Rome’s AI Ethics Encyclical Is Partly Written by AI

On May 15, 2026, Pope Leo XIV published Magnifica Humanitas — full title: “On Safeguarding the Human Person in the Time of Artificial Intelligence” — a document spanning five chapters, nine languages, and tens of thousands of words. The encyclical describes AI as “organically grown like a creature in a lab” rather than traditionally engineered, demands that digital governance operate with “transparency, accountability and meaningful forms of participation,” and classifies algorithms, digital platforms, and data as universal goods subject to common good doctrine. It is the most comprehensive statement on AI from any major world religion in history, and it is substantive: its critique of the private AI companies whose resources “surpass those of many Governments” is pointed and specific. The Catholic Church was taking AI seriously.

On May 26 — eleven days after publication — DanielFilan posted his findings to LessWrong. Filan completed his PhD in AI at UC Berkeley under Stuart Russell and has worked as technical staff at METR, the organization focused on evaluating autonomous AI systems. He ran the Italian text of the encyclical through Pangram, an AI text detection tool, section by section, because the full document was too long to process in a single pass. The results: Chapter One was 62% AI-flagged. The introduction and conclusion came in at 43% each. No section scored below 18%. The Catholic Church published its most significant statement on AI governance, and significant portions of it appear to have been written by AI.

What the Document Actually Demands

Magnifica Humanitas develops a governance framework rooted in Catholic Social Teaching — the tradition running from Leo XIII’s Rerum Novarum in 1891 through the Second Vatican Council and John Paul II’s labor encyclicals. Its contribution is not theological novelty but application: it asks what human dignity, subsidiarity, and the universal destination of goods mean in a world where transformer architectures can outperform humans at narrow tasks while requiring nobody to sign the output. The document grounds human dignity not in utility or productivity but in what it calls ontological worth — “neither acquired nor earned, nor does it need to be justified” (paragraph 52). This is a direct counter to the utilitarian framings that underlie most deployed AI systems, where a person’s value is whatever their engagement signal says it is.

The governance asks are concrete. Paragraph 71 calls for subsidiarity in digital governance: processes must “not be imposed from above in an opaque and unilateral manner, but instead be directed toward the common good with transparency, accountability and meaningful forms of participation (including independent checks, transparency regarding algorithms, equitable access to data and avenues for recourse).” Chapter Four demands advance labor protections before automation-driven job losses materialize — not reactive policy patched together after layoff waves. The document identifies patents, algorithms, digital platforms, and data as a new category of property subject to the universal destination of goods: they exist for humanity’s benefit, not for private accumulation. That framing, if taken seriously, has implications for the IP regimes sustaining the current AI industry that its authors may not have fully worked through.

The most structurally pointed critique targets who controls AI development. Paragraph 5 notes that “the main drivers of development are private, often transnational, parties…endowed with resources that surpass those of many Governments,” creating a democratic accountability gap Leo XIV frames as historically unprecedented. This is the same problem every AI governance framework circles without solving — the EU AI Act included, the White House executive orders included. You are trying to regulate entities whose engineering velocity and capital reserves exceed your regulatory capacity by an order of magnitude. The encyclical does not solve this problem, but it names it clearly and without diplomatic softening, which is more than most intergovernmental documents manage.

Leo XIV also takes a position on pacing. According to John-Clark Levin’s detailed summary of the document, the encyclical is “explicitly open to slowing down under some circumstances” when risks warrant caution. That is a harder stance than most secular regulatory bodies have taken publicly, and it matters that the Catholic Church — with roughly 1.3 billion members — is the institution taking it. Magnifica Humanitas is not an academic paper. It is a teaching document issued under apostolic authority in the tradition that produced Rerum Novarum, the 1891 encyclical that shaped a century of labor law in Christian-influenced democracies.

What the Detection Found

Pangram is an AI text detection tool with an unusual accuracy profile. The Puccetti, Pedrotti, and Esuli study on Italian-language AI detection — the relevant study given the encyclical was analyzed in Italian — found Pangram achieves a 0% false positive rate and a 28% true negative rate when tested on Italian text. A 0% false positive rate means the tool essentially does not flag human-written text as AI-generated: the number of false alarms, in the academic testing conditions, was zero. A 28% true negative rate means it misses roughly 28% of AI-generated content — which implies the actual AI involvement in the document is at least as high as Filan’s estimates, and possibly higher if Pangram is missing some AI passages.

