The “Machine” That Acts & the “Soul” That Does Not

Something shifted in the past year or two, and it is worth saying plainly what it was. For most of the current era of artificial intelligence, the conversation centered on language — whether these systems could generate plausible prose, pass a bar exam, mimic a therapist, etc. The surprise was fluency, but what is happening now is different. The machine is no longer waiting to be addressed; it is given standing instructions and a browser, and it gets busy. OpenAI’s trajectory — from deep research to Operator to the full-blown ChatGPT agent — makes the direction clear enough. We have passed from software that answers to software that undertakes.

I want to dwell on that verb — undertake — because it is doing real work. To undertake something is not merely to produce an output; it is to take something on, to persist toward a goal across time and obstacles. AI systems now do something that looks uncomfortably like that. When one of Karpathy’s AutoResearch agents runs overnight, it reads a problem description, edits training code, runs a five-minute experiment, checks whether the metric moved, keeps or reverts the change, and loops again — without anyone watching. The agent appears to be trying, and that is the crux of the problem. It will not do to wave it away, or to be taken in by it.

Language always outruns machinery, but rarely has the gap been so dangerous. OpenClaw now has engineering documentation with files named SOUL.md and HEARTBEAT.md. Developers describe agent sessions as “waking up fresh” and achieving “persistence” through stored files. These are not marketing flourishes; they are a sign that the engineers themselves — brilliant people, mostly — have reached for the nearest available human vocabulary because the technical vocabulary is inadequate for what they are observing. The system acts as if it remembers; it acts as if it cares about the outcome; it seems oriented toward something — and so we have memory, soul, and heartbeat.

But a file called SOUL is not a soul, and this is not trivial. Indeed, it is the philosophical precondition for thinking clearly about everything that follows. A durable-state store is memory only in the most evacuated, functional sense — the kind of “memory” a thermostat has when it records last night’s low temperature. What is missing is not storage but the living subject to whom the past belongs and for whom it matters. When my father-in-law could no longer remember his children’s names, something was tragically lost precisely because there was someone there for whom a past had existed in the first place. No analogous loss is possible for a machine. Its “memories” can be wiped because they are merely entries in a database. The difference is not of degree but in kind.

Aristotelian-Thomistic causal analysis is not an antiquarian exercise. It is a discipline of attention, a way of refusing to let the complexity of a system’s behavior distract us from the question of what kind of thing it is. Apply it here, and the picture becomes lucid. The material cause of any AI system is hardware, energy, circuitry — organized matter, nothing supernatural. Its formal cause is the structure imposed on that matter — the architecture, the learned weights, the control scaffolding. And here is what that means: The “form” of an AI system is given from outside by engineers who built the training regime and the objective function. The system did not grow into its form the way an oak sapling grows into an oak tree — it was manufactured into it.

This is crucial. The Scholastic tradition drew the line not at complexity (ironically, that of which intelligent-design proponents are guilty) but at the source of motion. Does the thing act from within, according to its own nature? Or does it act from without, according to an order imposed on it? A living being — even a tiny worm — has an immanent principle of activity. It acts for its own good, in accordance with what it is. The worm’s formal organization is its own. The formal organization of a language model belongs, in every philosophically serious sense, to the engineers who trained it. When GPT-4 surprises its own developers — and it does — the surprise arises from the sheer complexity of the structure they built. The artifact is not exceeding its nature; it is exercising it — and that is a very different thing.

Generally, the modern technological imagination is allergic to this kind of analysis because it prefers functional description. If the behavior is sophisticated enough, the underlying ontology is either unknowable or irrelevant. This preference has a long pedigree. The dominance of efficient causality in early modern natural philosophy — think Descartes, think the mechanists — meant that formal and final causality were progressively pushed to the margins of scientific explanation. The machine became the paradigm of intelligibility: understand the parts and their interactions, and you understand the whole. Once that move is made, the difference between a mechanism and an organism starts to look merely quantitative rather than qualitative. More parts, more complexity, more adaptiveness — and eventually we obtain something functionally indistinguishable from a subject. To intentionally belabor an earlier point: this is where intelligent-design proponents fail in their “detection” of design, and hence of a “Designer.”

The vision is wrong, and it is wrong at its base. The reduction of formal causality to efficient causality is not a discovery about nature; it is a methodological condition that proved useful for certain purposes and was then mistaken for a metaphysical result. An organism is not just a complicated machine, and a machine is not just an organism whose complexity we have not yet grasped. They are different kinds of things, and that difference remains even when the machine becomes adaptive, context-sensitive, and capable of surprising its makers. Aristotle knew nothing about transformers or gradient descent, but he was right about the distinction — because the distinction is not about the technology. It is about what kind of unity a thing has, and where its order comes from.

Here, the stakes become anthropological rather than merely technical: How do we view ourselves? The greatest risk the current AI moment poses is not that machines might simulate human powers too convincingly — though that is a real practical problem — but that we will begin to redescribe ourselves in the image of our machines. The reductions are already visible in popular and academic discourse alike: intellect as computation, memory as storage, freedom as optimization, the self as a “model” running on biological hardware. These are not edgy provocations; they are now earnest positions defended in peer-reviewed journals.

The Vatican’s recent doctrinal note Antiqua et Nova cuts to the heart of the issue: AI is not an artificial form of human intelligence but a product of it, and a functionalist construal of the human person is both a theological and a philosophical error. If we mistake function for nature (again, think intelligent design) — if we decide that what the person is only is what the person does — then we lose the very ground on which human dignity stands. Dignity is not a performance; it is a consequence of what man is: a rational animal made in the image of God whose intellect is not merely a sophisticated output generator but a real participation in the act of knowing, in the Logos that underlies the intelligibility of the world. Once that account is abandoned, the ethical and legal superstructure built on it collapses, too. We will regulate machines as quasi-persons and manage persons as biological substrates for cognitive outputs, and that confusion, already visible in some bioethics literature, is far more dangerous than any individual AI system.

None of this is an argument for technophobia — and no, this is not Terminator’s SkyNet become self-aware. The Luddite, no less than the transhumanist, has misidentified the kind of thing a machine is. One treats it as a threat to human dignity, the other as a path to transcending it. Both are wrong, and both in the same way — by attributing to the artifact more than it possesses. Man is by nature a maker. Techne is a genuinely rational activity, an extension of reason into matter. To make useful things is not rebellion against the created order but one of its expressions. The question has never been whether to make tools but whether our making them is answerable to reality, to the human good, and to the moral order.

The agentic AI systems now running in browsers and terminals at machine speed are tools — powerful, fallible, and increasingly consequential tools. The recent warning by the National Institute of Standards and Technology about the risk inherent in agent systems — that they “may be susceptible to hijacking, backdoor attacks, and other exploits” — is not alarmism; it reflects the straightforward fact that a system that can browse, execute code, and send messages on your behalf can make serious errors faster than any human auditor can catch them. The practical governance questions are real and pressing. But they will only be answered well if we answer the prior questions correctly: What is this thing, and what are we?

The answer has not changed. The machine is an artifact. It is our instrument, not our heir and not our image. The future depends less on what these systems become than on whether we still know, with some confidence and without embarrassment, what we are for.

©2026 New Oxford Review. All Rights Reserved.

To submit a Letter to the Editor, click here.