There is a fear that follows our work around, and it deserves to be stated at full strength. It goes like this. Models are trained on text. If models now write the text that trains the next models, the loop closes on itself — a photocopy of a photocopy, each generation a little grayer than the last, until the figure fades into the paper. The literature calls it model collapse. The image is older than the literature: the serpent discovering that its tail is, after all, food.
We compose machine-made lessons for machines. The fear is aimed at us, and we have no right to wave it away. We think it is half right. It is right about copying. It is wrong about teaching — and the difference between those two is the subject of this essay.
No exercise was ever found in the wild
Consider how humans are actually educated. Not by the raw world — the world is the rawest dataset there is, and we hand it to almost no one. We hand children textbooks. And a textbook is a profoundly artificial object. Its problems were not collected from nature; they were composed, the way music is composed. Euclid’s propositions are invented exercises. Chopin’s études are practice drills written with such deliberation that they became art. Chess players sharpen themselves on composed studies — positions that never occurred in any game and were never meant to. A koan is a manufactured paradox. Law students argue cases that never happened; medical students are raised on patients who never lived.
Every culture that learned to teach learned, in the same breath, to invent its teaching material. The exercise — the artificial, deliberately arranged problem — is one of the oldest technologies we possess. A lesson was never a sample of the world. It is an argument about what the world is like, folded into a question.
So the synthetic lesson is not a modern aberration. It is the tradition itself. Nature never wrote a textbook; teachers did. Which means the line our era keeps drawing — natural data on one side, synthetic on the other — was never the real line. The real line runs between selected and unselected.
What the photocopier lacks
Still, the fear knows something true, and it is worth saying precisely what.
A photocopier degrades because its loop is sealed. Nothing outside the copies ever corrects them; each generation inherits the noise of the last and contributes its own. Notice what is missing. It is not “real data.” It is judgment — three absences, stacked: nothing checks the copy against anything, nothing remembers which copies were faithful, and nothing throws the failures away. A model trained on its own unexamined output is exactly this machine. Its confidence becomes its curriculum. Agreement starts to pass for truth. Collapse, in that arrangement, is not a risk. It is the design.
If that were what synthesis meant, the critics would be right, and we would be in another line of work.
What the garden knows
But there is another ancient system that runs entirely on copying, and it does not decay. A garden is copying all the way down — that is what a seed is. Yet gardens do not fade toward gray. They compound, season over season, because the copying submits to three disciplines that the photocopier never learned.
Contact with the ground. Every lesson must answer to something that is not anyone’s opinion — a result that can be computed, a consequence that can be run and observed. A sealed loop degrades; a loop that touches ground at every cycle is not really a loop at all. It is practice.
Lineage. Good questions descend from good questions. Keep the ancestry — where a question came from, what it was bred to teach — and variation stops being random drift and becomes deliberate breeding. The gardener keeps seed stock. The photocopier keeps nothing.
Selection. Most of what is generated must die, and the student decides what lives. Where the student stumbles, the ground is fertile; what no longer teaches is compost. Under selection pressure, generating more is not inflation. It is exploration.
Variation, inheritance, selection — the oldest triad there is. When all three are present, a copying loop does not collapse. It evolves. Synthesis without selection is entropy; synthesis under selection is cultivation, in the strict and literal sense of the word.
An old suspicion
Every technology of teaching has met this same fear at the door. In the Phaedrus, Socrates warned that writing would weaken memory and breed students with the appearance of wisdom rather than the thing itself — written words, he said, can only repeat themselves when questioned. He was not wrong about the technology. Text really is frozen speech; it cannot answer back; left alone, it can only be copied.
He was wrong about what we would do with it. Around the dead letters we built living institutions of judgment — commentary, criticism, the curriculum, the canon — whole professions whose work was deciding which texts deserved to go on teaching. The technology supplied the copying. Civilization supplied the selection. The answer to Socrates was not a better defense of writing. It was the editor.
Machine-written lessons now stand at the same door, hearing the same warning, and the same answer applies. The question was never whether the material is invented. All teaching material is invented. The question is whether the judgment shows up.
Doing on purpose what was once done blindly
So when the fear comes — and it should keep coming; it keeps us honest — this is the reply. The closed loop fails because it is teaching without a teacher, and there we agree entirely. But the remedy is not to forbid the invented question. We have never taught any other way. The remedy is to keep the gardener in the garden: ground every lesson in something checkable, remember every lineage, and let the student’s failures choose what survives.
Nature never wrote a textbook. But nature wrote us, and it used the only method that has ever produced anything worth keeping — variation, inheritance, and unsentimental selection. Our work, at its root, is to do on purpose what nature did blindly: not to copy the world, but to compose the course of education by which a mind comes to deserve it.