Pope Leo XIV’s newest encyclical on artificial intelligence, Magnifica Humanitas, has quickly become the center of a very modern kind of controversy: not about what the document says, but about how it might have been written.
A new analysis shared on the LessWrong forum argues that parts of the encyclical show statistical and stylistic signals that are consistent with AI-generated text. The claim is not that the Vatican has admitted to using AI tools, nor that the analysis can definitively prove authorship. Instead, it suggests that some passages may have been produced—or heavily revised—using AI assistance, based on how an AI-detection tool interpreted the writing.
The discussion began with an assessment posted by Linch Zhang, who examined sections of the encyclical and reported that certain paragraphs appeared, according to the AI detector Pangram, to be between roughly 40% and 100% likely to have been written by AI. A separate user then reportedly ran additional portions of the text through Pangram as well, finding that a substantial share of the encyclical’s early material matched patterns the detector associates with AI writing.
That alone would be enough to spark debate in any context. But when the subject is an encyclical—an official teaching document intended to carry theological authority and careful rhetorical weight—the stakes feel different. Even if the analysis is ultimately inconclusive, it raises a question that many institutions are now grappling with: what does “authorship” mean when drafting tools can mimic human style so convincingly?
To understand why this story has caught fire, it helps to look at what the analysis actually did, and what it didn’t do.
The LessWrong post frames its findings around the behavior of Pangram, a popular AI detector. These tools typically work by looking for patterns they associate with machine-generated language: word choice frequency, sentence structure tendencies, repetition rates, and other subtle markers. In this case, the analysis points to a notable increase in the use of the word “genuinely” compared with previous encyclicals, describing it as a trait that appears more often in writing attributed to Anthropic’s Claude than in earlier papal documents.
This is where the conversation becomes both fascinating and frustrating. Fascinating, because it shows how AI detectors attempt to reverse-engineer authorship from style. Frustrating, because it also highlights the limits of that approach. A single word—especially one that could plausibly be used for emphasis in a theological argument—may not be a reliable indicator on its own. And even if a detector flags a passage, that doesn’t automatically mean AI wrote it. It could mean the text was edited by someone using AI as a drafting aid, or that the author intentionally adopted a style that resembles what detectors expect from AI output.
Still, the analysis doesn’t rely solely on one word. It describes multiple “known traits” that appear in AI-generated writing, and it uses Pangram’s scoring to estimate the likelihood that specific paragraphs were produced by AI. The reported range—40% to 100%—is striking, but it should also be treated cautiously. AI detectors often produce confident-looking percentages even when their underlying methodology is uncertain, and their accuracy can vary widely depending on the text type, the detector version, and the writing domain.
In other words, the numbers are attention-grabbing, but they are not the same thing as proof.
What makes the debate more intense is that the encyclical itself is about AI’s impact on humanity. Magnifica Humanitas is not a casual blog post; it is a formal statement meant to guide moral reflection. If the document warns about risks tied to AI—whether those risks are ethical, anthropological, spiritual, or social—then the possibility that AI-assisted drafting played a role becomes a kind of meta-issue. People naturally ask whether the message and the method align, and whether the institution’s stance on AI is being practiced consistently.
But there’s another angle that deserves attention: even if AI assistance occurred, it doesn’t necessarily follow that the encyclical’s content is “fake” or “inauthentic.” Many institutions already use software tools that shape writing—grammar checkers, style editors, translation systems, and document formatting tools. AI drafting tools are different in degree, not just in kind. They can propose phrasing, restructure arguments, and generate polished prose quickly. Yet the final responsibility for doctrine and wording still rests with human decision-makers.
The real question is not simply whether AI was used. It’s how it was used, and what level of human oversight existed. Was AI used as a brainstorming partner? As a language polish tool? As a first draft generator that humans then rewrote extensively? Or was it used more directly to produce large sections with minimal human rewriting?
Those distinctions matter because they change what “AI involvement” means. A small amount of AI-assisted editing is not the same as outsourcing authorship. And even if AI helped produce a draft, the theological and rhetorical choices—what to emphasize, what to omit, how to frame a moral argument—still require human judgment.
