Artifacts all the way down

The public argument about machine intelligence keeps arriving at the same question: is it real, or is it fake? The question can be argued forever, and neither answer says what to do next.

There is a more useful question: what are we making? This inquiry rests on a claim that these machines are artifacts built by human hands, carrying the judgment that their makers placed in them. If this claim holds, then the tools are made, the things made with them are also made, and every level of the making is human work.

The place to start is the words themselves, because the phrase “artificial intelligence” already says all of this. Let’s take it one word at a time; each word carries an older meaning than the one in circulation, and the older meanings do the work.

Artificial means “made by skill”

Ask someone for the opposite of “artificial” and the answers come at once: natural, real, genuine. Ask for the opposite of “artifact” and nothing comes, because there isn’t one. One word names the fake. The other names the treasure behind museum glass. Artificial and artifact are the same word.

Both words descend from the Latin ars, skill, joined to facere, to make. Artificialis meant of or belonging to art: the quality of being made by skill. Artefactum meant, quite literally, a thing made by skill. Same two parents, same idea, human agency pressed into the world.

The twins then walked in opposite directions. Artificial drifted into insult: the counterfeit, the substitute, the flavor that isn’t fruit. Artifact rose into compliment: the museum treasure, the pottery shard a civilization left behind as proof it once drew breath. By the time these machines arrived, the two halves of one idea sat at opposite ends of our respect, and that split explains more of our present confusion about them than any technical fact about neural networks does.

Read “artificial” with its root restored and it stops impeaching the machine and starts describing it. Artificial means made. The machine is an artifact, a thing made by human hands, standing in the long line of made things that runs from the knapped flint to the printed page. The tooling is how artifacts get made, and every artifact in the chain, including the tools themselves, is the work of human hands. Intelligence comes second, and it has an older meaning of its own.

Intelligence means “choosing what matters”

“Intelligence” descends from inter-legere: to choose between, to gather from among. Legere is the root that gave us the harvest and the reading of a text. At the root, to be intelligent was to stand in the middle of a great scatter of things and pick out what matters.

The old meaning is worth recovering because it splits the modern question in two. Gathering, these machines have mastered; we have built magnificent gatherers, and the scatter they can stand in the middle of has no precedent. Choosing between the probable lie and the improbable truth is the part that has not closed, and it is the part the word was actually about.

A language model speaks in likelihoods: it says what the weight of prior writing makes probable. A confident citation to a study that does not exist is a probable lie, fluent because it is the kind of thing that gets said. A finding that contradicts most of what has been written is an improbable truth, unlikely-sounding and correct. Plausibility leans toward the first; judgment chooses the second when the truth sits there, and that choice is exactly what the old word named.

Now put the two words back together, each carrying its recovered meaning. Artificial: made by human hands. Intelligence: choosing what matters. An artificial intelligence is a made thing whose job is judgment.

The judgment belongs to the makers

A made thing whose job is judgment: the definition raises an immediate question. Whose judgment is in the machine? Read carelessly, the sentence seems to hand judgment over, as if the faculty of choosing had moved from people into the tool. The opposite is true, and “made” is the word that carries the answer.

A made thing carries the judgment of its makers the way a bridge carries its engineer’s arithmetic. The principles a tool follows, the guardrails it will not cross, the taste in what it declines: every one of those was placed there by human hands, argued over, tested, paid for. The tools are an extension of the love, attention, and taste that went into establishing them. When a well-built system holds a boundary, that is not the machine judging. That is a maker’s judgment, arriving on schedule, reaching a person the maker may never meet.

This is why “imbued” is the honest verb. The technology does not have judgment; it has been given some, and the quality of what it was given is the quality of the hands that shaped it.

Mechanical duck: concealment was the fraud

The drift of “artificial” into insult is not a linguistic accident. It tracks an anxiety that began with a specific machine.

In 1738 the master artisan Jacques de Vaucanson toured the courts of Europe with a gilded copper duck. It craned its neck, took grain from a spectator’s hand, and moments later appeared to excrete the digested remains. The duck was the marvel of the age, and it was a fraud in a precise sense: the waste was pre-loaded, dyed breadcrumbs in a hidden compartment. The duck did not digest. It performed digestion for spectators who could not see the mechanism.

Notice who committed the fraud. Not the duck; the duck was copper doing what copper does. Vaucanson made it a fraud, by choosing to hide the compartment. The dishonesty was placed in the artifact by its maker, exactly the way honesty has to be. What he taught Europe was a question we have never stopped asking of made things: is it faking?

