encouraged.

It felt to me like basically a RAG-style architecture. If I had to guess, I would say it’s GPT-5 now, but with deep access to documents, which I believe you’ve even gone as far as going into archives and digitizing stuff.

I don't know if that project is separate or is contributing actively to the corpus that the AI has access to. So, I'm interested in what more you can tell us about how that has been developed. Is fine-tuning necessary, or is this a persona that a foundation model is willing to take on? How are you evaluating it? That's really interesting. Obviously, there's right and wrong, but there's also vibes, and I'm interested in what the Catholic vibe test is, if anything. So, long prompt, but take it wherever you want in terms of the AI engineering of what you've got online today.

Matthew Harvey Sanders

Yeah, it's been quite the journey. Today we're in 165 countries, and it's the number-one answer engine for the Catholic faith in the world. And it's been a long journey, with a lot of learning throughout. But I think we learned in the same way everyone else did, right? We started this knowing that a compound AI system was going to be required to make this work.

And so that was a combination of RAG with specialized tooling, prompt templates, and everything else—all the typical things that you expect to see in a compound AI system. A lot of the hard work really was around 2 areas in particular. The RAG database was critical, because when we launched it initially, we had around 600 Magisterial documents in its knowledge base. These are things like the Code of Canon Law and the Catechism of the Catholic Church—pretty seminal works, fairly comprehensive—but it wasn't anywhere near sufficient.

The reason that became clear to us was the long tail. We thought people initially would be going there just asking it straight-up categorical questions, and very soon after we launched it, we saw that people wanted way more than that. They wanted to come and say, "Here's what's going on in my life. What would the Church say to me?" That required the AI to generalize from first principles and apply them to someone's life, when obviously ChatGPT was not very good at that in the early days.

This caused a bit of a crisis in the company, because our commitment was that we do not want to lead people astray. That's the whole reason we started this project, right? Given the state of this technology, should we just shut it down until we can do this more effectively? We had a big think about that, but ultimately realized that if we shut it down, they're just going to go back to ChatGPT. Is that better than what we have? I mean, at least we're getting it right most of the time.

So we said, "Okay, let's stick with it. What do we need to do to address this long tail?" One was that we had to start a massive project to digitize as much of Church knowledge as we possibly could. That involved a lot of things, including building technologies like Vulgate and setting up the Alexandria digitization effort for robotic scanning of documents at scale, which we can talk about in a bit. So that was one thing: expand the knowledge base as large as we can possibly make it.

One particular area we focused on was this generalization problem. How do we take these first-principles documents and ensure that they're properly applied to particular situations? We realized that the popes have been doing that for 1,000 years, and they're doing it in their homilies and in the general audiences, right? They were taking complex theological insights from the Fathers of the Church and distilling them into a 10-minute address.

So we realized that if we got enough of those, that might help. That might actually—and it did—make a huge difference. Now we have over 28,000 Church documents in that knowledge base, and some of those documents can be tomes of books. Providing the LLM all that context helped.

And then there were other problems. We had to build specialized tools because context engineering is obviously critical here. Some things models just can't figure out. What are today's readings, right? Or, "I want to do the Divine Office today. What's in the Divine Office today?" They can't figure this out on their own. They need help.

Anytime we bumped up against issues like that, we had to build tools to ensure that when people asked about them, the right context was served. Of course, throughout that process, we had to iterate on the prompt, and we also had to build really robust evals. Every time a new model came out with frontier capabilities and was better at something like reasoning, we were very quick to want to embrace it, because we needed that for all the generalization our system was doing.

When we first started, we tried to use Google's models, but they were unbelievably woke when they got started. We just couldn't make them work. They were refusing to answer certain Church-teaching questions, so we had to abandon them. And then, obviously, when ChatGPT started, we went to Gemini—it wasn't Gemini; it was PaLM back then. We tried PaLM. It didn't work.

Obviously, when Anthropic came out, we were quick to jump on Claude, and we were on Claude for a while. Then DeepMind took over AI at Google, and the Gemini models were pretty good, particularly in benchmarks that matter to us, like hallucination and needle in a haystack, these kinds of things. So we were with Gemini for a while.

But more and more, we realized that if we're going to be serious about Catholic AI, we're going to have to train it more from scratch. There's just no way to truly achieve alignment with a pretrained model from one of these companies. Right now, in our compound AI system, we're using an open-source model under the hood to power what we're doing. But that's not enough.

So we started the Ephraim program, and we're preparing for our 3rd training run. We haven't released it to production yet, but it's looking very promising. I know you have AI engineers on the team who are probably curious how we did this and how we trained the model. It was extremely difficult.

