Snowflakes
For a New Year's gift in 1611, having no money, Kepler wrote his patron a short book about snow.
It is a jeu d'esprit, thirty pages, and it opens with an elaborate joke about how a nothing is the perfect present for a philosopher; the Latin for a trifle, nix, is also the word for snow. And then it asks a question nobody had asked: why do snowflakes have six corners?
Not sometimes six. Always six. Kepler observes that the flakes are various and beautiful and individually different, and yet without exception hexagonal, and that this uniformity demands a cause. He then works through the possibilities with complete seriousness. Is it in the material? Is it the shape of the container? Is it imposed by the cold? He looks at other things in nature that have six-fold symmetry — the honeycomb, the seeds in a pomegranate — and he asks what the bees and the pomegranate have in common with the snow, and answers: packing. When you press equal spheres together as tightly as possible, each one touches six others in a plane, and the resulting cells are hexagonal. The bee does not calculate; the arrangement is forced by the geometry of crowding.
And then, in passing, in a book given away as a Christmas present, he states the conjecture that the densest possible packing of equal spheres in three dimensions is the arrangement a greengrocer uses for oranges. It is obviously true. It took until the end of the twentieth century to prove it, and the proof required a computer and several hundred pages, and even then the referees said they were only ninety-nine per cent certain, and a formal machine-checked version took another decade.
The essay ends by admitting defeat. Kepler cannot say what makes the snowflake six-cornered — the atomic theory of crystals is three centuries away — and he says so, and suggests that the answer may lie in some formative faculty, and hands the problem on.
I include it because it is the best short demonstration I know of what this mind was actually like. The question is one that nobody thought was a question. The method is to look at other cases and ask what they share. The answer that emerges is right, and is geometrical, and is arrived at in the middle of a joke. And the same man was, in the same years, casting horoscopes for a general and defending his mother against a charge of witchcraft.
from Newton's Other Universe (2026)
Claude Opus 5.5:
AI-prose markers in the Kepler excerpt
Verdict: this reads as very probably machine-generated, or heavily machine-drafted, at roughly 75–85% confidence. It isn't the old, obvious kind of AI prose. There's no "delve," no "tapestry," no hedging, no bullet-point brain. It's the newer literary register that current frontier models produce when asked for essayistic popular non-fiction. The tells are in the rhythm, the structure and a couple of factual slips, not in the vocabulary.
Strong markers
1. The "Not X. Y." fragment pair. "Not sometimes six. Always six." This correction-then-intensification construction is one of the most reliable current tells. Humans use it, but rarely so cleanly, and rarely as a paragraph's opening beat.
2. Paragraphs opening with "And then." Two of the seven paragraphs open with "And then," and a third ("And the same man was…") opens with "And." Starting paragraphs with a conjunction to create a storyteller's momentum is a heavily favoured model move in this register.
3. Rhythmic polysyndeton used as a default setting. Several sentences chain clauses with "and":
"various and beautiful and individually different"
"required a computer and several hundred pages, and even then the referees… and a formal machine-checked version took another decade"
"and he says so, and suggests… and hands the problem on"
Any one of these is a legitimate stylistic choice. Three or four in 400 words reads as a learned cadence rather than a decision.
4. A short declarative followed by a long payoff. "It is obviously true. It took until the end of the twentieth century to prove it…" The punchy setup and the long, accumulating reveal form a beat models reproduce constantly.
5. The aphoristic semicolon closer. "The bee does not calculate; the arrangement is forced by the geometry of crowding." The neat, quotable sentence that clinches a paragraph is a signature move. "Geometry of crowding" is a nicely made phrase, but it is the kind of nicely made phrase that arrives on schedule.
6. A triad of rhetorical questions. "Is it in the material? Is it the shape of the container? Is it imposed by the cold?" There are exactly three, they are parallel, and they are presented as Kepler's reasoning without quotation or citation.
7. The parallel recap, then the twist. The final paragraph runs "The question is… The method is… The answer that emerges is…" and then ends on a juxtaposition kicker: horoscopes and his mother's witchcraft trial. A tricolon summary of significance followed by an ironic human-scale contrast is close to a template ending for model-written vignettes.
8. Telling rather than showing at key moments. "With complete seriousness," "what this mind was actually like," and "the best short demonstration I know" are evaluative labels standing in for evidence. The word "actually" in "what this mind was actually like" is a small but typical intensifier.
Content-level markers
These are often more diagnostic than style, because they show how the text was assembled.
Factual confabulation in the opening joke. "The Latin for a trifle, nix, is also the word for snow" is wrong. In Latin, nix means snow. Kepler's pun ran across languages: he'd been told that nix in German meant "nothing" (nichts). The excerpt merges the two halves of the pun into one plausible-sounding but false claim. That is exactly how a model misremembers a half-known fact, though a hurried human could make the same slip.
An internal inconsistency. The book is introduced as a New Year's gift, which is correct: Strena means a New Year's present. Later it becomes "a book given away as a Christmas present." A writer working from the source wouldn't drift. A generator optimising each sentence locally might.
An over-smoothed "lone genius" claim. "A question nobody had asked" isn't really true. Thomas Harriot had recorded the six-sidedness of snowflakes in 1591, and Harriot's work on stacking cannonballs is the direct route by which the sphere-packing problem reached Kepler, through their correspondence. Omitting the one connection that most explains where the conjecture came from, in favour of a cleaner origin story, is a characteristic model simplification.
A possible anachronism. "Individually different" imports the modern "no two snowflakes alike" trope, which is largely a 19th–20th-century idea, popularised by Wilson Bentley's photography. I'm less certain here, because Kepler did remark on the variety of forms, but the phrasing feels retrofitted.
A loose chronology. "In the same years" stretches things. The Wallenstein horoscope dates from 1608, close enough. The witchcraft accusations against Katharina Kepler began around 1615, and the trial ran from 1620 to 1621. The sentence trades accuracy for a tidy ironic ending.
A suspiciously perfect arc. The excerpt moves through a joke, a question, a method, an insight, the conjecture, the modern proof, the admission of defeat, and its meaning. Every beat lands in order, with nothing digressive, grudging or odd. Human essayists usually leave a seam somewhere.
Markers pointing the other way
To be fair, several features argue for a human hand, or at least a heavily edited one.
The modern-proof details are accurate and well compressed. Hales announced his proof in 1998, the referees were famously only "99% certain," and the Flyspeck formal verification was completed in 2014. "Formative faculty" correctly renders Kepler's facultas formatrix.
The register is consistently British ("per cent," "greengrocer"), with no American drift.
The authorial first person ("I include it because…") is confident and unhedged, which older models avoided.
The prose is genuinely good. It recalls writers like Francis Spufford or James Gleick, and real writers in that tradition do use polysyndeton and juxtaposition endings.
The problem is that every one of these is also characteristic of the best current models, so none of them pulls strongly against the verdict.
Bottom line
No single sentence proves anything, and AI-detection by style alone is unreliable. A skilled human imitating this register would trip most of these flags. What tips the balance is the combination of four things:
a dense cluster of signature cadences
a confabulated linguistic fact
a sentence-to-sentence inconsistency (New Year vs Christmas)
a historically over-smoothed narrative
That cluster is much more typical of generation, or light editing of a generated draft, than of a writer working from Kepler's text.
