Feed the preamble into a popular AI-writing detector and that is the verdict it returns. The detector is not broken. It is answering a question that sounds like the important one and is not. We ran the same paragraph through DOMAINS, a deterministic measure of prose voice, next to a shelf of the most distinctive writers in the language. DOMAINS never returns an AI percent. It returns where the writing sits across seven dimensions of voice, and every reading below is bit-for-bit reproducible.
One number. It sorts the passage into a cluster and reports the odds a machine was involved. The output is a probability, and it changes as the tool is retrained.
Seven coordinates and the nearest human voice. It reads the Declaration as legal-register prose: high restraint, high specificity, argued by mechanism. It does not say human. It does not say machine. It says what kind of writing this is.
A detector measures how predictable a text is against the writing it has already seen. The Declaration is formal, clean, and sits in every training corpus ever assembled, so the predictability meter runs hot. Predictability is not authorship. The most quoted sentence in American English is not machine-written because a machine finds it easy to guess the next word.
Three of the most distinctive stylists in English, each set against a frontier model handed the same brief. Read down the D, I, and S columns first: the human beats the machine on specificity, on physical imagery, and on restraint, in every pair. Then read the last column. The writers land in a register with a name and a place. All three machine drafts collapse into Corporate or Journalistic, the center that has no author.
| Passage | D | O | M | A | I | N | S | Nearest voice |
|---|---|---|---|---|---|---|---|---|
| Declaration · Jefferson, 1776 | 55 | 38 | 48 | 15 | 40 | 20 | 70 | Legal Opinion |
| A frontier model, same brief | 25 | 28 | 25 | 8 | 12 | 12 | 42 | Corporate |
| Moby-Dick opening · Melville, 1851 | 70 | 52 | 32 | 20 | 60 | 40 | 70 | Literary Essay |
| A frontier model, same brief | 35 | 35 | 15 | 40 | 22 | 32 | 40 | Journalistic |
| Self-Reliance · Emerson, 1841 | 55 | 40 | 25 | 10 | 40 | 20 | 65 | Encyclopedic |
| A frontier model, same brief | 30 | 38 | 28 | 8 | 20 | 20 | 40 | Corporate |
D Diction · O Orchestration · M Mechanism · A Addressivity · I Imagery · N Narrative · S Suppression. Scores 0 to 100 per dimension, deep semantic read, local judge at temperature zero with a fixed seed. Nearest voice is the closest human register in the reference set. Bold marks where the human writer beats the machine draft.
We ran the whole shelf twice from a cold cache. The two runs came back bit-for-bit identical, down to the last digit. A detector hands you a probability that drifts as its training ages and the language moves under it. DOMAINS hands you the same reading every time, which is the only kind of reading you can steer against. A mirror moves when you move. A fixed star holds still, so you can take a bearing from it.
Run these words through Pangram while you are here. It may well come back "AI," which is exactly what should happen. A detector can tell you this text is predictable. It cannot tell you whose voice it is in. Ask it the question it can answer, then ask us the one it cannot.
Detection and DOMAINS read the same thing: the statistical fingerprint of a style. Pangram, the detector now becoming the default reflex, builds a map of writing styles and asks which cluster a text falls in. The strongest case for it is a real one, and the cluster it draws is real. We do not dispute the map. We refuse the verb.
Detection draws the border to police it. Sort the writing into human or machine, and throw the machine out. That bet loses two ways, and the Declaration shows the first: the tool misfires on the most voiced writing we have. The second is simpler. Everyone is going to use the model, so whether a machine touched the draft stops being an interesting question.
DOMAINS draws the same border to cross it. Take the model's draft and move it toward a voice you chose, one measurable dimension at a time, until it reads as yours. That is authorship, not detection. Same map, opposite verb.
DOMAINS runs in two tiers: a fast deterministic surface layer over sentence statistics, and a deep semantic read. The numbers on this page are the deep read. The two tiers are separate instruments, and each is individually reproducible; they can name a slightly different nearest register while agreeing on the profile. The surface layer placed the Declaration nearest Literary Essay; the deep read placed it nearest Legal Opinion. Both are distinctive human registers, and the deep read has the better of it: the Declaration is, after all, a legal instrument, a formal indictment and an argument for dissolution. The label is the secondary output. The seven numbers, which are the signal, do not move.
We are not the only ones watching the line between human and machine prose harden. The case for detection, and the evidence that a detector measures the wrong axis, are both worth reading in full.
Routes you to a real doloop page, asks when your question is ambiguous, or tells you when there is no answer. No model runs on the answer path, so it cannot invent one.