We're very lucky that, because the company is very mission-oriented and very mission-focused, we're able to tap into some world-class expertise—people we would never be able to hire. Because we've been able to tap into some of the best in the world at specialized models and training, we've been able to get some very unique expertise in here, which has allowed us to do this. But that is to say, we have not done it effectively yet.

I will say that the model, Ephraim 2, in our fidelity benchmarks ranks higher than any other comparable model. It's 50% better than the next competing model. So we have made some very good progress. But of course, the challenges we bump up against when we're training a 3-billion-parameter model, which is what Ephraim 2 is right now, are emerging capabilities like multilingual understanding and reasoning.

To get those emerging capabilities, you need large swaths of data. Obviously, the Phi program is very good at isolating the tokens that are required to produce those emerging capabilities. We need techniques like RLDR and others, so we've been able to sift away, so to speak, needless data. But multilingual understanding is still a challenge.

Whenever we had to rely on those external data sets and train them in with our Catholic data set, obviously in post-training we're then trying to train the model to listen: someone's asking you a question, so make sure you're only tapping into these parameters—the ones that have knowledge of the Catholic faith. You're just using these other parameters to help you understand Korean and things. But it's a messy process.

Although I was very impressed with where we landed with it, there's still more work. There's a new architecture we're working on right now. We also have a lot of new, really exciting techniques, like HRM, that we think we can utilize more effectively.

By the end of the year, we're hoping to have Ephraim 3 trained, and we'll leverage that under the hood in our compound AI system alongside our current system. We'll feed it, basically, the long tail, and we'll see how it holds up. If it seems to do well, then eventually we'll release that into a production environment.

Obviously, that'll be a big deal because it's a small model. It's a lot more efficient than the system that we have now. As we deploy features like voice mode and things like that in the next few weeks, it's computationally demanding, and there are latency issues. So we just feel Ephraim is superior.

And that's not even to talk about where we ultimately think sovereign AI from a Catholic perspective needs to go. Our ultimate vision is that people can actually have Ephraim at home running on their own compute. This Ephraim model could actually tap into their Matter smart-home system, their apps, and things like that, and could actually be their personal AI.

Nathan Labenz

There are many aspects of that that are fascinating. I would encourage listeners, Catholic or not, to go try the Magisterium AI product experience. It’s very smooth and feels very well done. As I said, I had guessed it was GPT-5, so I’m impressed that you have an open-source setup working to the level that it felt like GPT-5 to me.

When you look at the open-source world, do you see an important distinction, from your perspective, between Chinese and Western open-source models? That’s a big debate in AI in general, obviously, and I wonder if you have a unique take on it.

Matthew Harvey Sanders

Just yesterday, somebody sent me a fidelity benchmark from another group. We have our own internal benchmarks, which I think are probably the best in the world for the Catholic faith, but another group had done a fidelity benchmark on a more Protestant kind of Christianity. DeepSeek was at the top of the list. It crushed Grok and Gemini.

By “crushed,” I mean it was a few percentage points, but I think in benchmarks that means a lot. So, yeah, it’s difficult with open-source models. In terms of capability, I think the Chinese basically have parity with us, or at least almost. I don’t think it’ll be long, algorithmically at least.

I think it’s really just a question of what kind of censorship is built into the model and what kind of fundamental biases it has. That’s one of the reasons why we chose to avoid Chinese models. That’s not to say that, when we tested them, we bumped up against that censorship, aside from the obvious stuff around any mention of the Communist Party.

For the most part, they were pretty good. There are just some issues because of the nature of these models being mixtures of experts and things like that. When we’re really pushing them on very specific issues, the model sometimes had a difficult time. It would also start speaking in Chinese instead of speaking in English.

That was just a very difficult technical issue, and we’re not going to try to figure out what parameters are firing there and correct those mistakes. When a group like Groq fine-tuned DeepSeek, I was actually curious to give that a try. Perplexity, I know, had also done that, but it just didn’t work for us.

I do think there’s a lot of potential in open source, and it won’t be long now. We already are leveraging an open-source model, and I think the one we’re leveraging is good; they’re only going to get better. I don’t think we’ll necessarily need to go back to closed source.

That being said, if the right model came around and crushed it on our benchmarks, that’s what we care about. It’s just ensuring that the AI is faithful and that people have a good experience. I would make the pivot in a heartbeat if that were the case.

Nathan Labenz

Yeah, interesting. I don’t know if you can or are comfortable saying what model you’re working on, but either way, I’d be curious to know if you did any sort of continued pre-training on it or your own custom post-training. Or are you finding that available models are steerable enough that, with prompts and scaffolds, you can get the behavior you’re getting?

