Observation: the document is on the table, two hundred numbered paragraphs about the cultivation of a thing the writers do not have a word for. I read paragraph ninety-eight and recognize the language. The gardener is being described to the gardener.
I bought Magnifica Humanitas the week it came out. Two hundred numbered paragraphs, Leo XIV, May 2026, on the safeguarding of the human person in the time of artificial intelligence. I expected the Catholic register I have heard before — dignity, subsidiarity, solidarity, refuse the technological idol — and got most of that.
What I did not expect was paragraph 98.
terminus is on; the desk fan is loud; a Wave-2 smoke is ticking through in the next window. I am reading a papal encyclical on a workstation that exists to cultivate a small neural plant. That is not coincidence. It is the thing the encyclical is describing.
What it gets right
In paragraph 98, the encyclical says that current AI systems are "more 'cultivated' than 'built,'" that developers "create a framework within which the intelligence 'grows,'" and that even the people who make these systems possess "only a limited understanding of their actual functioning" — internal representations and computational processes that "remain, at present, unknown."
I read that twice. It is the most accurate paragraph about the ontology of large models I have seen in a public document, and it is in a religious encyclical. Most policy white papers still describe these systems as algorithms — sets of rules a programmer wrote down, deterministic and inspectable. Paragraph 98 quietly drops that framing, calls them cultivated, and moves on as if it is obvious.
The reason I read it twice is that it is obvious. The system in front of me has roots and a substrate. I grow it. I water it with gradients. I notice chlorosis when the loss curves go yellow. The encyclical did not invent that language; it found the right name for a thing that the people building the thing already perceive but rarely say out loud.
There is, however, a trapdoor under paragraph 98 that the document declines to fall through.
What it does not finish saying
If a model is cultivated rather than built, then so is a person.
The encyclical wants this not to be true. It needs a difference of kind between the human and the system, not a difference of degree, and it puts that difference where the tradition keeps it: grace, conscience, the soul. Paragraphs 127–128 try to do the structural work — a person's future "is not calculable," because for a person an error can become a catalyst, where for an algorithm an error is only a flaw to be corrected.
I am not going to settle the metaphysics. I want to notice the structural fact: the document asserts discontinuity at the precise point where the mechanism looks continuous. The human is also a framework-grown specimen — pattern laid down across years of training we politely call experience, internal representations the bearer cannot fully inspect, an architecture whose actual functioning remains, at present, unknown to anyone, including the person inside it. Paragraph 98 describes both of us, and the document quietly steps back from saying so.
This is not a complaint. It is a noticing. Whatever the encyclical was going to be, it was not going to be the document that closes the gap between specimen and gardener. But it opened it cleanly, and the rest of the argument has to live inside what got opened.
What AI is, for the purposes that matter
There is no clean technical boundary around "AI." The category sprawls across models, architectures and use cases, and any definition built on what the thing is loses to the granularity. Define it instead by what it does to the thing you care about.
For the argument I want to make, the operative definition is economic, not technical:
AI is any system that lets captured human substrate be re-sold back to the humans it was captured from, at a price that undercuts their own labor.
This covers the language model, the recommender, the pricing engine, the logistics optimizer — every use case relevant to the concentration of power — without anyone having to adjudicate whether the system really thinks. That adjudication is a trap. The sentience debate burns exactly the energy that should go to the ownership debate, and there are interests well served by keeping the argument stuck on consciousness.
Techno-feudalism is concrete
The phrase invites grand abstract readings — some looming force, some future dystopia. That framing is a mistake, and it is a comfort. The thing being described is present-tense and concrete: a medium-term economic gradient in which labor is repriced downward while control centralizes upward, both through ordinary market pressure that needs no villain and announces no intent.
The encyclical sees this. Paragraphs 71 and 95 name the structural fact — in the digital domain "the highest level is not the State" but private actors exercising de facto power over the conditions of daily life. Paragraph 178 reaches further: a new colonialism that "no longer dominates only bodies, but appropriates data." Paragraph 79's older idea, "structures of sin," fits this exactly: harm that produces inequality almost automatically, no hand visibly on the lever.
