What is Catholic AI? Technology Meets Theology, with Matthew Harvey Sanders, CEO of Longbeard
Summary
- 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.
Deep dive
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.”