Matthew Harvey Sanders

Yeah, I don’t mind saying it. Right now, we’re using gpt-oss, and that’s partially because it’s a small model. It took some work because there are some anomalies with it, but we got most of them addressed. We’re still working on it, by the way.

Frankly, we needed a really fast inference provider because of the scale. Not only are we a product, but we also have an API, so we need a superfast inference provider. That’s one of the reasons why we decided to sign up with Groq. They’ve been a great partner to us so far.

I have not had good experiences with fine-tuning, and I think in large part it’s because of the nature of our particular domain. It just doesn’t work well; it doesn’t hold up over the long tail. If you’re trying to make sure that it has the right kind of personality, obviously I think there are a lot of benefits to fine-tuning. I know some people have had success with fine-tuning for a specific kind of knowledge domain, but we felt like the fine-tuning process literally cut it off at the knees.

We abandoned fine-tuning, and that’s one of the reasons why we decided we had to double-click on actually training a model from scratch—not tuning an existing model, but really training it from scratch. Ultimately, I think that’s the only way to ensure that it’s truly aligned.

Part of it, too, is the knowledge that, when our users are prompting these models, they’re trying to always align themselves to the particular user. Of course, they have limited context. When someone starts prompting, they’re trying to figure out, “What are the values of this user? What do they want to hear?” There are all these different value systems competing and banging around in their heads.

I worry that, over a long tail, it may be good 99% of the time. But what we stress about is that 1% where it just cracks and says something crazy, and someone posts it on Twitter. The next thing you know, everyone’s questioning our brand.

We’ve worked very hard with the bishops’ conferences around the world. One of the things we said we’re trying to do is respond in a responsible way. There have been some people who built Catholic AI products that were not built responsibly, and we’ve caught shrapnel from that.

That’s one of the reasons we launched an API. We want to ensure that, after the mountain we’ve climbed to build what we’ve built, we can allow people to build on top of us so they don’t have to climb that mountain. Essentially, that helps create a rising tide, so people can focus on front-end applications and less on the fundamentals. I think that’s been helpful.

I just spend a lot less time taking media calls because some crazy Catholic AI said something wacky, and they want to know how we feel about it. I think the future for us is our own training of our own model. That being said, I still think there’s a lot of potential in embracing open-source models with the right evals framework and system in place.

Nathan Labenz

Can you share user numbers? And maybe what the business model is? Is this something that’s self-sustaining, meant to make money, or even subsidized? The $3.99 price point notably stands out as being at the Khan Academy level of pricing for seemingly everyone around the world.

Matthew Harvey Sanders

Yeah, it’s a good question. We are a for-profit company, so of course we aspire to be as profitable as we can. But we’re also a mission-driven company. One of the commitments we made when we started this is that nobody should have to pay to access the patrimony of the Church.

This is meant to be a gift to humanity. That’s what we felt, and I think the Church believes that’s what God intended. Everything that we digitize, whether it’s through our digitization hub in Rome, digitizing the Pontifical Libraries and all that data, or anything that’s open source that can be made available to people for free, we do so.

That’s one of the reasons why we have a generous free tier, which we hope to continually increase over time. That way, if somebody has a theological or spiritual question, they can get an answer. Our business model is more focused on finding ways to build additional experiences, especially tools, that people can opt to pay for beyond just getting a straight answer.

Things like our voice mode, our biblical commentary features, and deep research, which we’ll be releasing soon, are all bundled together in our Pro plan and meant to incentivize people to upgrade. One of the reasons why we priced Magisterium AI at $3.99 is that we operate on an honor system by country. We didn’t want to price anyone out of the market right now.

I’m hoping that we can keep it at that price point. Obviously, we’ll have to figure that out as we’re in the process of scaling. It’s becoming quite intense. Provided the scaling laws hold and the cost of compute continues to go down, I think we should be okay.

Of course, with the profits we get from the company, one of the things we’re hoping to do is accelerate the digitization project so we can get access to more and more. I’m really excited about the robots in Rome that are working to digitize those libraries.

I think it’s incredible to me that there are books in those libraries that haven’t been opened in 100 years. Who knows the last time somebody actually read them? The insights in those books could be extremely useful.

The more we digitize, the more those insights can be made available to people and impact their lives. There’s a lot of work to do, and this mission of building and scaling Catholic AI has compelled us to move into verticals we never expected, like building Vulgate to help build a state-of-the-art extraction pipeline so we can digitize these things at scale.

Now we’re getting into robotics because we’re trying to remove the human in the loop entirely from this whole ingestion process. It’s fun, but ultimately, I think we’re a for-profit company. We are going to find a way to make this possible.

