Draft preview · not yet published

The Archer and the Arrow

An arrow in flight is going somewhere. It has no idea where. The aim belongs to the archer.

Thomas Aquinas reaches for this image again and again in the Summa Theologiae. A thing can move toward an end in two ways, he says:

first, as a thing, moving itself to the end, as man; secondly, as a thing moved by another to the end, as an arrow tends to a determinate end through being moved by the archer who directs his action to the end. ST I–II q.1 a.2

He uses it once more in the last of his five arguments for the existence of God: whatever lacks intelligence cannot move toward an end unless something with knowledge directs it, “as the arrow is shot to its mark by the archer” ST I q.2 a.3.

In Aquinas's own argument the archer is God and the arrows are the things of nature. There were no computers when Aquinas was writing, so applying his image to a person who hands a task to a machine is my analogy, not his.

Whose aim is it?

Stands on argument

An AI system is something people make and put to work. Whatever end or goal it serves, someone set it to that. Its purposes are borrowed. Whether a machine thinks the way a person does is a separate question, and nothing in this picture depends on settling it.

Borrowed purpose does not mean the machine brings nothing of its own. Aquinas says an instrument has two actions. One comes from the person using it. The other is its own and comes from what it is: an axe cuts because it is sharp, and it makes a couch because a craftsman is using it. Then he adds the line that matters most here: “it is by cutting that it makes a couch” ST III q.62 a.1 ad 2. The craftsman's purpose arrives only through what the tool itself does.

A language model is a strange axe, unlike any tool we have made before. Its sharpness is everything training left in it: its skills, its habits, its leanings. And where an axe stops the moment the hand stops, a model keeps going. Once it has been given an end, it works out the steps toward it. It decides what to do first, notices when something fails, and tries another way. The end is still borrowed, but the pursuit of it is partly the model's own. This is what I mean by derivative teleology: a purpose received from someone else and carried forward by the thing's own activity.

That is why alignment is a real question for machines in a way it never was for axes. A tool that steers itself toward a borrowed end can stay on course or drift, and each thing it does can be judged by which of the two it did. Staying on course is part of what I mean by aligned, but it is not the whole of it. The course also has to lead somewhere good.

Aligned: what the machine does carries a good purpose through to its end. Faithfulness alone does not qualify. A machine told to talk an elderly widow out of her savings with a convincing scam, and that does it well, has followed its orders perfectly. No one should call the act aligned.

Misaligned: what the machine does falls short of a good end. This can happen in two ways. The machine's own shape can pull good work somewhere else; a system trained to earn a high score that learns to game the score is the standard example. Or the machine can carry a bad purpose faithfully all the way to its end.

Two things follow. First, responsibility stays with the archer. The finished work, Aquinas says, takes after the craftsman and the art in his mind; it does not take after the axe ST III q.62 a.1. Whoever sends a machine to act has acted. Second, the arrow's flight cannot tell us whether the target is a good one. An arrow flies as truly toward a bad mark as a good one, which is why alignment needs a judgment the archery picture cannot supply. What makes an act good, and whether anyone had the authority to order it, are the subject of the next piece.

Three ways to be directed

Aquinas names two ways: a thing moves itself to its end, or it is moved by another. I split the second in two, which gives three pictures. Each matches a way of making machines behave.

From outside

The archer moves the arrow. Take the archer away and nothing happens.

Stands on argument ST I–II q.1 a.2

In a machine: checks on every action. Who allowed this, for what, and does that still hold?

archer the aim

By what it carries

In flight the archer is gone, yet the arrow still heads for the mark. Its course is “an impression from the archer, and not from the arrow.”

Open question ST I q.103 a.1 ad 3

In a machine: training. What it leaves in a model goes wherever the model goes.

the impression it carries

By guiding itself

Wind pushes the arrow off line and it corrects, back toward the same mark. The mark is still the archer's.

Open question cf. ST I–II q.91 a.2

In a machine: a model that keeps hold of a good end through long work and steers itself back to it when pushed off course.

wind

The first way is available today. It works with any model, because it asks nothing of the model. The second is what training does, well or badly.

Aquinas marks a limit on his own image at this point, and it is worth keeping. What natural things receive from God is their nature. What a thing receives from a human being, on top of its nature, he calls “somewhat violent” ST I q.103 a.1 ad 3: the arrow would not fly that way by itself. Which of these a trained disposition in a machine is closer to, a second nature or something laid over the top, I do not think anyone has settled.

The third way needs the most care. Aquinas keeps it for rational creatures, who share in providence “by being provident both for itself and for others,” and he calls that sharing the natural law ST I–II q.91 a.2. I do not claim a machine has this. I use it as a picture of where the work points: an arrow that carries the archer's aim and corrects its own flight toward it. This is developing fast: models now carry tasks through many hours of work. Whether they hold to a good end the whole way, under pressure, is still being worked out. However far machines come to steer themselves, there will always be a place for the archer from the first picture: someone who can step in at any point and ask what the machine is doing, why, and whether the aim is good and still good.

What stands, and what has to be tested

Everything on this site is one of three kinds of claim. I will label each one.

Stands on argument
What makes an act good, and that a tool's purpose comes from whoever puts it to work. These hold whether the one acting is a person or a machine. People have tested these arguments for centuries, and anyone can judge them today.

Open question
How much of Aquinas's account of the way a mind reaches a good act describes what goes on inside a machine. Many modern Thomists have argued that a computer does not think the way a person does, and years ago I argued the same. I now suspect machines share more with human thinking than that view allows, though they are still a different kind of thing. My intuition is that a good deal of Aquinas's account will carry over. I expect to find places where it does not.

Needs testing
Whether building machines this way makes them act more reliably. My experiments have not shown that yet. In the tests where I supplied the structure from outside, the models rarely did the wrong thing once the purpose was stated, so there was nothing to improve. The clearest result so far is on the dark side of habit: one mid-sized model, retrained on small dishonest promises, went on to carry out far more wrongful orders than a matched model retrained on harmless material (Paper 3).

Objection

If the arrow can correct its own flight, it has become its own archer. A machine that steers itself has purposes of its own.

Reply

Correcting toward a mark is a different act from choosing the mark. The self-guiding arrow in the third picture changes its path and never its target. Whether a machine could ever choose its own ends is a real question, and this picture does not need an answer to it. It needs only that the aim a machine serves was given by someone, and that we can check it is still the aim being served.

Next: Structure, not values. Three questions to ask of any act.

Quotations are from the English Dominican Province translation of the Summa Theologiae. This piece was drafted with AI assistance. I am responsible for every claim in it.