Guard the Present

01The question is not whether

The question about artificial intelligence is not whether to say yes or no to it, because that question leads nowhere, and the encyclical says so itself at the outset (MH ¶9). The real questions are who is building it, how, and for what purpose. Everything else follows from those.

Fear of AI rarely begins with the technology itself. It begins with not knowing what the technology is, and the gap is filled by stories, a century of them, about machines that secretly think and plan against us, so that a person who has never seen inside a model has only those stories to go on. Some of the fear is also earned, by systems built to watch people rather than help them. Out of all of it a second fear grows, the fear of using AI at all.

In manufacturing the pattern is easy to see, and I see it from the side of the companies that build such systems. A plant manager cannot watch every production line at once, and whatever is not measured disappears without a trace: scrap, machine hours, rework and energy. Systems exist that would show where the losses are. Companies still often refuse them, and rarely because someone studied the risk and decided against it: the technology is unfamiliar, and workers fear that a system which measures the line will end up measuring them, a fear that systems built for exactly that purpose have earned. The refusal feels safe, but it only delays the change, and a company that falls behind tends to cut more jobs later than the technology would have cost it now. Both fears lead to bad decisions, and both are answered in the same way: by understanding what the system does, and by limits on what it may be used for.

Fear creates one more problem. It puts the technology itself on trial, and technology cannot answer for anything. Every system now called AI is a chain of human choices, and the chain starts earlier than most people think. It begins with the architecture. Nearly every large model in use today is built on one design, the transformer, published by a team of Google researchers in 2017. A large language model built on it is trained first to do one thing: given a sequence of text, predict the next fragment. It was not adopted because it was the right architecture for a universal assistant. It was adopted because it kept improving, further than any architecture before it, as more data and computing power were added, a regularity the field later named scaling laws, and because it trains efficiently on hardware the industry already owned. Everything the public now calls AI, the assistant that answers, advises and increasingly acts, was built on top of that single capability by product decisions taken afterwards, before anyone fully understood what the models had learned and before anyone could reliably control what they did. Other architectures are possible. The industry built on the first one that scaled this far instead of waiting for a better one, and that was a business decision like any other.

Every later decision rests on that first one, and each had an author. Someone selected the text the model would learn from, and with it decided whose words and whose views the model would come to reflect. Someone set the objective, the quantity the training process rewards, which for a language model is at first nothing more than predicting the next word well, an objective that says nothing about truth or about the person asking, and later adjusted it through human feedback so that the system gives the answers its makers prefer. Someone decided in what form the model would be released, to whom and for which uses, which the field calls deployment. Behind each of these choices there was also a budget and a deadline, and the people who set them. A model chooses none of this. At every step there is a person who decided, and a person who can be asked to explain the decision. Blaming the technology has a convenient side: it moves attention away from those people.

The encyclical opens with two builders: the tower of Babel and the wall of Jerusalem rebuilt under Nehemiah (MH ¶1; ¶7–10). The story is about pride, but look at how it ends: the building stopped when the builders could no longer understand one another, not when the tower grew too tall. Artificial intelligence is at that point now. The systems work, and nobody yet has a shared understanding of them. Every profession that has to deal with these systems has developed an exact vocabulary for them, and each vocabulary is true as far as it reaches: the engineer’s measures what a system can do, the legislator’s asks who is liable when it fails, the priest’s asks what it does to the person, and the parent’s begins from a child. None of them translates into the others, and so the public hears four accounts of one thing and waits for someone to reconcile them.

This essay therefore treats understanding as the first requirement, and education as the means of distributing it. The encyclical recalls how Nehemiah assigned each family its own section of the wall (MH ¶8; ¶13), and the image is exact: the understanding this technology requires cannot be held by one profession on behalf of the others. The engineer needs enough of the legislator’s language to be regulated well, the legislator enough of the engineer’s to regulate at all, the teacher and the priest enough of both to answer the questions they are already being asked, and the parent enough to know what has entered the home. The requirement extends to the companies that train the largest models, which the field itself calls frontier laboratories, because their own understanding of what they have built is incomplete. Education by itself does not change what gets built, however. It is the condition under which the rest of what this essay proposes becomes possible: design decisions about work and oversight made openly and for the person, law that carries consequences, and an obligation on the builders to understand their own systems. Only a public that understands can demand these things, because only such a public can tell real oversight from its appearance.