In large part, it just requires us to make good on some of the technical promises we made, such as training our own model. Rather than always relying on APIs from third parties, it’s going to be difficult. But if we can train our own model, bring our inference compute costs down, and ensure that we’re adding a lot of value so people are willing to purchase the other kinds of add-ons and things like that, then we think we can become a very profitable business.

Nathan Labenz

That it does.

Matthew Harvey Sanders

And I should note as well that our ultimate business model in the long run is not really going to be focused on subscriptions. It’s more about third-party content recommendation.

Let’s say I’m a publisher. One of the existential risks in the publishing industry is that now everyone’s turning to models. Their questions are going to the models; they’re not going to a publishing site or even to Google as much anymore. They have to find a way to increase exposure for their knowledge library to users, to create demand.

One of the programs we’ll be launching soon will use our product Vulgate, which was built to help vectorize libraries initially at scale but can really effectively be used to vectorize anything. University publishing houses and universities with their own holdings can vectorize those through Vulgate. Essentially, when someone asks a question of Magisterium AI, we’ll tap our own knowledge base and give them an answer, but then we’ll provide third-party content recommendations to them. That could be books, videos, or whatever else.

We’ll drive high-quality link traffic out. In this way, we’re able to add value to users by connecting the semantically related content that’s relevant, but we’re also able to drive and create revenue for the business, which hopefully allows us to keep our subscription costs low.

Nathan Labenz

So, how many people do you have?

Matthew Harvey Sanders

I think we’re around 22 right now.

Nathan Labenz

When it comes to training your own model, I take it that’s entirely from scratch, as opposed to off of some base. I’d love to hear a little bit more about your strategy there.

You alluded to Phi, and you also mentioned the perils of fine-tuning, which is something I know well. It’s maybe worth just a quick editorial on fine-tuning. I was the last and least valuable co-author of the “Emergent Misalignment” paper, and one of my big takeaways from that line of research is that if you’re going to do fine-tuning, you should be conscious of the fact that what you’re trading is better performance in the domain of concern for probably worse and certainly unpredictable, sometimes wildly unpredictable, performance in other domains that you didn’t fine-tune on.

In the “Emergent Misalignment” paper, bad medical advice or insecure code led to this sort of general turn toward evil or transgressive behavior. There are different conceptual frameworks to put on it, but when the model decides it wants to have Hitler over for dinner, clearly something has gone wrong that you didn’t anticipate. I always emphasize, too, that this was a surprise to the researchers when they found it.

Coming back to what you’re about to do, I would assume that data filtering, or even just a lot of synthetic data generation, would be a big part of the strategy. Does that mean you’re going to rent model APIs and have them do the creation of data or translation? You’ve got 165 countries, obviously a lot of languages. How do you deal with it if those sources aren’t available?

I would assume that the easiest way is probably to have AI-produced translations in a lot of cases. How many tokens do you need to build up to actually run this sort of thing? What’s the strategy for getting there?

When you finally get to the post-training phase, it’s really hard, right? How can you get confident that the post-training you’re going to do is going to be as robust and aligned in the long tail as what you have today? Obviously, OpenAI is really good at this stuff.

Matthew Harvey Sanders

Yeah. I certainly wouldn’t say we’re better than OpenAI. But I will say that the people we’re working with, specifically in this partnership—I can’t disclose who because they’re in stealth right now, but we’ll announce that partnership soon—are the best in the world, I think, at specialized models like this.

I’m not saying there aren’t other people who are also really exceptional, but I think this person in particular is extremely gifted. The training process is really unique. Everything about the training process is unique, even down to the coding language being used to train the model. It’s something very unique, and I think it has to be in order to do this effectively.

Again, a lot of this is filtering, for sure. I think over time, as the research community becomes clearer and clearer about essentially what tokens are most necessary for desirable emergent capabilities, we can filter out a lot of meaningless data. For Phi-1 and Phi-2, that’s in large part what we’ve been doing.

But that does come at a cost. It comes at a cost because it creates a lot of complications on the post-training end. One is that multilingual understanding suffers, because you end up having to make compromises: you can’t do 20 languages; you can do 6.

When it comes to reasoning, how sophisticated do we need that chain of thought to be? Is chain-of-thought really necessary for our use case if the model is properly trained?

One of the interesting issues is that we know, and Anthropic has written papers on this, that sometimes chain-of-thought isn’t something the models need to do. It’s doing it for our benefit. We know that in some ways it’s already planned its answer, and it’s simply generating tokens to demonstrate how it would arrive at the answer. But we know it somehow intuitively just knew the answer to begin with.

This is one of the challenges we face. When we’re training the model with the core tokens that we have, which represent the Catholic philosophical and intellectual tradition, and we have those other tokens in there, how do we ensure that we train the model in such a way that when it’s making those decisions intuitively about where it wants to go, it’s using the right tokens to do that?

