doloop.io
DOMAINS · the shape AI writes toward

Ask seven models to write with a voice. They hand back the same one.

You spent years learning how you sound. A model imitates it in one prompt, and a dozen tools will sell you that imitation. None of them measure the thing that matters: whether your real writing still holds its own shape, or has started sliding toward the one shape AI writes toward. DOMAINS measures that distance. From outside the model, same reading every run.

The shape test

Every AI piece lands near the same shape, regardless of topic or model. Human writers each have their own.

Each shape is a writing fingerprint: the balance of diction, orchestration, mechanism, addressivity, imagery, narrative, and suppression in one piece. Four AI pieces from seven models including GPT-4o, Claude, and Gemini cluster tightly. Four human literary works look nothing alike. AI isn’t bad at voice. It’s optimized against having one. It is trained to be useful to everyone, which means it writes toward no one in particular.

AI writing: 4 pieces, 3 models
Llama-3.3 · literary prompt A
GLM-5.2 · literary prompt
Llama-3.1 · literary prompt
Llama-3.3 · literary prompt B
Human writing: 4 literary works
high-A essayist
imagery-rich
plain / spare
narrative
AI cross-doc spread16.5
Human cross-doc spread20.6
AI distance-to-centroid11.4
Human distance-to-centroid14.0
Fingerprints are real scored vectors from the deep corpus (35 literary works vs 59 AI pieces, apples-to-apples: literary register only). Cross-doc spread = mean pairwise distance between per-work fingerprints; distance-to-centroid = mean distance from each fingerprint to the corpus mean.
The addressivity gap

The writers readers call irreplaceable score high on A. AI never does.

A
12–15
AI range (all models)

Addressivity measures writing with a specific reader in mind: direct second-person address, imperatives, confession, a felt presence on the other side of the page. AI mean A (15.3) and human literary mean A (16.4) are nearly equal. The difference is range.

AI addressivity sits in the 1–16 range across seven models. Human literary essayists span 12 to 50. The writers at the high end are confrontational, intimate, unmistakably themselves. AI is trained to be useful to everyone. Useful to everyone is addressed to no one. That ceiling does not move.

Feel the difference · illustrative examples, not scored corpus passages

A human essay. An AI essay on the same topic. What the contrast feels like before you see the numbers.

Human literary essay
Opening
The market at dawn smells of cut celery and diesel. Three old men play chess at the corner table while vendors arrive like a tide, their carts trailing mud into the dry square. By eight the noise has become a physical thing, something you lean against.
Later in the same essay
What I want to argue is that markets are not neutral mechanisms for price discovery. They are communities of practice, spaces where trust accrues and is spent, where reputation is the only currency that holds across seasons.
AI essay: same topic, literary prompt
Opening
Markets are important economic institutions where buyers and sellers come together to exchange goods and services. They have existed for thousands of years and continue to play a significant role in modern society.
Later in the same essay
The participants in markets engage in transactions that create value for both parties. These exchanges are governed by principles of supply and demand and are essential to the functioning of economies around the world.

The human passages do different things: one puts you in a place, one makes an argument. The AI passages do the same thing twice. Change the topic, change the model: the AI passages will still do the same thing twice. Human writers have a fingerprint that travels with them. The question DOMAINS asks is whether yours still does.

Seven dimensions

What the fingerprint measures. Each cell: what it looks like high vs. low.

D
Diction
High
“the glaciers’ inexorable ablation”
Low
“the glaciers melting slowly”
O
Orchestration
High
“She ran. The corridor stretched, endless, its fluorescents buzzing; behind her the door swung on its hinge, marking time.”
Low
“She ran down the corridor. The lights were on. The door was open.”
M
Mechanism
High
“Too rich and the engine floods; too lean and it starves. The difference is a needle’s turn.”
Low
“The carburetor helps the engine work correctly.”
A
Addressivity
High
“You have felt this. The moment a room changes, not the people in it, but the room itself.”
Low
“People often feel that their environment affects their mood.”
I
Imagery
High
“wings pale as old linen, fringed with the color of dried grass, trembling against the glass”
Low
“the moth had light-colored wings and was near the window”
N
Narrative
High
“She had been waiting since before the strike, since the morning her father came home without his jacket, since the year the mill went silent.”
Low
“The strike had been ongoing for several weeks and had affected many workers.”
S
Suppression
High
“He said nothing. Later, washing dishes, he broke a glass.”
Low
“He was very upset and found it difficult to control his emotions.”