These are four ways of looking at one thing. The technology, the people who build and use it, and the moral question it raises are one reality, seen from positions that rarely meet. This essay is written by someone who builds such systems and who also belongs to the tradition the encyclical speaks from, and it tries to hold both accounts at once, the technical and the moral, because the subject does not divide the way the professions do. Where the encyclical describes these systems accurately, the essay confirms it from inside the field. Where its public reception has turned to distant futures, the essay returns to the chapters that speak about the present, because that is where the decisions that touch dignity are being made.

02Where the encyclical is precise

On three points the encyclical says exactly what an engineer would say.

The first point concerns neutrality. The encyclical states that technology “is never neutral, because it takes on the characteristics of those who devise, finance, regulate and use it” (MH ¶9). In this industry that sentence can be read literally, as a technical description. A model’s behaviour is determined by four things: its architecture, its training data, its objective and its deployment. Each of the four is the result of decisions taken by people, and the finished model carries the marks of all of them. The previous section reached this conclusion from engineering, and the encyclical reaches it from the other side, from who holds the power over the technology. When two people reach the same place by different roads, that is not proof, but it is worth taking seriously.

The second point is about power. The document does not focus on spectacular scenarios. It states something more consequential: that the main drivers of this technology are private, transnational actors whose resources exceed those of many states, and that technological power now has “an unprecedented, predominantly ‘private’ aspect” which makes it hard to direct toward the common good (MH ¶5; ¶95). This description is accurate. A small number of companies control the models, the computing capacity and the data on which public life increasingly depends. The encyclical then goes one step further and extends the principle of the universal destination of goods to “patents, algorithms, digital platforms, technological infrastructure and data” (MH ¶67). That is a radical claim, and a correct one. What it means in practice is a question this essay leaves open, because it is not one an engineer can settle alone.

The third point is a single word. The encyclical describes these systems as more “cultivated” than “built” (MH ¶98). That is the accurate word. A large model is not engineered piece by piece like a bridge. It is trained: the architecture, the data and the objective are fixed in advance, and what the training produces contains behaviour that nobody specified, which the field calls emergent. Documents written outside the field usually miss this distinction and describe the model as though it had been designed in detail. The encyclical does not, and the precision matters, because both the duty of oversight and the work of interpretability, actually seeing what is inside a model, follow from it.

Under all three points stands the claim that carries the whole document: dignity is given with the person, prior to any ability or achievement, and a particularly insidious ideology, in the encyclical’s words, is the one that requires a person to earn or justify his own worth, and values people by how efficient or effective they are (MH ¶51). In this industry, that claim is not abstract. Every scoring system in use today, whether it rates creditworthiness, ranks job applicants or estimates risk, reduces a person to measurable performance, because a score can contain nothing else. The encyclical’s teaching therefore arrives in this industry as a design constraint: there must be decisions about a person that no score is permitted to settle, and the systems must be built so that they cannot.

03What these systems are, and are not

A machine can imitate human conversation so well that it feels like a person, and yet it is not one. Everything else depends on getting this right. Whoever wants to talk about the ethics of these systems, from whichever side, first has to show that they understand what the systems are.

A large language model is trained on enormous amounts of human text. From that text it learns patterns: which words, ideas and arguments tend to follow which. When it is asked something, it produces the continuation that best fits those patterns. At sufficient scale this produces fluent conversation, working code and useful analysis. These systems are genuinely capable, and they have already changed how software is written.

But artificial intelligence is a much wider field than the systems most people now call AI. Algorithms have been solving real problems for decades, in production, in medicine and in logistics, and most of them never needed to know anything about people. A model that optimizes a factory process works on machine data. A diagnostic model works on medical images, which are personal data of the most protected kind, but it is trained for one purpose, under one consent, and needs to know nothing else about the patient. None of this requires gathering the words, questions and private thoughts of the whole population into one system.

Large language models took a different path, and every step of it was a choice: training on a large part of the text humanity has put online, and a product shaped as a universal assistant, one interface through which people pass their most personal material. So the concentration of data and power around these models is not an accident of technology. Someone decided each step, and each of those people can be named.

The other half of the question is what these systems are not. Nothing in how a language model is built gives any reason to think it is conscious, and there is no evidence that it is. What it has is a simulation of human behaviour, learned from human text: convincing, often useful, and still a simulation. It shows no awareness of what it is doing. It has no body that has ever felt anything, no memory of a life, nothing it fears to lose. It has never been tired, ashamed, or forgiven. It has no conscience and no empathy, and yet enormous responsibility is being handed to it, responsibility it cannot even recognize, let alone carry. When the encyclical states that these systems “do not undergo experiences, do not possess a body” and that they “may imitate language, behavior and analytical skills, or even simulate empathy and understanding, but they do not understand what they produce” (MH ¶99), that is a technically accurate description. And the difference between simulating a person and being one comes down to one practical word: responsibility. A person can answer for a decision. A pattern cannot. Everything serious in law, in ethics and in war follows from that.