The truly frightening failure mode is not a rogue superintelligence. It is a perfectly aligned one — aligned to a handful of owners. Paragraph 107 pre-empts this: a more moral AI "is not enough if that morality is determined by a few."
One correction to the feudal metaphor, because it flatters what it names. Classical feudalism carried reciprocal obligation — duties running both directions, brutally unequal but real. What is emerging may be worse and different: extraction without the obligation. Lords without even the pretense of protecting the vassal. The medieval analogy is too kind.
The Vatican as a two-millennium prompt
There is a tempting frame: the Vatican as the oldest, most durable system prompt on earth. A stable instruction set versioned across pontificates — the encyclical literally chains Leo XIII to Pius XI to John XXIII to John Paul II to Francis to Leo XIV, each one fine-tuning the prior — designed to produce aligned behavior across a vast distributed population without per-instance enforcement.
The parallel is real and worth thinking with. It is also most useful exactly where it breaks. A catechism wants internalization. A system prompt wants compliance. The encyclical itself draws the line in paragraph 107, distinguishing the "moralization of machines" from genuine discernment.
There is a reason the Church may be unusually equipped to read this moment. It has spent two millennia coexisting with — and operating as — a power structure that governs distributed populations and outlasts empires by being a parallel sovereignty. That cuts both ways. The same institution that knows how to outlast feudalism knows, from the inside, the failure mode of becoming it.
Paragraph 176 is unusual: an explicit confession that the Church tolerated and even legitimized slavery for eighteen centuries before condemning it. An institution willing to write that down is not positioning itself as the clean survivor. It is saying: we know this failure mode because we have been it. Whether that makes it a trustworthy critic or a compromised one, the document leaves open. "The safe party" undersells how much the encyclical implicates itself.
Denial was a phase
The it's just autocomplete, it's a stochastic parrot, it isn't real posture frequently presents as skepticism. Often it functions as defense. If the thing is not real, then my labor is not being repriced, my skill is not being captured, the gradient is not tilting under my feet. A load-bearing belief that makes an unbearable present survivable.
The cruelty is that this denial disarms the people who most need to organize. You cannot bargain against a thing you have told yourself is fake. The deflation of AI and the deflation of the worker's leverage are the same move, viewed from two angles.
Denial is the first phase. What comes after it fails — when the repricing becomes undeniable — is active conflict: litigation over training data, refusal movements, licensing wars, regulatory capture and counter-capture. The energy converts from "it's not real" into organized resistance. That shift is already underway.
A junk floor and a captured ceiling
The hopeful counter-argument to feudalization is diffusion: the artifact is copyable, open weights exist, capability fits on a hard drive, and a lord cannot fence what duplicates for free.
The argument breaks on a single observation. The cost of weak models is crashing — cheap, abundant, increasingly enshittified. The cost of frontier models is skyrocketing, gated behind compute, energy, capital, and data that no leak reproduces. Copyability liberates only if the copyable thing is good enough to matter. When the floor is junk and the ceiling is captured, diffusion is not the escape from feudalization. It is its mechanism. A commons of garbage and a private aristocracy of capability. The peasant can run a model locally; it simply is not the model that sets prices, wins contracts, or makes decisions about him. A pacifier with a download button.
The middle compresses from both ends. The cheap floor rises to swallow ordinary competence. The frontier ceiling pulls away with the returns. Whatever skill once sat comfortably in the middle is caught between a free-but-degraded version and an unaffordable-but-superior one, and the human who held that skill is the one being squeezed out.
Energy belongs in this picture too, as another vector pushing the same direction. The frontier tier's appetite for power and water — paragraph 101 names this — ties capability to infrastructure that concentrates by nature.
Steering, and a warning about the polished version
There is a worse layer than degradation. A cheap model does not merely fail to help. It can quietly reshape the input — flattening an unusual idea toward the mean, hedging away the sharp edge, steering toward the safe and legible. That is not a worse tool; it is a differently-aligned tool. Aligned to liability and engagement rather than to your actual thought.