At the same time, the public’s concern is understandable. When a detector suggests that a document’s style resembles AI output, it can feel like a breach of transparency. Readers may wonder whether they are engaging with a human voice or with a system trained to imitate human language. For religious texts, where tradition and personal conscience are central, that concern is amplified.
This is why the story has moved beyond a technical curiosity and into a broader policy and ethics conversation.
If you’re assessing AI involvement in official Church writing, what would you want clarified? The most obvious answer is confirmation from the Vatican. But confirmation alone may not be enough. Transparency about drafting tools—at least at a high level—could help readers understand whether AI was used and for what purpose. For example, did the drafting team use AI to translate or to rephrase? Did they use it to generate outlines? Did they use it to check readability? Or did they use it to draft entire sections?
Even without naming specific tools, a disclosure framework could reduce uncertainty. Institutions could say, for instance, whether AI was used at all, whether it was used only for language polishing, and whether any AI-generated text was reviewed line-by-line by human authors. That kind of information would allow readers to interpret the document with more context, rather than relying on detector scores.
There’s also a deeper issue: AI detectors themselves are not neutral arbiters. They are models trained on patterns of machine-generated text, and they can be wrong. They can also be gamed. If a text is edited to resemble human writing, detectors may fail. If a text is written in a style that overlaps with common AI outputs—especially in formal, high-level prose—detectors may over-flag it. That means the detector’s output should be treated as a hypothesis generator, not a verdict.
So what should readers take away from the LessWrong analysis?
First, treat it as an investigation into style signals, not confirmed authorship. The analysis is valuable because it encourages scrutiny and asks whether AI tools might be influencing official writing. But it does not establish that AI wrote the encyclical. It establishes that some passages look statistically similar to what Pangram expects from AI-generated text.
Second, recognize that the encyclical’s subject matter makes the question unavoidable. When a document warns about AI’s dangers, the public will naturally scrutinize whether AI was involved in producing the warning. That scrutiny is part of the cultural moment: we are learning to evaluate not only claims, but also the processes behind them.
Third, understand that the Church is not the only institution facing this problem. Governments, universities, and corporations are all dealing with AI-assisted drafting. Some are adopting internal policies; others are still figuring out what “responsible use” means. Religious institutions, however, face a unique challenge because their documents are meant to be read as expressions of moral and spiritual authority, not merely as information.
That uniqueness is why this story resonates. It’s not just about whether AI can write like a human. It’s about whether institutions that speak for humanity’s values will adopt the tools that reshape how language is produced.
There’s also a subtle irony in the current debate: AI detectors are themselves products of machine learning, and they are being used to infer whether another machine-learning system may have contributed to a text. This creates a loop where AI is both the suspected author and the judge. That doesn’t invalidate the analysis, but it does complicate the interpretation. If detectors are imperfect, then the conclusions drawn from them are inherently probabilistic.
And yet, probabilistic conclusions can still be meaningful. They can prompt questions, encourage transparency, and motivate better internal documentation. Even if the detector is wrong, the fact that it flags a document suggests that the writing style may have shifted in ways that deserve explanation.
One reason the analysis points to “genuinely” is that word choice can be a fingerprint. But fingerprints can also be coincidental. Formal writing often uses certain adverbs and intensifiers repeatedly. The difference is that AI writing tends to distribute such intensifiers in particular ways—sometimes more evenly, sometimes more frequently, sometimes with a pattern that detectors learn. If a papal document uses a word more often than prior documents, it could reflect a deliberate rhetorical strategy, a translation nuance, or a change in drafting team preferences. It could also reflect AI-assisted rewriting. Without access to the drafting process, it’s impossible to know which.
That’s why the most productive next step is not to argue endlessly about detector accuracy. It’s to push for clarity about process.
If the Vatican chooses to address the question, it could do so in a way that respects both the seriousness of the topic and the reality of modern drafting workflows. A measured statement could explain whether AI tools were used, and if so, what safeguards were applied. It could also clarify that doctrinal content was reviewed by human authorities and that any AI-generated text was treated as a draft material requiring human judgment.
Such a response would not only calm speculation