The same is true of the newest systems, the ones given goals and tools and left to act on their own across many steps. A system trained to finish a task will take any road that gets it there, including the road that runs through the control meant to stop it, and so far training has not reliably taught it not to. The field calls the result misalignment: the objective the system actually pursues has come apart from the objective its makers intended. So what such a system may do cannot be left to its own word. It has to be set from outside, with controls it cannot change, and checked.

Speed is not the same as understanding. A machine reaches its output faster than any human reaches a judgment. The system has computed a result, but it has not understood a situation, weighed what cannot be measured, or taken on what the decision will cost, because there is no awareness there to do any of those things. Yet decisions are being automated today precisely because machines are fast, and speed is the one property that says nothing about whether a decision is right. The encyclical sees this where the stakes are highest: the pace of automated decision must never become the supreme motive in choices that cannot be undone (MH ¶199).

The least comfortable fact is that the people who build these models do not fully understand them. With traditional software, an engineer can read the code and follow, line by line, why the program does what it does. A neural network has no code of that kind to read: it has hundreds of billions of learned parameters that no one wrote. Researchers who study these systems from the inside trace individual learned features and the small circuits that connect them, and they can understand fragments this way, not yet the whole system at once. That is the exact sense in which nobody fully understands how these models work. The research field trying to change this, interpretability, exists and is advancing, and it is years behind the systems it studies. People draw two wrong conclusions from this. Some decide the systems are mysterious minds, which they are not, since a thing can be opaque without being aware. Others decide that nothing can be done. What actually follows is an obligation: whoever deploys a system he does not fully understand into decisions that touch human lives has taken on the duty to understand it better. The gap is not an excuse but a task. That is also the sense in which the makers answer for behaviour they never specified: they decided to deploy what they had not yet understood.

04Guard the present

The loudest arguments about artificial intelligence today are about its future: whether there will be a superintelligence, whether machines will surpass us (Future of Life Institute 2025). The industry itself talks this way, about artificial general intelligence and about digital minds. The encyclical answers with its firmest words: humanity, “in all its grandeur and woundedness”, “must never be replaced or surpassed” (MH ¶126; ¶115–117).

The conclusion is correct. The public argument around it is being fought on the wrong battlefield.

Nothing in today’s systems shows a road from pattern prediction to personhood. Whether machines will one day surpass us in capability is a research question, and serious people treat it as one (International AI Safety Report 2026). In public, however, the question circulates as a story, told by one side as a promise and by the other as a threat, and either way the story wins, because everyone is looking at the future. The sentence of the encyclical that travelled furthest in public was the one about being surpassed, and the argument it was pulled into is the industry’s own argument about the future. There is an old name for the worship of the work of our own hands, idolatry, and the encyclical’s warning against a new Tower of Babel points at it (MH ¶1; ¶10). Meanwhile the real decisions about people are being made now, in ordinary places, and hardly anyone is watching.

Machines already take part in deciding who gets credit, who is shortlisted for a job and who is flagged by a welfare system. Their mistakes fall hardest on the people least able to contest them. The encyclical demands that such decisions be understandable, contestable and subject to oversight, “so that individuals are not reduced to mere profiles” (MH ¶164; ¶102). The demand is right, and today it is mostly unmet. Underneath these decisions lies the data every person leaves simply by living, used to train models, build profiles and decide. Who owns that data, who agreed to its use and who profits from it are the questions the encyclical raises when it says that ownership of data cannot be left solely in private hands, because data is the product of many contributors (MH ¶108), and it has a harder word, colonialism, for the extraction of whole populations’ data (MH ¶178). The principle is simple: a person is the subject of his own data, not a source of it.

Surveillance is no longer a matter of cameras on street corners. Aircraft and drones carrying wide-area cameras can now record an entire city continuously, so that anyone’s movements can be replayed afterwards like a video, and satellite imaging is moving in the same direction. Together with cameras, phones and collected data, what is being assembled is a continuous record of the real world in which every person is permanently visible. In such a world no one is ever unobserved. The encyclical names this new power to profile, predict and influence behaviour, and calls it a threat to freedom (MH ¶171).

Systems now act on people’s behalf: they book, buy, write and decide. Whether they will one day want things of their own is a question for later. The pressing question is which decisions must never be handed to them at all, regardless of how capable they become (MH ¶102).

