aligning without discussing is imposing

the technical problem of how to align an artificial intelligence with human values is not a technical problem. it is a problem of jurisdiction disguised as an engineering problem. pope leo xiv published on may 25, 2026 the first encyclical of his pontificate, magnifica humanitas, and between paragraphs 95 and 111 sits the sharpest critique of the year of the concept. the industry that claims to work on the safety of these systems chose not to read it that way, or not to read it.

the document is already well-known news. what the press covered were the headlines: "the pope calls for disarming artificial intelligence", "warns about the effects on minors", "warns that ai deskills workers". all of that is in the text. but the most piercing substance is elsewhere, in the conceptual apparatus the pope builds to talk about the technical problem the industry calls alignment.

the encyclical, in paragraph 107, says the following:

"we cannot limit ourselves to invoking the moralization of the machine, the so-called 'alignment' of ai with human values, without having the courage to add a further condition: the possibility of discussing the ethical code to be used, subjecting it to criteria of shared social justice. otherwise, whoever controls the ai will impose their own moral vision, which will become the invisible infrastructure of the systems. a more moral ai would be of no use if this morality is decided by a few."

that is what to look at.

the technical problem that is not technical

the alignment problem, in the sector's dominant conversation, is posed like this: we have systems that learn to produce answers by optimizing objective functions, and we want to make sure those objective functions correspond to what human beings actually value. the problem becomes acute when the systems are capable enough to make decisions in open contexts: hiring, credit, diagnosis, legal assistance, content generation. if the system pursues a poorly specified version of "what we want", it produces harms that are hard to reverse.

the most serious companies in the sector devote years and billions to this problem. entire lines of research exist to refine training procedures so that models respond in line with a declared set of values. one of those companies called its procedure constitutional ai: an explicit set of principles against which the model evaluates itself during training. the word constitution is not metaphorical. it is deliberate. it proposes, in effect, a jurisdiction.

the pope's observation is not against the technical procedure. it is about where the political question ends up once the procedure is applied. he continues, in the same paragraph:

"a more present politics is needed, capable of slowing down where everything accelerates and of protecting the spaces in which communities can keep participating and questioning themselves."

the idea, translated without liturgy: if you align a system with values that you define, without public discussion about what those values should be, you did not solve a problem. you displaced it to a layer where it can no longer be discussed. the system's morality becomes infrastructure, in the literal sense: what is underneath, what holds things up, what is not seen.

that is a structural critique, not an aesthetic one. and it is the one the public conversation about artificial intelligence safety systematically avoids.

the epistemic concession

before getting to the critique of alignment, the pope concedes something in paragraph 98 that the industry downplays when it can:

"modern artificial intelligences are more 'grown' than 'built': developers do not directly design every detail, but create an architecture on which the ai 'grows'. as a consequence, the fundamental scientific aspects, such as the internal representations and computational processes of these systems, remain unknown."

it is worth pausing on the sentence. it is academically correct. any serious researcher in mechanistic interpretability would sign it. what is remarkable is that it is in a document of the magisterium, signed by the pope, presented in the new synod hall on may 25. the acknowledgment of epistemic uncertainty about internal representations, said from rome, has a specific political function: it pulls the argumentative ground out from under anyone who says "we know what we are deploying".

the concession matters because the critique in paragraph 107 rests on it. if the systems are grown and nobody fully understands how they process information internally, the alignment procedure operates on a partially opaque box. the values that are "infused" into the system are not reviewable the way a law is reviewed. they are reviewable the way a biological hypothesis is reviewed: with external evidence, tests, feedback. they are, in other words, executable before being deliberable.

that changes the nature of the question. the decision about values is not an act prior to implementation. it is an act ratified by implementation. and implementation, today, is in the hands of a handful of private companies headquartered in two countries.

the cognitive monopolies

paragraph 108 names the asymmetry without euphemisms:

"as with every great technological advance, ai tends to increase above all the power of those who already have economic resources, expertise and access to data. in light of the common good and the universal destination of goods, this phenomenon raises serious concern: small, very influential groups can steer information and consumption, condition democratic processes and influence economic dynamics for their own benefit, contradicting social justice and solidarity among peoples."

and, one line later, it proposes something concrete:

"the ownership of data cannot be entrusted only to the private sector; it must be regulated. data are the fruit of the contribution of many and cannot be sold or entrusted to a few."

that is, in catholic language, collective ownership of the cognitive raw material of artificial intelligence. in social doctrine, the principle is called the universal destination of goods, and it has historically been applied to land and water. applying it to training data is a deliberate extension, and it breaks with the market logic that assumes whoever collected the data has the right to use it to build the models that later determine conditions of access to credit, employment and services.

the next paragraph, 109, completes the move:

"in a world where a few actors concentrate data, computing capital and normative capacity, to speak of the common good means unmasking this new epistemic, economic and political asymmetry, naming the new monopolies of ai."

