What is it For?

The question of teleology is one our vocabulary for technology is woefully unfit to ask. We know how to talk about whether a system works—about scale, optimization, deployment, adoption, alignment, reliability, safety, robustness, latency, and throughput. We know how to ask whether the model is multimodal, whether the platform is general-purpose, and whether the human remains “in the loop.” We know how to say that a tool increases productivity, reduces friction, expands capability, closes the gap, or sharpens the edge. What we do not know how to say, at least not in any language that sounds public and adult and available to more than a few philosophers and priests, is what the thing is for. Not what it does or what it enables. What it is for.

Even ethical language, once it is imported into the technology sector, tends to be translated at once into engineering language: “bias metrics,” “red-teaming,” “safeguards,” “provenance,” “thresholds,” “audits.” These things matter, but they belong to the management of means. They do not tell us whether the system is ordered toward anything worth wanting.

This is part of what makes Gideon Lewis-Kraus’s April 15 piece in The New Yorker, “How Project Maven Put A.I. Into the Kill Chain,” so clarifying. The essay does something more unsettling than merely summarize Katrina Manson’s book on the Pentagon’s automation project. It begins with the lurid, headline-ready question of whether Anthropic’s Claude had any role in a military operation, then unexpectedly turns that question inside out. The real story, Lewis-Kraus suggests, is not a chatbot’s cameo in war but the longer construction of Project Maven itself: a Palantir-supplied system meant to collect, analyze, and streamline intelligence, compressing the distance between seeing and striking. Through Manson’s reporting on Drew Cukor, the Marine officer who pushed the project forward, Lewis-Kraus traces how something described as an aid to analysis hardened into part of an automated targeting architecture—one in which software inherits, accelerates, and conceals institutional decisions about violence.

The sentence that matters most in the essay arrives later, and Lewis-Kraus presents it as a Weberian problem: “Bureaucracies are efficient, but they cannot determine what ends our efficiency ought to serve.” He gets there by way of Kevin Baker’s argument that the glamor of A.I. can distract from the fact that procedure, however rational, cannot answer the prior and more basic question. Efficient for what, and to what end?

Criticism of technology feels reductive because the argument always collapses into two familiar positions, one proclaiming the tools powerful, inevitable, and potentially beneficial, the other warning that they are dangerous, destabilizing, and insufficiently regulated. Both may be correct. Yet both often remain trapped inside the same grammar. They quarrel over rate, risk, and control, asking how fast, how accurate, how scalable, and how safe. They ask who should supervise, what standards should apply, and what safeguards can be installed. But the decisive question comes earlier and sounds almost childish beside them: Why do we want this at all? What human end does it serve? What conception of the person, polity, neighbor, enemy, citizen, child, has already been smuggled into the machine before anyone begins to argue about guardrails?

The modern answer is that the machine has no end of its own. It is only a tool, taking its purpose from the hand that uses it. This is meant to reassure us, because it places moral responsibility somewhere else—with the legislature, consumer, or battlefield commander. But it is false. Tools are made to favor one action over another. Their interfaces hasten certain judgments and deter others. Their defaults assign authority. Their metrics teach institutions what to prize. A search engine is not neutral because it answers many questions, no more than a social network is neutral because it hosts many opinions. A targeting platform is not neutral because, somewhere in the chain, a human being still clicks “accept.” The shape of the instrument is already an argument about the direction in which we mean to go.

War makes this impossible to ignore, because war strips the euphemism from the room. In consumer technology, teleology hides behind convenience: a recommendation engine is for “engagement”; a delivery app is for “efficiency”; a navigation system is for “frictionless experience.” These phrases are ugly, but they are at least bloodless. In war, the old fact returns. The system is for finding, sorting, selecting, and killing at speed. Lewis-Kraus notes, following Manson, that in Maven the path from target identification to target destruction can be reduced to four clicks. This sounds like a technical detail, which is precisely the problem: It is a moral decision made to pass as workflow.