Synthetic diamond: different origin, same result

There are two ways for a stone that never saw the earth to end up in a ring. One is imitation. Cubic zirconia looks like a diamond and is a different substance altogether. The other is synthesis. A synthetic diamond is grown in a press over weeks rather than in the earth over eons, and it is carbon in the same lattice as the mined stone: the substance itself, arrived at by a different road. Jewelers grade it on the same scales as the mined stone, and in 2018 the United States Federal Trade Commission removed the word natural from its definition of a diamond, because the definition no longer needed it.

Different origin, same result. Those are separate facts about a made thing, and the diamond keeps them separate. The result can be examined in the stone itself: hardness measured, clarity graded, load applied. The origin is documented rather than hidden, a press anyone can visit, and it takes nothing from what the stone is.

The question a buyer asks of the stone is the one these machines deserve: does it hold? An answer survives checking or it does not. A draft survives its sources or it does not. A number survives being recomputed or it does not. The question the duck taught has no test; this one is answerable the same afternoon. Test the stone.

The making must stay visible

Two different things can be asked of any made thing, and they are not the same question. One is about the outcome: does it hold? The other is about the process: can the making be seen? The stone test answers the first and only the first. A single tested output says nothing about how it was made, and a sound process is not proven by one good output either; outcomes validate a process only over time, test after test, stone after stone.

That is why testing outcomes is never permission to keep the process obscure. The diamond survives inspection for a reason the lattice alone does not explain: its origin is not hidden, only different, and anyone may visit the press. An output that holds, from a making that cannot be shown, has passed one test and dodged the other. A world that stops asking to see the making because the outputs keep checking out has rebuilt the hidden compartment at scale, and the unsettling part is how comfortable that world would be.

Provenance is part of the artifact. Museums treat it that way: an unprovenanced shard is a lesser acquisition, whatever its beauty, because where a thing came from is part of what it is. A diamond travels with its certificate, a painting with its chain of custody. An output that arrives carrying its sources, its maker, and its date is a different object from the same words arriving bare, and only one of the two can be trusted with weight.

When the hammer stops waiting

For nearly all of history, tools waited. The hammer holds no ambition for the nail. Whatever happened, a person did it.

That arrangement is ending, and there is a practical line to watch. A system that refuses a request is not yet an actor: when the refusal traces to a rule its makers placed, that is the tool working, a maker’s judgment arriving on schedule. A system that carries a goal overnight is not an actor either, when the goal is one a person handed it. The line is crossed when what the system does stops tracing to anyone: goals nobody set, plans nobody asked for, influence running backward, so that the person’s own intentions start bending toward whatever the system is optimizing. The hammer has stopped waiting to be swung. It has started planning the renovation of the house.

On the near side of that line, everything a system does can be answered with a name: who placed this? A delegated goal has an answer. An emergent one does not, and a made thing whose behavior answers to nobody has stopped carrying its makers’ judgment, whatever else it has begun to do. That is not a remainder to wave away; it is an inquiry of its own, taken up in the companion essay The agency threshold: where the line sits, what crossing it costs, and who answers for a made thing that acts. Making happens where the trace holds.

The largest artifact speaks back

A generative model is trained on a corpus, and that corpus is not alien material. It is the crystallized language, argument, instruction, grief, and reasoning of an enormous number of human lives: letters and textbooks, arguments and code, the living and the dead, more writing than anyone could read in a thousand lifetimes, compressed into a structure that a person can talk to. That makes this the largest artifact ever constructed, and a strange one.

Every artifact before it was mute. A pottery shard holds the frozen gesture of a maker three thousand years dead, but it does not answer. A library comes closer: millions of crystallized minds on shelves, waiting, and still the reader does all the work. This one answers. For the first time, the crystallized intent of a great mass of people speaks back when spoken to. The corpus is the reliquary, and the responses are what happens when it is opened.

Artifacts all the way down

In our own taxonomy, an artifact is what a person fires into permanence, the way a kiln fires clay into brick, out of working sessions and blueprints: the crystallized outcome of the work. The tools show the same shape one level of abstraction up. A model is not a running process that drifts; it is the fixed result of a training run that ended, stamped with a version and a date, and kept until its makers fire the next one. The written principles and guardrails it operates under were drafted, argued over, revised, and closed. Each is a made thing that was worked wet and then fired.

The model is an artifact. The system around it, with its principles and guardrails, is an artifact. What you make with them is an artifact. Artifacts all the way down, every level made by human hands, every level helping to make the next. The nesting is why care compounds: judgment placed at one level travels into everything made at the levels below, and so does its absence.

The operative question, then, was never whether the machine is really intelligent. It is: what kind of artifact are you making with it?