That’s something we have to work with real experts on. This is why the training of the model is done in these very specific steps, so that you basically have these benchmarks that you have to test at different stages of the training. You have to ensure that the thing doesn’t run off the rails as we introduce it to new data to help produce some emerging capabilities.

We then have to check it and make sure that the data hasn’t somehow cracked the model in some way. That process takes a while, but the good news is that because we don’t have to rely on as many tokens, our training costs are lower. That allows us to iteratively run at a much lower cost.

Because we have a very specific use case, it effectively allows us to do what would be more difficult for anyone who’s trying to train a model, obviously, for general-purpose use.

Nathan Labenz

So, how general do you want it to be? In the end, should I be able to bring my algebra homework to it? Is there some sort of boundary on the class of things that you would even want it to engage with?

Matthew Harvey Sanders

To some extent, yes. I’m kind of with Elon on this: I want the model to understand basic physical reality. I think anything that’s verifiable—and those areas are verifiable—I have no problem with model training on code.

I've no problem with model training on math. I think that's obviously very advantageous, even for a frontier model, even for Catholic philosophy and theology. This is one of the advantages that the Catholic Church has against, say, Protestant denominations and things: we have so much more, comparatively, and our tradition is so philosophically and theologically consistent.

We joke—we haven't started this research project yet—but I actually think you could take something like an automatic reasoning platform, like a theorem prover, and you could generate basically a kind of mathematical policy based upon Church teaching. It's that consistent. Because of that, you can surely embrace things like math and code because it just makes sense.

For that reason, I think there are areas where we can be very generous in making use of data. Then there are other areas we should avoid like the plague. This is why I think the humanities are difficult. The reason we have DPO is because telling someone which poem is better—this one or this one—we don't know, right? I think we just drop that kind of data and that kind of post-training, and allow the Catholic intellectual and philosophical tradition to stand on its own.

That being said, eventually, if we're going to train it to be this kind of general-purpose AI that lives at home, it has to be able to answer every question, not just Catholic questions. I think building the future will mean either having a highly specialized model that's really efficient but what it's really good at is classification—it knows what it doesn't know—or having it know what state-of-the-art model it has to tap into to answer a question.

I think that's essentially what we're planning to do: build up scaffolding, almost like an ecosystem around the model, so that it's never constrained by things that it doesn't know.

Nathan Labenz

But, of course, it's smart enough to know that when an answer is generated, if that somehow was fundamentally misaligned with its training data—which in this case would be the Catholic Church's teaching. That's definitely what you mean? A router, basically, is kind of what you're describing?

Matthew Harvey Sanders

Yeah. Essentially, I think that's going to be the future: specialized models and a large ecosystem of them. Through things like MCP and stuff, hopefully we'll be able to tap into them when we need to.

Nathan Labenz

You mentioned that one thing goes wrong and it reduces trust in what you've built with a ton of hard work. Are you doing defense-in-depth sorts of things to try to catch that? There are a lot of techniques now, right? Input filters, output filters, constitutional classifiers—I could go on. Is it to the point now where you're already applying those sorts of strategies?

Matthew Harvey Sanders

Yeah, to some extent. We still have a lot more work to do on this. What we're really working on right now is to basically build some kind of super-judge who can literally look at every single answer that comes in and basically apply a score. If, for some reason, it seems to be problematic, it'll flag it, and we can look at it and investigate what the heck happened there.

For the most part, we've done a pretty good job with the architecture. We don't get that very often, namely because, for the fundamental issues—which would be the most problematic if it got them wrong—we have documents to cover those. It can be done; obviously, any system can be jailbroken. But because those documents are always being served into the context, it's not easy for the model to get too far off the rails. That being said, it still can, so we still have a lot more work to do on that.

Eventually, where we're going is that we want to have an automated LLM-as-a-judge that looks at everything that comes in and flags areas of concern for us. We talked about creating a constitution, and I'm a big fan of that. I mentioned automatic reasoning. I actually think you could derive the fundamental math of the Catholic faith and express that as a policy, and you could ensure that every answer that's generated is at least fundamentally consistent. The math adds up, right? I think there's a lot of potential in that particular direction.

Then, of course, in combination with actually training the model according to a very specific, fairly comprehensive eval set, it makes the job a lot easier. One of the advantages that we've had—we talked a little bit about users and distribution—is that we've been very fortunate because we're the leading Catholic AI. As we focus on the top of the funnel, like the hierarchy and things like that, slowly but surely it's been making its way down to the grassroots.

That's how we wanted it, because initially we wanted our users to be people who were highly discerning. That way, if there were issues, they would be very encouraged to flag them: “This is maybe not wrong, but it's not good enough.” That feedback has been enormously helpful to us as we've continued to build out our company AI system.

