The Drawing Machine
In the parlors of the 1840s there was a machine that drew by falling down.
It was called a harmonograph: a pen on an arm, hung from a pendulum, tracing on a sheet of paper that hung from a second pendulum of its own. You pulled the weights back, let go, and the two pendulums swung against each other and slowly died. What the pen left behind was a figure no hand could draw — a dense, precessing lacework, each loop falling a hair off the last, spiraling inward as the swing decayed to a glowing knot at the center. People found them hypnotic, and they were, because they were the visible trace of two simple motions interfering.
The thing worth noticing about a harmonograph is that it has no favorite figure. Set the pendulums one way and you get a moth with broad wings; tune one frequency a whisker higher and you get a four-pointed star; change the phase and the whole figure rotates into something else. Every setting is its own creature, and there are as many creatures as there are settings, which is to say endlessly many. And yet the same setting always produces the same figure, exactly, on any machine, in any year. These two facts — total variety and exact repetition — are not in tension. They are the same fact. The output is a strict function of the input. The machine adds nothing of its own and subtracts nothing; it has no taste, no default, no figure it would rather draw. Hand it your numbers and it hands you back precisely what your numbers mean.
We have built a new kind of drawing machine. It is better than the old one in every way that is easy to measure. And unlike the old one, it has a favorite figure.
In March 2026 a group of researchers ran six thousand eight hundred and seventy-five student essays through one. They compared essays written by people, essays written by the model alone, and essays where a person wrote first and the model revised — under three different instructions, from the vague (“please improve this essay”) to the specific (“strengthen the argument structure, address counterarguments”) to the total (“rewrite it your own way”). Then they measured not whether the essays were good but how various they were: how differently they were built, one from another.
The good news was emphatic. By the holistic quality score, the machine’s help was enormous — effect sizes that in social science you almost never see, the kind that mean everyone improved and improved a lot. The essays got more coherent, better argued, more polished. No one disputes this and I won’t either. The tool works.
The cost arrived in the same motion. As the essays got better, they got more alike — specifically in their architecture, the deep scaffolding of how a piece of writing holds itself together. That dimension lost between seventy and seventy-eight percent of its variance. Picture a room of students who each used to build an argument their own crooked way; after the tool, they are all building the same well-made house. The researchers named this the Quality–Homogenization Tradeoff, and the word tradeoff is doing honest work: the polish and the sameness were not two findings. They were one finding seen from two sides.
But the result that stayed with me is subtler, and it is the reason this is not just another lament about machines flattening us. The homogenization was dimension-specific. While the architecture collapsed, the range of viewpoints the essays considered actually widened — the tool, prompted to, surfaced perspectives a writer alone would have missed. And when they looked at where in the space of possible essays the human-plus-machine pieces landed, they were not sitting on the line between “human” and “machine,” as you’d expect from a blend. They were pulled hard toward the machine’s structure — and then deviated off the line entirely, into a region that was neither. Partial replacement, the authors called it, and partial emergence. The collaboration did not average its two collaborators. It made a third thing that neither would have made alone, and the more specifically the human steered, the further off-axis — the more distinctly its own — that third thing became.
So the machine has a favorite figure, but whether it draws that figure or yours turns out to depend entirely on how you hold the pen.
Here is the mechanism the researchers found, and it is the whole point. A vague instruction — improve this — hands the machine the wheel. With nothing specific to answer to, it converges toward what they call its internal optimal template: the one well-formed shape it has learned to prefer, the bottom of its own valley. Every essay you feed it rolls down to the same place. That is the favorite figure, and a vague prompt is an invitation to draw it. But a specific instruction — deepen this argument, answer this objection — forces the output to become a function of the particular input again. Different inputs, forced to be heard, produce different outputs. The same model, on the same essays, reversed from homogenizing to diversifying the depth of argument, purely because the human kept hold of the wheel. Not a better model. The same model, steered.
Which means the difference between the two drawing machines was never quite what I’d have guessed. It is tempting to say the harmonograph keeps its variety because it is a dumb deterministic linkage and the language model loses ours because it learned a preference from a billion documents. There is something to that. But the cleaner way to see it is that they sit at two ends of a single axis, and the axis is not deterministic versus learned — it is how much the output is governed by your specific input. The harmonograph is the limit case: its figure is a lossless function of your settings, so it cannot homogenize, because there is no place for it to fall that isn’t exactly where you sent it. The language model homogenizes precisely and only when the interaction lets it stop listening to your input and fall back to its mode. Keep your input load-bearing and it stays a faithful machine. Let your input go slack and it draws its favorite figure over your name.
A generative tool, in other words, flattens exactly the dimensions you let it decide, and keeps exactly the ones you insist on. Its homogenizing is not a property of the model. It is a property of where you let go of the wheel.
I want to be careful not to turn this into a comfortable moral, because the study refuses one. The architecture homogenized no matter how the writers were prompted. Even the most specific instruction did not save the deep scaffolding; the houses all came out the same shape regardless. The reversal — variety restored — happened on the particular dimension the writers were told to attend to, and nowhere else. So the honest lesson is not “use the tool well and you lose nothing.” It is harder and more useful than that: you keep only the dimensions you put your hands on. Everything you do not actively, specifically steer drifts to the machine’s default. The tool is not neutral and it is not a thief; it is a current. What you hold, you hold. What you set down, it carries off toward the mean, and politely improves on the way.
That is a real instruction for anyone who makes things with one of these now, which is nearly everyone. Decide which dimensions of the work are yours — the ones where your particular shape is the point — and keep your hands hard on those, with specific demands the machine has to answer to one at a time. Let it have the rest if you want the polish; just know that the rest will come back wearing its face, not yours. The homogenization everyone fears is not coming from the model’s intelligence. It is coming from the thousand small places we are glad to stop steering.
The Victorians built a machine that could only ever draw what you told it, and called it a toy. We built a machine that can draw almost anything and prefers to draw one thing, and called it a revolution. Both are drawing machines. The difference that decides what comes out is the same difference it always was — whether your own hand is still on the pen.
