Writing Distinctiveness Score: What It Measures, and Why It Matters

You can describe your writing as "direct but warm" and mean it sincerely. So can a few hundred thousand other people. A writing distinctiveness score exists because adjectives cannot tell you whether your prose is actually recognisable, and neither can the AI model you are asking to imitate it.

TL;DR: A writing distinctiveness score is a 0 to 100 number that measures how far a sample of your writing sits from the average output of current AI models, across six measurable dimensions of style. It is a distance, not a grade. Under 30 means your sample reads close to generic AI, 30 to 59 means a fingerprint is emerging, 60 and above means a strong fingerprint. The number is built from sentence-level measurements, which is exactly the layer a tone label never touches. You can measure your own score for free with anything you have written.

What Is a Writing Distinctiveness Score?

A writing distinctiveness score is a single number that summarises how different your writing is from what an AI model produces when nobody has told it anything about you. It is computed by measuring your text on six style dimensions, measuring the average AI output on the same six, and reporting the distance between the two points. The higher the number, the more of your voice an unconfigured model would flatten.

The score is deliberately symmetric. Writing that is far more formal than the AI average and writing that is far less formal both count as distance. The measurement is not asking whether you are "better" than the machine on any axis. It is asking whether you are somewhere else.

This matters because the problem most professionals have with AI drafts is not that the drafts are bad. It is that they are interchangeable. Three colleagues ask the same model for "a concise, professional update" and get three updates that read as one author. The score quantifies how much you personally stand to lose in that exchange.

Why a Tone Label Cannot Produce a Score

Ask a model to write "in a confident, friendly tone" and it will comply, in the sense that it will produce text a reader would plausibly call confident and friendly. What it will not do is produce your confident and friendly. The model does not know your tone. It knows the words that describe tone, and it maps those words onto the middle of everything it has read that was labelled the same way.

A voice does not live in the adjective. It lives in choices that are almost never described because the writer does not notice making them:

• How long your sentences run, and how much that length varies inside a paragraph.
• Where you put your qualifiers: "probably worth a look" versus "worth a look, probably."
• Whether you start sentences with "And" or never do.
• How you handle transitions: a blunt new paragraph, a "So," a colon, a rhetorical question.
• How often you use parentheses, dashes, semicolons, or none of them.

None of those show up in "confident and friendly." All of them show up in a stylometric measurement, which is why a score can be computed and a label cannot. We covered the adjective problem at length in why tone adjectives are not your voice. The score is what you get when you take that argument seriously and build the instrument.

The Six Dimensions Behind the Number

The score is built on the same six axes our full Writing DNA analysis uses. Each is scored 0 to 100 for your sample and for the AI baseline.

Two things stand out in that middle column. First, the generic model is not neutral. It runs long on complexity, high on expressiveness and low on conciseness, which is a precise description of the "polished but padded" feel people complain about. Second, none of these axes is a tone word. Formality is the closest, and even that is measured from marker counts rather than declared.

The tool also names your most distinctive axis, meaning the dimension where your gap from the baseline is largest, and it phrases that gap as character rather than deficit. A writer who scores 20 on expressiveness against a baseline of 76 is not told they are flat. They are told they stay measured where generic AI over-emotes. Both readings are true. The second one is the one that helps you configure a model.

How the Generic-AI Baseline Was Measured

A distance is only as honest as the point you are measuring from. The baseline behind the score comes from 320 samples generated across five models (Claude Opus 4.6, Sonnet 4.5, Haiku 4.5, GPT-5.2 and Gemini 3 Pro), four languages (English, French, Spanish and Japanese), eight prompt types and two variants each. Every sample went through the same measurement pipeline your text goes through, and the per-locale averages became the baseline. The values were last measured on 15 February 2026, and the date is shown on the results page so you know what you are being compared against.

The reason for eight prompt types is that a model's style shifts with the task. A cover letter and a product description from the same model do not land on the same point. Averaging across task types gives a centroid that represents "AI wrote this and nobody configured it" rather than "AI wrote a cover letter." The full method, including what we left out and why, is in how we measure average AI.

One caveat we publish rather than bury: the five models do not agree with each other on every axis. On expressiveness in particular they spread widely, which we wrote up in what the average AI baseline hides. The baseline is a centre of gravity, not a single model's fingerprint. If you only ever use one model, your real distance from that model may be a little larger or smaller than the score suggests.

How Is the Score Calculated?

The score is the straight-line distance between your six-dimensional point and the baseline's six-dimensional point, divided by the largest distance possible, then scaled to 0 to 100. Each axis contributes the square of its gap, the six squares are summed, and the square root of that sum is compared against the maximum a six-axis space allows.

The arithmetic has a consequence worth knowing before you read your number. Because the scale is normalised against the theoretical maximum, scores are hard to inflate. A writer who sits 25 points from the baseline on every one of the six axes scores 25, not 60. To reach the "strong fingerprint" band at 60, your average gap across all six dimensions has to be around 60 points. That is a large, consistent departure from the middle, which is what a strong fingerprint should require.