AI writing clusters in a specific region: high D and I (fluent vocabulary), A in the 1–16 range (where human literary essayists reach 12–50), moderate O and N. The consistent AI signature is not low scores across the board, but a specific shape that appears regardless of topic.

Within a single piece

Frontier AI varies within a piece. The convergence is cross-document.

Important distinction: Per-passage variation within a single work is not the differentiator for the models tested. Llama-3.3-70b and similar models can vary meaningfully passage to passage when writing a long essay. The convergence finding operates at the cross-document level: each AI piece individually may breathe, but all AI pieces across different topics converge to the same fingerprint. GLM-5.2 shows stronger per-passage flatness as well. The cross-doc collapse applies to all models tested.

The chart below is for a smaller model (gemma3:12b), where per-passage flatness is visible within a single work. Human passage SD 9.0 vs AI passage SD 4.2 in this sample.

75 60 45 Human literary essay 135710 passage AI essay (gemma3:12b, literary prompt) 1510 passage

Human: Emerson, Essays First Series (Project Gutenberg 2945), ten consecutive passages. AI: gemma3:12b, literary-persona prompt, same scorer and seed. Human passage variance (SD 9.0) is more than twice AI passage variance (SD 4.2).

Methodology

How we computed the fingerprint and the spread

For a source work divided into passages p1…pn (n ≥ 4), each passage is scored on seven dimensions (0–100) by a locally-run language model with a fixed random seed. The within-work movement score is the sum of per-dimension standard deviations across passages:

score(W) = Σk ∈ {D,O,M,A,I,N,S} σ(k across p1…pn)

The fingerprint of a work is its mean per-dimension vector: (mean(D), mean(O), …, mean(S)). Cross-doc spread is the mean pairwise distance between fingerprints across works in a corpus; distance-to-centroid is the mean distance from each fingerprint to the corpus mean.

Reproducibility

Apple Silicon (Metal)

120 of 120 paired runs produced identical scores. All results on this page use this substrate.

Cloud API (single provider)

83% of runs produced identical scores. Not used for reported results.

Cloud GPU (rented)

50% of runs produced identical scores. Renting a GPU type does not guarantee the same physical chip across runs.

Corpus. 112,475 passages across 11 human-writing registers, used to train the fast scorer. Of these, literary essay and narrative fiction (10,000 passages each) are fully deep-scored; remaining registers are approximated by a regression model calibrated on the fully scored set. The numbers reported on this page come from the 35 human literary works and 59 AI pieces fully deep-scored, not from the full 112k.

AI corpus. 550 deep-scored passages from Llama-3.3-70b, Llama-3.1-70b, and GLM-5.2 (cross-doc spread and fingerprint grid). Bare-condition generation runs across four additional frontier models (GPT-4o, Claude Opus 4.5, Gemini 2.5 Flash, Kimi K2) confirm the A finding: frontier models without a literary system prompt produce A in the 1–4 range. Cross-doc spread computed on 59 AI pieces vs 35 human literary works, register held constant.

What we built

A measurement. Not a generator.

Jasper, Notion AI, and Claude Styles take samples of your writing and teach AI to sound like you. DOMAINS works in the other direction. It scores what you actually write, across seven dimensions, and shows you where your writing sits relative to where AI consistently lands. If the gap is closing, you would see it here first.

You could try to get the same result by prompting better. You would get better writing. You would not get a consistent voice.

Every prompt call starts cold. The model has no memory of the last piece, no record of what your profile looked like six months ago, no comparison to your best-register cluster. It produces something that sounds like the instruction. Next time you prompt, it produces something that sounds like the instruction again. The two outputs converge on the instruction, not on you.

Voice is a consistency claim. It means this piece sounds like the person who wrote the last one. That comparison requires a stored reference. Prompting has no stored reference. DOMAINS does.

Score your draft →