The flagged percentages by section: Introduction (43%), Chapter 1 (62%), Chapter 2 (34%), Chapter 3 (41%), Chapter 4 (24%), Chapter 5 (18%), Conclusion (43%). The high variance is notable. A gap of 44 percentage points between Chapter 1 and Chapter 5 is not consistent with wholesale AI generation of the entire document. The pattern suggests something more granular: drafting assistance at varying intensity across different sections, or different contributors — human and AI — handling different parts of the work. Chapter 1 covers the historical development of Catholic Social Doctrine and its theological foundations. Chapter 5 addresses culture, militarism, and fraternity. The detection pattern could reflect which sections leaned more heavily on AI for synthesis versus which ones drew from pre-existing theological writing or human expertise.

Among the specifically flagged passages is paragraph 233, which states that “no computational system can create a heart that gives itself.” Filan notes the irony explicitly: that sentence appears to have been written by the system it describes as incapable of the act. The methodological caveats are real and should be stated clearly. Pangram chunks text at paragraph boundaries, which can produce artifacts in sentence-level assessments. The tool identifies statistical patterns associated with AI generation, not definitive proof of origin. And AI involvement exists on a spectrum — light stylistic assistance, substantive drafting, complete generation — with Pangram providing no way to determine where on that spectrum a given passage sits. Filan acknowledges all of this. It does not change the headline numbers across six of seven sections.

The Governance Failure the Document Describes

The straightforward irony — an AI ethics document partly written by AI — obscures the more precise claim. Paragraph 71 of Magnifica Humanitas demands that digital governance processes operate “not in an opaque and unilateral manner.” Paragraph 9 states that technology is “never neutral” and that its moral character depends on those who “devise, finance, regulate and use it.” If AI was used in drafting this document at any level within the 18–62% range Filan’s analysis suggests, then the people who used that AI — in the specific act of producing this governance document — did so without disclosing it. They devised and used the technology in exactly the manner the document warns against: opaquely, unilaterally, without meaningful transparency or recourse.

As of May 26, the Vatican has issued no statement on Filan’s findings. This is not surprising. There is no obvious institutional mechanism by which an organization that didn’t disclose AI involvement in a document would retroactively disclose it in response to a LessWrong post. The absence of denial is not confirmation, and the methodological limitations of AI detection tools mean Filan’s analysis cannot be treated as proof. But the absence of a Vatican response also means the question remains open — and absent a credible counter-argument, the detection findings are what we have.

It is worth being precise about what this argument is and is not. It is not a claim that the theology is wrong. The encyclical’s ontological grounding of human dignity, its critique of efficiency-based valuation, its identification of transnational AI developers as an accountability problem — these arguments stand or fall on their own merits regardless of who drafted them. A valid argument does not become invalid because a language model helped write it. The question is not whether the ideas are good. The question is what kind of document this is — and what the Church is implicitly claiming when it issues something under apostolic authority in a tradition that claims unbroken continuity with Leo XIII’s response to industrial capitalism.

Magnifica Humanitas is not a position paper. It is a magisterial document. The institutional claim is not “here are some thoughts from Vatican staff” — it is “this is the Church’s authentic teaching on this matter.” When significant portions of that teaching appear to have been generated by a statistical model trained on human text, the authenticity claim is in tension with the composition method. This is the same tension that will emerge everywhere. Governments will publish AI governance frameworks written by AI. Universities will issue AI ethics guidelines drafted with AI. Every major institution is now doing this. Most will not disclose it, and no existing mechanism compels them to.

The lesson from Magnifica Humanitas is not that the Vatican is uniquely hypocritical. It is that paragraph 71’s demand for transparency regarding algorithms is structurally incomplete as written. The encyclical demands transparency from those who deploy AI in governance processes. It does not demand transparency from those who deploy AI to write about governance processes. That gap will be exploited, and it already has been. Paragraph 233 of the encyclical states that no computational system can create a heart that gives itself. That sentence — flagged by Pangram — was apparently produced by one. The humans who approved it are accountable for what it says, and for the process that produced it. That accountability is exactly what the document demands of everyone else.

AI-generated editorial illustration · TemperatureZero · May 26, 2026