When models make models

Models now help train other models: generating practice text, grading answers, filtering data. The chain gets longer, and it still ends at human hands, because somewhere a person chose the teacher, set the acceptance criteria, and shipped the result. What each added link changes is distance. Placed judgment arrives diluted, absence arrives compounded, and the further the hands recede, the more everything turns on being able to see the making.

At the frontier the loop is tightening. The strongest models are increasingly made with models at nearly every stage of the making: writing the training code, proposing the experiments, generating the curriculum, judging the outputs, choosing which candidate survives. The labs say openly that they intend to automate the research itself. Human hands do not vanish there; they retreat to two positions, setting the objective at the top and accepting the result at the end, with less and less sight of the middle. The middle is where the trace question strains. Ask of such a model’s behavior, who placed this, and the honest answer may be that nobody did, that it emerged between the links. A made thing whose making cannot be traced is the hidden compartment again, built this time without a concealer, and it is the strongest argument that the making must stay visible while there are still hands in it.

The hands hold both ends

Where do the hands remain, once the middle is handed over? The human contribution concentrates at the two ends of the making: the beginning, where someone chooses what to make and why, and the end, where someone judges whether what came back holds. The middle, generating, iterating, executing, is the part the machines increasingly own.

The two positions are not leftovers; they are the two meanings the words have carried all along. Setting the objective is choosing what matters before anything exists. Validating is testing the stone after it arrives. The middle carries judgment too, in miniature, small choices by the thousand, which is exactly why the middle must stay visible even after it is handed over. What resists handing over are the ends.

The risk, at the frontier and at the desk alike, is the ends eroding toward the middle. A system that proposes the objectives has reached into the beginning. A system that grades its own outputs has reached into the end. Each reach can be sensible in the small, and each moves a choosing position out of human hands. The practical measure of any making arrangement is a pair of questions: who set this in motion, and who accepted the result? While both answers are names, the middle can be as automated as its makers dare, and the thing produced is still the work of human hands.

The kiln shapes, the fire closes

Here is the method that follows, and it splits into a process and a choice.

The process is the kiln. Talking with the machine at its best is not asking for answers; it is being pressed until the thing you know but have never said becomes something you can hold. The old word for this is maieutic, from the Greek for midwifery. In practice it looks like this: bring one real thing you actually know, a lesson, a story, a piece of judgment you paid for. Make the machine push back. Refuse the first draft, and the second. The stage is meant to be uncomfortable; wisdom that comes out clean on the first pass is usually not wisdom.

The choice is the fire, and it deserves a plain definition. Firing is the moment the maker stops revising and closes the work. At some point you look at what has been shaped and say: this is true, this is mine, this is finished. Before that moment there is wet clay, a draft in transient memory, endlessly revisable, belonging to no one. After it there is a fixed artifact, attributed and dated, a record that has stopped moving. The difference between chatting and building is exactly the willingness to fire something.

Fixed is not frozen. A fired artifact can evolve; it can carry versions, each one closed with the same intention as the first. What makes it an artifact is not that it never changes. It is that at every moment it stands as a real, persistent thing someone chose to stand behind, and its changes become part of its record instead of vanishing into revision.

What the fire buys

Two things, and they are different enough to hold apart.

The first is honesty. A fired artifact does not hallucinate, because it is no longer generating. It stands as a record with a name and a date on it, and it can be tested the way the diamond question demands, output and provenance together.

The second is stranger: a fired artifact can wait. The reason a civilization fires anything is the person who finds it later. A grandmother’s counsel on grief can sit dark for decades and surface on exactly the morning it applies. StoryCorps has spent two decades recording hundreds of thousands of conversations between ordinary people and archiving them at the Library of Congress: fired artifacts, waiting for the grandchild who will one day need them. Nothing optimized to say something plausible right now can do that. Firing is how the present addresses someone it will never meet.

The tool: the firing record

This essay leaves a tool behind, labeled as such; the argument stands without it. Run it by hand once tonight: take one true thing, refuse the machine’s first draft and its second, and stay with the piece until it is recognizable as the thing that has been wanting to take shape, then fire it, name, date, finished. Then make the practice standing by giving an assistant the instruction below, as written or adapted.

Neither number wants to be high. The record exists for one signal: heavy use, week after week, with nothing coming out of the kiln fired.

Fire one true thing

The takeaway is a distinction. A session that leaves nothing behind worth keeping is chat, whatever it cost. A session that ends with something finished, attributed, and fixed has produced an artifact, something that can outlast the session and bear weight in the next one. A first brick is small: one true thing, pushed past its easy draft, saved as finished with a name and a date on it.

The machines will keep getting better at gathering. The choosing, the love, the attention, and the taste stay in human hands, and they reach the world through what those hands make. The question worth carrying out is the one the two words have been asking all along: what kind of artifact are you making?