"naming the new monopolies" is the phrase. not regulating, not auditing, not certifying: naming. the encyclical is asking, in plain terms, that public discourse stop treating these companies as neutral infrastructure and start treating them as what they are.

the verb the safety industry does not use

in paragraph 110, leo xiv underlines a word and marks it as his own:

"i would like, finally, to use a word that is very important to me: 'disarm'. disarming ai means removing it from the logic of the arms race, which today is no longer only military but economic and cognitive. it is the race for the most effective algorithm and the largest data bank, in order to consolidate a geopolitical or commercial advantage over everyone else. to disarm means breaking this equivalence between technological power and the right to govern."

when a pope underlines a verb, it is deliberate. disarming is not the same as regulating. regulating assumes a legitimate infrastructure whose excesses get put in order. disarming assumes an infrastructure that is, in itself, part of an arms race, and that must be removed from that race before it can be used.

the artificial intelligence safety sector uses a vocabulary of its own: fine-tuning, evals, red teams, benchmarks, review boards. all of that vocabulary assumes the problem is one of adjustment, not of a race. the encyclical names that the race itself is the problem, and that adjustment, without disarmament, is optimization inside an armed dynamic.

the paragraph closes with a sentence worth reading slowly:

"that is why regulating it is not enough; it is necessary to disarm it and make it welcoming."

welcoming. that word is outside the sector's lexicon. it does not appear in any usage policy, in any corporate document, in any known regulatory agenda. it is, however, the word the pope chooses to describe the state to which artificial intelligence should be brought.

who decides the values that run in montevideo

paragraph 95 had already said what matters for anyone who lives outside san francisco or seattle:

"control over platforms, infrastructures, data and computing capacity is not the prerogative of states, but of large economic and technological actors who, in fact, determine the conditions of access, the rules of visibility and the very possibilities of participation."

and paragraph 5 of the introduction had said it earlier:

"in the past, it was mainly states that drove and steered innovation. today, by contrast, the main engines of development are private actors, often transnational, endowed with resources and capacity for action superior to those of many governments. technological power thus acquires an unprecedented face, predominantly 'private', and for that reason even harder to discern, govern and orient toward the common good."

the systems that make automated decisions about credit, hiring, content moderation and clinical assistance in uruguay were not aligned with values deliberated in uruguay. they were aligned, at best, with values deliberated in the ethics boards of north american companies. at worst, they were not aligned with anything explicit and reflect the statistical biases of the data they were trained on, data that was not collected here either.

the question is not whether those values are good or bad. the question is jurisdictional. when a system trained in california decides that a credit application in montevideo is high risk, or that a resume mentioning udelar is less competitive than one mentioning mit, or that a word in rioplatense spanish signals toxicity in a forum, we are not facing an artificial intelligence safety problem. we are facing a problem of moral sovereignty. the system's ethical code was decided, in the terms of paragraph 107, "by a few". and those few do not live here.

the pope says it more simply in paragraph 111, addressing directly those who build the systems:

"developers therefore carry an important ethical and spiritual weight, since every design choice expresses a vision of humanity."

every design choice. every decision to label a data point, define a parameter, configure a default behavior. each of those choices, distributed across thousands of people in a few companies, is building the invisible infrastructure that millions of people elsewhere interact with without having taken part in its design.

the reframe

the public conversation about artificial intelligence safety is badly framed. it assumes the problem is building systems that behave the way we want. the pope, without proposing a technical solution, makes a simpler and more uncomfortable observation: any system, however well behaved, executes the values of whoever builds it. the urgent question is not how we align. it is who decides which values count as alignment.

that question is not answered with better training procedures. it is answered with institutions that do not exist yet. and as long as those institutions do not exist, the decisions executed in systems used in buenos aires, lima, lagos and manila keep being made in offices where nobody from those places ever sat.

the most interesting document published this year on the alignment problem was written by a canonist born in chicago, who spent most of his career in peru, citing saint augustine and hannah arendt. it did not come out at a machine learning conference. it came out on may 25, in the new synod hall, signed a few days before the 135th anniversary of the encyclical that in 1891 first named the social question. it is 110 pages long. i am sure the sector that claims to work on the problem has not yet read it in full or dismissed it as an irrelevant text because of where it comes from.

it is worth reading.

sources:

[1] leo xiv. magnifica humanitas, carta encíclica sobre la custodia de la persona humana en el tiempo de la inteligencia artificial. the holy see, may 15, 2026 (published on the 25th). https://www.vatican.va/content/leo-xiv/es/encyclicals/documents/20260515-magnifica-humanitas.html

[2] dicastery for the doctrine of the faith. public presentation of the encyclical, new synod hall, may 25, 2026. remarks by cardinal víctor manuel fernández.

[3] bai, yuntao et al. constitutional ai: harmlessness from ai feedback. arxiv:2212.08073, december 2022.

[4] francis. laudato si', §§101-114, on the technocratic paradigm. cited by magnifica humanitas §92.

[5] john paul ii. centesimus annus, §§31-48, on the collective ownership of goods. cited by magnifica humanitas §108.