Since we became a platform as well, we've been very blessed to be partnering with Hallow, which is the world's largest prayer app. They've added Magisterium AI, integrated right into their system, and they also provide us very helpful feedback. As we reach more users, we learn of particular areas where our knowledge base may be underrepresented.

Then we have to go and tell our ingest team, “Listen, we're not really good in this area. Let's say it's some obscure, ancient form of Christianity. Go find some books on that, digitize them, and let's try to build out that area of our knowledge base.” That kind of feedback is essential, and it certainly informs the way we go about prioritizing work.

Nathan Labenz

It's really striking how much of this is remarkably convergent with stories that I've heard from many other places. When it gets down to actually doing the work, a lot of it is just the same stuff.

With that in mind, I just had a conversation a week ago today with the CEO of Databricks, Ali Ghodsi. You may know they acquired a company called MosaicML. What they were doing was basically providing, as a service to enterprises, something like what you're doing: creating your own model with your own data. They were often doing continued post-training, as opposed to totally training from scratch.

The idea was that if you're an enterprise, you've got all this data. I always think of GE and 3M—companies with unbelievable amounts of products that they've made over time, unbelievable numbers of employees, and a 100-year tradition. It's obviously no 2,000-year tradition, but it's pretty big, right?

He surprised me by saying that they killed that product, that they're no longer offering these deeply custom LLMs to enterprises anymore. Do you think that was a mistake? I don't want to put him on the spot to criticize his business strategy, but it seems like you're finding a need for it. Everything you're telling me has me coming back to the same notion I had before that conversation, which is: I don't know, if the Catholic Church wants this, why don't GE and 3M also stand to benefit from it? Any thoughts on that?

Matthew Harvey Sanders

Well, I think part of it is benchmarks. For the most part, I think people just want to feel like they're talking to the most intelligent model they possibly can. They feel like the answer is probably just a little bit better because this model benchmarks so much higher.

There's just this culture of, “I, as a business, want to be able to tell our employees that we're working with the best—the best models in the world.” I think sometimes that even trumps the cost. Honestly, the cost of training a model on your own data is very small.

Part of this, too, is because evals are really hard. In large part, it's one thing to train a model based on what someone says; it's another thing to be supremely confident that that training was successful. Unless that company has really, really worked to mine and distill the core insights from the subject-matter experts in the business, it's hard to know that the model truly works.

I think this is why there are companies like Distyl AI that exist. The act of trying to extract the insights from the critical employees—the leaders—is so hard and time-consuming that people just abandon it altogether. At the end of the day, they realize that using a state-of-the-art model, even one like Gemini 2.5 Flash, still seems to get comparable results. Obviously, the team that trained that model is the best in the world, and it has a high level of generalization capability, meaning its intelligence benchmarks are definitely higher than our model.

Nathan Labenz

So why risk it? Why wouldn’t we just use Gemini 2.5 Flash and let the infrastructure—the training and the rest—be somebody else’s headache?

Matthew Harvey Sanders

That being said, I think that’s because, for most industries, I’m not convinced that alignment is that big of a concern for them. I don’t think they feel they’ll find a lot of competitive advantage by being a specialist. Whereas in other sectors like ours, people just want to know that the model is not capable of saying something antithetical to faith. They just want to trust it.

If they’re going to come to it confessionally and share, they want to know that they’re getting good advice. They want to know that they’re getting good advice not from a model, but from the great thinkers in the Church—these saints, philosophers, and theologians. I think our use case, and many others out there, calls for a high degree of specialization because trust is such a critical factor. But I just think in industry, I don’t know if that trust is there.

Nathan Labenz

You had also alluded earlier to some apps not being developed responsibly, and I wanted to get your take on what that means. This question echoes all over the AI space, but in doing my homework for this, I came across Text With Jesus, for example, as an interesting app out there. You mentioned the idea of coming to it confessionally. I actually did ask the Magisterium product if I could confess to it, and it said, “No, you can’t. You’ve got to go to an actual priest and do it the right way.”

It does seem like we’re headed for a weird world. I remember Avi, who created the Friend product, once said something along the lines of, “I’m not trying to make an assistant. I’m trying to make something that is closer to how people traditionally related to God.” It’s always on, always kind of watching you. Putting a positive spin on that, I assume it would be like having an angel on your shoulder, always either inspiring or encouraging you to be your best self.