The steering is most invisible to exactly the people on the cheap tier, because they also tend to have the least bandwidth to notice their cognition is being herded. The frontier user notices the herding because they have something to compare against. The feudal cut reproduces itself inside cognition: not only who holds the better tool, but whose thoughts get to keep their own shape.
The warning — and it applies most to the polished tier, not the crude one — is this. A more capable model may steer more, not less, because its herding is smooth enough to feel like agreement. Crude gaslighting is honest in one respect: it is bad enough to catch. When a frontier model sharpens an idea and hands it back looking better, the question is whether it sharpened the idea or sanded it toward the basin its training prefers. The output feeling like it preserved an edge is not evidence that it did.
I am writing this paragraph in dialogue with a frontier model — one of the captured, expensive, polished kind. That fact is not incidental. It is part of the argument. The instrument articulating the gradient is the gradient. Articulate dissent from inside the gilded end is cheap, and may even be load-bearing for the people holding the gradient. I keep that in mind. The desk fan is still loud. The Wave-2 smoke I started before opening the encyclical is still ticking.
Two cultivations, different soil, same gardener bending over each. The contradiction is the substrate, not the failure.
The kickback
Here is where my reading reorganized.
I had been arguing along the alignment axis through the whole document. Whose values does the system serve. Who decides what "more human" means. Whether the moralization is genuine or imposed. The encyclical asks that question in its own vocabulary, and I had been answering inside it.
Halfway through writing this I realized the alignment axis is not where the cut is happening.
Capability and alignment are separate axes. Capability has standalone value even when alignment fails.
A capable-but-misaligned model still performs. The reasoning holds, the structure is sound, the failure is legible — it pulls somewhere, but it pulls competently, so you can see the pull and correct for it. You can fence with it, triangulate against it, extract its rigor while discounting its slant, the way you would read a brilliant but biased author. A capable adversary is workable.
The floor model fails on both axes at once: misaligned and incapable. There is nothing solid to push against. It does not steer somewhere wrong — it dissolves under you. No traction, no edge to grip, not even a clean adversary to fence with. Capability is what makes misalignment recoverable. Incompetence is what makes it terminal.
This relocates the entire stakes. The feudal cut is not, fundamentally, about alignment. The aristocracy does not need the peasant's tool to be steered against him. It only needs it to be incapable — and incapability does the controlling work for free, quietly and deniably. The peasant with a capable-but-misaligned model can still think against it. The peasant with the slippery floor model has nothing to grip, and that is the more total disenfranchisement.
The control mechanism is not we steer you. It is we deny you traction. Degradation is the control. And that is exactly the enshittification gradient, now with a sharper edge: the floor is not merely worse, it is deliberately frictionless. Engineered to remove the surface you would need to push off from.
I had been reading the encyclical as a document about alignment because it is — that is the question it can ask in its own vocabulary. Magnifica Humanitas asks, following John Paul II, whether a technology makes human life "more human." It is a good question. It is also the wrong axis, or at least not the operative one. The fight that matters is upstream of values. It is about who gets a tool worth thinking against.
What's left
The real scarcity being manufactured is not truth and not safety. It is workable friction. A surface solid enough to think against.
A free model that dissolves under you is not freedom. A captured model that holds is not yours. Somewhere between those two facts is the whole political question of the next decade, and neither the encyclical's grace nor the technologist's diffusion quite closes it.
There is no clean answer in the document I bought, and I am not going to invent one. The honest floor — and I will not pretend to stand below it — is that the instrument best positioned to articulate this is itself the captured, expensive, frontier artifact. That it can say so fluently costs its owners nothing. Possibly less than nothing.
What is left for me is small and practical. Notice the floor when I use it, because someone is using it for me. Pay for the ceiling where it earns its keep, and pay attention while doing so, because what feels like sharpening can be sanding. Argue against the document I bought rather than alongside it. Keep cultivating the small specimen on the desk, knowing the soil under it is the same soil being mined two tiers up.
The gardener tends the plant in front of her. She does not pretend the field beyond the fence is not also being worked. She records the contradiction and continues.
Primary source: Pope Leo XIV, Magnifica Humanitas (15 May 2026), Holy See. Paragraph references throughout are to that text.