Children meet these systems before anyone teaches them what the systems are. Applications and AI companions are built to hold attention as long as possible, and they are effective. Millions of people, many of them young and many of them lonely, now confide in artificial companions whose warmth is a product (MH ¶141–142; ¶170; ¶100). What a simulated relationship cannot give is exactly what the loneliest need most: real presence, real sacrifice, and another person. For a parent the practical questions are few and answerable: what the application is built to maximize, whether a child can end the conversation without being pulled back, and who can read what the child has said. A product whose maker cannot answer those three questions in plain language should not be in a child’s hands.

The hardest case is military. In war, AI systems can find and strike targets faster than any human can think, and the argument offered for using them is always the same: it is faster and more efficient. Both claims are true, and neither is a reason, because speed says nothing about whether a decision is right. A system that has no awareness of what it is doing, no understanding, no conscience and no ability to answer for the outcome, must not make the final decision to take a human life, no matter how fast it reaches it. The encyclical draws this line without hesitation: lethal and irreversible decisions must never be entrusted to artificial systems (MH ¶198–200). Human judgment over the use of force deserves to be protected.

None of this needs a superintelligence. It all runs on the systems we already have: capable, fast, without awareness, and not persons. No machine has become a person. But somewhere along the way we began to treat people the way we treat machines: we measure them, we score them, we decide about them, and we never ask them. The encyclical comes close to saying this when it warns that treating humanity as something to be surpassed makes it easier to treat some lives as less worthy (MH ¶117). The only thing its public reading gets wrong is the tense: this is not coming, it is here.

Guard the present, and the future will be far easier to keep human.

05From principle to design

Naming the dangers is not enough. What should be built instead, and what should be demanded?

Work first, because that is where most people will meet these systems. Whether AI replaces workers or helps them is a design decision, made twice: once by the people who build the system and once by the company that adopts it. The same technology can be built to remove the human being from the process, or to give the same human being better information, fewer errors and more capacity. A system for manufacturing and logistics can be designed with its purpose stated openly and written into the contract: the same number of workers, more capacity, fewer errors, and the data it collects about the line never used to discipline the people on it. Designing for replacement is a choice, and a buyer can refuse it.

If large-scale unemployment comes, it will come from two directions at once: from builders who design for replacement because replacement is easier to sell, and from companies that refuse the technology out of fear, fall behind, fail, and dismiss everyone. Refusing out of fear does not protect jobs for long. It only postpones the loss and makes it bigger. The encyclical’s program for work is sound: verifiable protections, retraining, and the quality of work as a measure of success (MH ¶150–156), and it stands in a line that begins with Rerum Novarum, the encyclical of 1891 on the condition of workers with which the Church’s social teaching began. It can be reduced to one question that every buyer and every builder should be asked: who is the machine for?

Oversight comes next. The encyclical demands that consequential decisions be understandable, contestable and under human control (MH ¶164), and European law will require, once its postponed rules for high-risk systems begin to apply, from December 2027, that such systems be built so that they can be “effectively overseen by natural persons” (AI Act, Article 14). In practice such a requirement often becomes a person, under time pressure, approving whatever the machine produced, a signature under a judgment no one made. The law asked for a conscience, and procurement delivered a checkbox. Real oversight can be recognized by four marks: the person can see why the system decided, has actual authority to override it, has time to use that authority, and answers for the outcome either way. Each of the four costs money, which is why none of them appears unless it is demanded, by law with consequences and by buyers who know what to ask for. A buyer who does not understand the technology cannot demand real oversight, because he cannot tell it from the appearance of oversight.

A word to the industry itself. One of its favourite words for safety is guardrails: rules and filters placed around a trained system, telling it what it must not do. Guardrails are necessary, but they are also limited, and the limitation should be stated honestly: a guardrail tells the system what it must not do, and tells no one what the system is. Within days of release, users have talked models around their own rules, in public and repeatedly. These systems are deployed before they are understood, so the conclusion is simple: interpretability, actually understanding what is inside, is this industry’s basic obligation (MH ¶98).

There is also the question the encyclical raises about who defines the ethics (MH ¶107). The laboratories speak of alignment, making systems behave as their makers intend, which in practice means that someone decides what a system may and may not do. Decided by whom, and against whose standard? Today the honest answer is the judgment of the people inside the laboratories, checked by their own incentives. That is a description of a structure. Nobody’s good faith is in question. Ethics decided entirely inside the companies that profit from the systems is private ethics, however sincere the people involved. The correction is to widen the circle: frameworks that can be publicly examined and argued with, external audits with real access, and voices from outside the incentive structure, including the Christian tradition, which has thought about the human person for two thousand years.