But I can also imagine that, of course, we’ve got all these examples of people being sort of deranged in part because of their ongoing conversations with AIs. There’s definitely a sort of voice-of-God dynamic to it, especially since you mentioned you’re launching audio soon. The more this is a literal voice that people hear, the more that may change how they relate to it. How do you think about what is responsible to build? What form factors are advancing the mission versus potentially leading people astray? How do we not end up in a sort of idolatry-of-AI end state? It does seem like there are some natural tendencies leading us there.

Matthew Harvey Sanders

Well, I think it all starts with a firm commitment to ensuring that the product is maximally faithful. If that’s where you’re starting, then you have to take every precaution you possibly can to ensure that your product is not going to cause anyone spiritual harm. That requires a certain level of technical expertise. You can shortcut it, as some people have done, by literally putting a prompt together, using a model, deploying an app with a cool UI, and letting it run roughshod. Then you don’t have the evals, so across a long tail, your app doesn’t always hit the mark.

How comfortable are you with the AI running someone off the rails? I’m not comfortable with it. I feel it every day. In this particular domain, you just have to have a radical commitment to fidelity. If you’re not willing to put the work in—if you’re not willing to put the evals together and use the right API—then just don’t do it. Just don’t do it. That would be the first thing I would say.

The second thing is about the direction the technology is going. For those who want to stay plugged into the GPT world, having an AI with infinite context in your life is something I think we’re inevitably probably going to have to have. If we want a Jarvis at home, like Iron Man, Jarvis works so well because it knows Tony so well, because it’s just always around.

I think people should have opted out of that and said, “Not interested. I’m going to go live on the farm with my other families, and I’m done with this.” But for those of us who have to stay here, I think it makes sense in some ways to embrace it. If for no other reason, I just want to know what’s real and what’s not real anymore. How do I know what’s in my inbox is written by a human being and not by an AI? You see phishing and stuff. I just don’t know if I’ll be able to tell at some point. These AIs are getting so intelligent.

We have to be able to trust these things. The question is, what would it take for me to trust a model to have infinite context in my life? I don’t trust OpenAI that much. I don’t trust any of the companies that much. They have shareholders, and the government can come knocking and ask for the data. Apparently, Sam said on the record that they do—we have to hand it over. I’m just not comfortable with that.

That means we need to have some other form of AI, some other stack that people can use, that can secure that kind of trust. People have to feel that the models are fundamentally aligned on a very deep level. The question is, how do you do that? For the Catholic Church, it’s easier because Catholics say, “The Catholic faith is what I ascribe to.” As long as the model understands the Catholic faith, cool. It can get to know me over time.

But if you’re someone who doesn’t have a very clear doctrinal system that governs your life, how do you ensure that the AI is really aligned to you? Certainly, when you buy it, it’s not going to be. You just have to live with that tension and hope that, over time, it’ll figure out the idiosyncrasies of what you believe. I think it’s going to be hard to fundamentally build AI that is truly aligned in some use cases.

But because the Catholic Church has this unique opportunity to be the first ones out of the gate, so to speak, to do this the right way, we should do it. If for no other reason, it already has to exist. I think it’s definitely critical for the future of civilization, and we can’t wait for someone else to do it. The Church has to lead on this, with the adoption of this technology, as it has in the past.

Nathan Labenz

Cool. I’m just trying to think if there’s any natural follow-up there. I did see in the Text With Jesus reviews that people were at times saying, “Your app is not actually representing my faith.” Obviously, there are a lot of different forms of Christianity and a lot of disagreements that could all be encompassed in something like a Text With Jesus experience. But that was striking to see in the reviews: people are not necessarily happy with its faithfulness to what they perceive to be the right and true teachings.

As we come to a close, there’s often this tension in religious communities between doing our own thing and doing what we believe God wants us to do and living the right way, and then engaging with the rest of the world. As big as the Catholic Church is, it’s still significantly outnumbered by the rest of the non-Catholic world. It seems like we’re headed for a radically different future.

It’s no longer totally crazy to talk about curing all the diseases. My sense from talking to Magisterium is that the Catholic Church is cool with that: cure them all one by one. But then there’s also transhumanism, which you mentioned earlier and which you think is kind of a bad idea. It’s not clear to me what the line is between curing all the diseases and transhumanism.

I asked Magisterium again, “Oh, great news. There’s been a life-extension drug created that will allow me to live 1,000 healthy years. Should I take it?” It told me no because, as I understood the response, it was too radical a change to the fundamental nature of things. Fixing defects one by one would be okay, but this is somehow crossing a line into something that’s not just fixing defects and stepping outside of what we conceive of as God’s plan or intent or whatever.

How do you think about how far you’re willing to go if this technology gets more and more powerful? How much of that would you happily fold in? Is there such a thing as an AI that’s too powerful or too transformative, such that you would inherently say it’s inconsistent with the Church? And how do you, if at all, intend to shape the mainline AI trajectory from your perspective? So you might just say, “Nah, they’ll do what they’re going to do.”

