Musings, Meanderings, & Loops · Essay
August 14, 2026
I went down a rabbit hole this week that started in one place and ended somewhere I did not expect, which is becoming a habit.
The starting question was mechanical: why can you tell? Machine-written prose has a texture. Most people can feel it before they can name it — a certain smoothness, an evenness of tone, a paragraph that arrives exactly where you assumed it would. I wanted to know what that texture actually is.
The answer turns out to be simple and slightly deflating. A language model does not write sentences. It writes one token at a time — a token being roughly three-quarters of a word — and for each one it calculates which token is most likely to come next given everything before it. That is the whole mechanism. Enormously sophisticated in execution, almost embarrassingly plain in principle.
And if you always choose the most likely next thing, you produce prose with a specific and recognizable defect: it is never surprised by itself.
There is now a small economy built on hiding this. Tools that take machine text and rough it up — vary the sentence lengths, swap the predictable adverbs, run the whole thing through a persona — so that it passes as human. The technical vocabulary is perplexity and burstiness: how surprising the word choices are, and how much the rhythm varies.
I understand the market and I am not interested in it, for a reason that has nothing to do with propriety. The detectors do not work very well. They produce false positives constantly, and they are especially prone to flagging formal writing and non-native English. Building an industry to evade an unreliable judge means you have accepted the judge's premise — that the goal is to pass.
But passing was never the interesting problem. The interesting problem is that the writing is dead. Fooling a detector does not fix that. It just makes the corpse harder to identify.
The obvious fix is to make the model choose less predictable words. There is a dial for this — temperature — and you can turn it up, and the text does become less predictable.
It also becomes worse.
Push it far enough and the model loses the thread, contradicts itself, invents things. You have traded smooth deadness for chaotic deadness. Nothing has been gained, because the randomness is not doing anything. It is noise added to an average.
Here is what I think the actual insight is, and it took me a while to get to. Human unpredictability is not random. It is intentional.
When a good writer drops a three-word sentence after a long, winding one, that is not statistical variance. It is a decision. Something in the argument needed a hard stop, and the writer knew it. The short sentence is low probability in the purely local sense — after eighty words of subordinate clauses, another long clause is far more likely — and the writer chose it anyway, because of something operating at a level the sentence itself cannot see.
That is the gap. A model optimizing the next token is reasoning locally with no theory of the whole. It cannot make the three-word decision because the reason for it does not exist inside the window it is looking at.
Which brings me back to Christopher Alexander, who opened the first essay I wrote this week and has been standing behind all of them since.
Alexander's argument in A Pattern Language was that good building does not come from applying templates. It comes from resolving forces — the genuine, opposing pressures of a situation. A room needs light and it needs privacy, and those two things fight; a pattern is a solution that resolves the fight in a way that stays alive. He catalogued 253 of them, at wildly different scales, from the layout of a region down to where you put a windowsill.
The thing to notice is that a pattern is a reason. It is an answer to a question about the whole. And it operates above the level of the individual brick.
So the fix for machine prose is not to shake the bricks. It is to give the system a layer that knows what the building is for.
That is not a metaphor I invented. Nikos Salingaros and Bruno Posel published a tool last year that does exactly this: teams assemble a context-specific subset of Alexander's patterns for a project, and those patterns go to the language model as structural constraints rather than as a prompt. The reported result is that the model produces designs nobody described to it — arrived at structurally, from the forces, rather than retrieved from the average of everything it has seen. Software people are building the same thing under a different name, treating agent capabilities as modular patterns with declared intents and declared forces.
Both are doing the same thing: putting something above the probability layer that can say this is what we are making, and here is the conflict it has to resolve.
I wrote yesterday about Huygens and his two pendulum clocks, and how they fell into rhythm not through the air but through the beam they were both mounted on. Alignment turned out to be a property of the shared substrate, not of the oscillators.
This is the same structure, and I did not see it until I had written both.
Token probabilities are the clocks. Turning up the temperature is shaking them harder and hoping coherence emerges. It does not, and it cannot, because the clocks were never going to synchronize through the air. The pattern layer is the beam — the thing underneath that lets local choices add up to something whole instead of something merely smooth.
And notice what this implies about the writing itself. The reason a human three-word sentence lands is not that it is improbable. It is that it is improbable and load-bearing. Surprise without structure is noise. Structure without surprise is the machine. What you want is the surprise that the structure required.
The complaint about AI writing is usually that it is too predictable, and the proposed cure is to make it less so. I think that diagnosis has the causation backwards.
Machine prose is not dead because it is predictable. It is predictable because it is dead — because there is no intention operating above the sentence, nothing that knows why this paragraph exists and what it is supposed to cost the reader. Predictability is the symptom. Absence of purpose is the disease.
Which is a hopeful conclusion, if you are someone who uses these tools to build things. It means the leverage is not in better sampling, or a cleverer humanizer, or a bigger model. It is in being able to say precisely what you are making and which forces it has to resolve — and that has always been the hard part, and it has always been ours.
The machine can find the next word. It cannot tell you why the paragraph is there. That question is still yours, and I suspect it is going to stay that way longer than most people currently expect.