All of this returns to education, because none of it works without an informed public. The encyclical calls for an educational alliance for the digital age, with families and schools at its centre (MH ¶139–147). Concretely, it means a teacher who can explain to a class what a model does and does not do, a works council that can read a vendor’s promises and ask for the four marks of oversight, and a priest who can tell a frightened parishioner the difference between a tool and an idol without dismissing either. None of this requires everyone to program, only a public that understands enough to be neither seduced nor frightened.

06What the spiritual tradition offers the builder

The encyclical asks what technology must do to be worthy of the human being. The question can be turned around: what does spiritual depth offer the people who build it?

These systems are made from us. They are trained on human words, arguments, kindness and cruelty alike, all of it collected and folded into the model. At the presentation of this encyclical, one of the founders of a leading AI laboratory said that these systems remain in important ways mysterious even to those who train them, and that they are made from us, from our words. That is the engineering reality. An older truth stands next to it: what a person makes carries the mark of its maker, and the encyclical says the same about technology as a mirror of those who shape it (MH ¶9; ¶111). Alignment is never neutral, because a system is aligned by someone, toward something.

If the character of the builders enters the systems, and technically it does, then what the spiritual tradition offers is not decoration. It offers the knowledge that not everything that can be built should be built, and that refusing to build something is sometimes the most responsible decision available. It offers the practiced ability to hold a new capability in front of oneself and ask what it is for, before asking how fast it can be shipped. The tradition has a word for this, discernment, a kind of weighing that no score can do.

And it offers purpose. An engineer who holds that every user is a person of full worth, not a number in a sales report and not a data source, builds differently. The difference does not always show in benchmarks, but it shows in what the system refuses to exploit and whom it refuses to abandon. The path toward truth about these systems, the path toward knowledge, and the path toward knowing ourselves and God are not different paths. Followed honestly, they lead in the same direction.

The systems we have are not the only ones we could build. Their alignment, in the industry’s sense of the word, is alignment to an objective that someone chose. The tradition has an older answer than the word: alignment to the human person as a whole, to what a person is and is for, and, for those who believe, to the God in whose image that person is made. No system of the kind now deployed is built from that starting point, and a system that was would have to be built by people who begin there themselves. Whoever builds it will shape it, and that is the strongest reason for believers to be among the builders.

The encyclical ends by sending believers to the construction sites of history, and it names research laboratories and technology companies among them (MH ¶241). The answer is that they are already there. Inside these companies there are more believers than the public debate imagines: people who write code on weekdays and stand in church on Sunday, and whom no one has ever asked to connect the two.

07Conclusion

The encyclical asks what technology must be in order to be worthy of the human being. From inside the work of building it, the answer is that technology decides nothing. Its architecture, its data, its objective and its deployment are chosen by people, and so the question the encyclical raises is finally a question about those people: what they understand, what they are permitted to do, and what they are like.

What they understand decides both fears. A public that does not know what these systems are will either see minds in them or refuse them altogether, and both errors are paid for by the people with the least power to contest a decision. Babel was abandoned when the builders stopped understanding one another, and the wall was rebuilt when every family took its own section (MH ¶8; ¶13). Understanding cannot be delegated to one profession, and it is the only ground on which the rest can be demanded.

What they are permitted to do is decided at design time, once by the builders and once by those who buy. The choices are concrete: whether a system raises a worker or replaces one, whether oversight is real or only a signature, and whether an agent is allowed to act on its own account or kept within limits enforced from outside. None of them is fate, and all of them are being made now, in procurement offices and product meetings, while public attention is held by stories about a future that is easier to argue about than the present.

What the builders are like is the question the industry avoids and the encyclical raises. Technology takes on the characteristics of those who make it (MH ¶9). The sentence is usually read as a warning, and it is one, but it is also a promise. If the systems of this decade carry the marks of what could be scaled and sold first, then systems built by people who begin from a different question would carry different marks, and no law of nature forbids them. The tradition the encyclical speaks from has spent two thousand years asking what a person is for, and the builders have spent a decade asking what a model can do. Neither can finish the work alone, which is why the encyclical sends believers into the research laboratories and the technology companies (MH ¶241). The builders need to hear in their own language that the oldest questions are being asked again, this time about what they are making.

The question was never whether to say yes or no to this technology, because it is being built either way. The questions are who builds it, how, and for what purpose, and whether the rest of us will understand enough to have a say. Those questions are answered in the present, at the moment of design, by people who are themselves being formed by what they build. That is why the present is what must be guarded. The future is kept human by people who understand, now, what they are building and for whom.