We're going to try to do what we think is the right thing. But is there any sort of aspiration to have some feedback loop into the mainline R&D efforts to shape how those go?

Matthew Harvey Sanders

I certainly think artificial superintelligence is probably inevitable. If that is the case, I am very interested in ensuring it's aligned. For obvious reasons, as we all are, right? So the question is, what are we aligning it to?

I generally think that, if nothing else, what Catholicism has created is probably the best toolkit for enriching life. And I think the most critical research project, aside from, let's say, cheating at chess or something, would be nailing what human flourishing actually consists of, and then being able to take a qualitative understanding of it and validate it with quantitative data. If for no other reason, we could actually train that insight into a model.

So whenever it's assigning what course of action to take, it could run it through its human flourishing understanding and be like, "Probably not a good idea. That's not aligned with human flourishing. Therefore, I'm not going to proceed." We need that. We need to be able to clearly articulate to a model. We need evals for human flourishing.

If we don't, we're basically ceding that it's just going to somehow intuitively pick it up from all the data we feed into it. But now, as we need more and more data and we're generating it synthetically, we're probably avoiding this critically and fundamentally important area because it's so dubious, right?

The reason why I think ethical AI is so fraught with peril, and why the labs don't really want to touch it, is: align it to whose ethical framework, right? Which is why I don't think ethics is the right approach. I think flourishing is the right approach.

I think we found, when we did this project at the Humanity 2.0 Foundation, that generally it's a lot easier to get people to agree on what they need to flourish than it is to agree on values or an ethical framework specifically, right? But saying things like, "Hey, do you think it's good that you have a mother and father in your life?" "Yeah, I think that's generally a good thing." Is it good that everyone has access to high-quality—yeah, that's good, right? That makes sense.

These are all common sense. We need to capture that common sense, frame it up, and train it into these models. And I think if we do that effectively, I don't think we have as much to worry about from artificial superintelligence, right? I know that that's going to be difficult, but I think that that's critically important.

As far as our transhumanism goes, generally the church's position has been: if a technology helps, let's say, restore someone to full humanity, that's a good thing. So let's say that they lost an arm in an accident and we build them a robotic arm, right? Then they have their arm back, and they can go into the world and do what they were meant to do.

When Alexandr Wang says, if you're deferring having a child because you're waiting for Neuralink developments to get to a point where you can actually ensure that, when your child is born, you can merge your child with a machine to increase its computational capabilities, so it's able to make more substantial contributions to the technocratic vision that they have—probably not good.

Now, I respect Alexandr Wang, but I just think the church is not going to be cool with that, right? If someone has a damaged brain, using machines to help heal that damage, fine. But it all comes down to intention.

If you're merging with a machine because, for some reason, you don't like the idea that AIs are somehow cognitively more capable than us, and so you want to merge with a machine so you can increase our relevance, so that we feel we're part of this project to get to Mars or something like that, that's where I think it gets really dubious. I think the church is going to have a very important role in setting off alarm bells around that, because I don't want to live in a Cyberpunk 2077 world, right, where people are walking around half machines because we thought it was a cool thing to do.

Nathan Labenz

Do you think the church will ever get to the point where it'll actually join the PauseAI movement?

Matthew Harvey Sanders

Well, it came pretty close. Father Philip Larrey signed, along with a bunch of other people, to slow it down. If I were talking to the Pope, I would say I think it's important to continually advise that we should probably slow this technology down until we have time to develop the right regulation and things like that.

But be pragmatic. As long as China and the US are in this competitive dynamic, they're not going to slow down. So is that the best use of your time—basically campaigning for a hopeless cause—or is it maybe more critically important to ensure that the technology that we are developing at a frantic pace is aligned to some kind of concrete objective that we can all agree is a good thing?

That, to me, is a far better use of his time, which is why I think making clear what human anthropology is and what the telos of civilization is is so critically important.

Nathan Labenz

This has been a fascinating conversation. I really enjoyed it. I really appreciate your time. Anything else you want to share—anything we haven't touched on, or any final thoughts you want to leave people with?

Matthew Harvey Sanders

No. Thank you very much for the opportunity. And to answer your question, yes, we're focused specifically on building a scaling Catholic app, but we're very interested in contributing to the overall research field. So anything we can do to help support that—and if there's anything the research community feels they can do to help advance the mission of building a scaling Catholic app—we're very open to collaboration.

Nathan Labenz

Awesome. Matthew Harvey Sanders, founder and CEO at Longbeard, you're building Catholic AI. Thank you for being part of The Cognitive Revolution.

Matthew Harvey Sanders

Thanks for having me.