The clean screen is part of the problem. It gives us white dots, legible interfaces, and dashboards. What it removes is not only mess but encounter. The old bureaucratic drag that technologists love to mock—the forms, contradictory reports, waiting, committee review, second-guessing, the chance for a timid person somewhere in the chain to say “not yet” or “not this one”—may indeed be inefficient. It may also be where conscience sneaks in. Lewis-Kraus quotes Baker’s argument that friction is where judgment forms. Compress time enough and you do not eliminate uncertainty; you merely lose the interval in which anyone can feel its weight.

A bureaucratic order can be exquisitely competent at carrying out a mission whose legitimacy it cannot itself examine. It can reduce arbitrariness and still magnify horror, and it can standardize procedure and still leave untouched the prior catastrophe: that the shared end has not been argued, or has been argued only in the coarse language of necessity, deterrence, competition, superiority, and national interest. The difficulty is not just that we have immoral ends but that our dominant public language for technology scarcely allows ends to appear as ends (they arrive already disguised as imperatives).

One notices this in the word “alignment,” which has come to do a great deal of work. Alignment with whom? With what? With user intent? Investor interest? State policy? Existing law? Military objectives? Human flourishing? The word assumes an objective waiting somewhere offstage, stable and legitimate, to which the machine need only be adjusted. But that objective is exactly what must be contested. A perfectly aligned system can be a perfectly obedient instrument of an unexamined purpose. To say that a model is safe because it follows instructions is like praising a missile for good manners!

We therefore need more than regulation in the narrow sense, though we certainly need that. We need a public vocabulary capable of handling teleology without blushing or dissolving into cant. We need words such as purpose, limit, legitimacy, proportion, dignity, and enough. We need to recover the difference between what can be done and what should not be built into the ordinary structure of action. We need ways to ask not only whether a technology is accurate but what habits of thought it trains; not only who controls it but what forms of life it presupposes; not only whether there is oversight but whether the institution using it has become addicted to speed, abstraction, and distance. We need words for the goods that are not commensurable with throughput.

Shared vocabulary matters because the problem is shared. It is not enough for one engineer, colonel, ethicist, or dissenter in a conference room to entertain private misgivings. Procurement has its own momentum, and bureaucracy has its own appetite. Once a system becomes infrastructure, individual conscience arrives too late. What is needed is language sturdy enough to circulate before the contract is signed and before the system becomes normal. A polity that cannot talk about ends will always be governed by people who talk only about means. And then the only public arguments left will concern efficiency, security, competitiveness, and resilience. And what the technology is for will be decided after the point at which anyone can object.

There is also a deeper reason this vocabulary matters. Teleology is not merely about final causes in some academic sense. It is about what kind of creatures we take ourselves to be. The machines we build are mirrors with ambitions, revealing which human capacities we find cumbersome and which we find sacred. If we build systems above all to remove hesitation, ambiguity, dependence, and the burden of interpretation, then we are saying that we increasingly regard judgment as drag, conscience as latency, and mortality itself as a design flaw.

And it is here that the older literature begins to seem less quaint than our contemporaries do. The central terror in the myth of dangerous creation was never simply that matter could move; it was that human beings, intoxicated by mastery, would set loose powers whose purposes they had never learned to name. We are very good now at speaking about emergence, capability, autonomy, and disruption. We are not very good at speaking about hubris, sacrilege, or the distortion of ends. Those words sound melodramatic, until one looks at a system built to make killing scalable and hears it described in the flat idiom of product design.

The task, then, is not to abandon technical language but to force it to answer to another kind of language: a shared vocabulary for ends, not in philosophy class but in product design, where abstractions become systems and systems become habits. Means must answer to ends, and governance must answer to some prior account of the human good. Without that order of questions, public debate becomes little more than an argument over how to steer a machine whose destination no one wishes to name. That is how whole societies slide into moral sleep—not because cruelty was ever argued for plainly, but because it was translated into process.

So let the first question be the simplest one, the one that sounds almost naive and is therefore hardest to ask in rooms full of experts: What is it for? Not what can it do or how soon can it ship. Not whether it keeps a human hand somewhere on the apparatus. What is it for? Until we can answer that in a language ordinary people can recognize as their own, technology will go on being discussed as though it were fate, or weather, or some impersonal force moving over us from elsewhere, when in fact it is made, decided, purchased, and arranged. And once what is made enters war, the cost of that evasiveness is no longer abstract. The screen and the language might be shiny and clean. The dead are not.