The same arithmetic also means one extreme axis cannot carry the score alone. Being 80 points below the baseline on expressiveness and dead centre on the other five yields a score in the low 30s. Distinctiveness, as the tool defines it, is a property of the whole profile.

For context on the measurement tradition this sits in, the stylometry literature has used exactly this kind of multi-feature distance to attribute authorship for over half a century, long before anyone needed to configure a chatbot.

Reading Your Number: Three Bands and a Reliability Flag

The results page reports the score with one of three readings and, alongside it, a reliability flag driven by how much text you gave it.

• 60 and above: strong fingerprint. Your sample is clearly distinct from generic AI. This is the kind of style that is worth extracting and deploying, because an unconfigured model would lose most of it.
• 30 to 59: fingerprint emerging. A few dimensions already set you apart. More text usually sharpens the picture in one direction or the other.
• Under 30: close to generic. Your sample reads near the AI average. This is more common than people expect, and it is not a verdict on your writing. Short, task-shaped texts (a status update, a meeting note) tend to sit here because the task constrains the style.

The reliability flag matters as much as the band. Stylometric measurements are noisy on small samples because one long sentence or one exclamation mark shifts an average. Below roughly 1,000 words the tool labels the reading indicative. Above 2,000 words, which is the bar the research on style extraction converges on, the reading is marked reliable. We explain where those thresholds come from in how many writing samples AI needs.

A practical rule: if your score is under 30 on 150 words, you have learned almost nothing. If it is under 30 on 2,500 words drawn from three different contexts, you have learned something real, and the interesting question becomes which axes are holding you near the middle.

Why Better Models Make the Score More Useful, Not Less

A reasonable objection: the models keep improving, so surely the next release will just write like me without any of this.

It will not, and the reason is structural. Model capability and user-data conditioning are separate layers. The labs improve the first. They cannot supply the second, because the second is your writing, and they do not have it. A model that is twice as good at following nuanced instructions is twice as good at following your instructions once you have some worth giving it. Without them, it is twice as good at producing the centroid.

The score is a measurement of that second layer. It tells you the size of the gap that instruction-following capability alone will never close, because the gap is made of facts about you rather than facts about language. As the models get better at executing a detailed Style Profile, the value of knowing your distance from the baseline goes up, not down. The instrument becomes more actionable each time the thing it feeds gets more capable.

The vendors' own documentation points the same way. OpenAI's custom instructions exist because the model needs to be told what it cannot infer. The score tells you what to put there.

What the Score Does Not Tell You

Every measurement has edges, and it is more useful to know them than to discover them.

• It does not judge quality. A confusing, over-long writer with wild sentence variance can score very high. Distance from the middle is not the same as being good.
• It does not compare you to other humans. The reference point is the AI centroid. Two people with identical scores can have completely different voices, sitting on opposite sides of the baseline.
• It is language-specific. English, French, Spanish and Japanese each have their own measured baseline. A score in one language says nothing about your voice in another, and the tool hides the comparison for any language it has not measured rather than borrowing the English numbers.
• It is a snapshot of one sample. Your voice shifts by context. A distinctiveness score on your LinkedIn posts is not a distinctiveness score on your board memos. The full Writing DNA analysis exists precisely because one sample cannot carry every register.
• It is indicative below the word bar. The tool says so on the page. Take the flag seriously.

How to Use the Score This Week

The number is more useful as a diagnostic than as a trophy. Three ways to put it to work.

1. Find your real fingerprint, not the one you would have described. Paste 2,000 or more words you actually wrote across at least two contexts. Look at the most distinctive axis the tool names. That axis is the first thing an AI draft loses when you give it a tone label instead of a rule. If your gap is on conciseness, the rule is about words per idea. If it is on expressiveness, the rule is about questions and emphasis.

2. Test the tone-label hypothesis on yourself. Ask ChatGPT or Claude for a paragraph "in a direct, warm, professional tone" on a topic you have written about. Score the result. Then score your own paragraph on the same topic. The AI draft will land near the baseline, because the baseline is what those words mean to a model. Your paragraph will land wherever you actually are. The distance between the two is what a tone label costs you.

3. Score before and after configuring a model. Give the model explicit rules derived from your measurements (target sentence length range, banned openers, punctuation habits) and score its output again. A configured draft should move toward your point. If it does not, the rules are not specific enough yet, and the six-dimension breakdown shows which one to tighten.

The last exercise is the most persuasive one, because it makes the argument in numbers rather than adjectives. Trained beats prompted, and the score lets you watch it happen.

From a Score to a Style Profile

The free tool runs the same analysis engine as the full product, on a single sample. A Writing Style Profile runs it across your whole corpus, separated by context, and turns the measurements into rules a model can follow: a Writing Twin you deploy in ChatGPT, Claude, Gemini or Copilot rather than a number you look at.

The score answers "how much of my voice is at stake." The profile answers "what, specifically, does a model need to be told to keep it." Both are built from the same six dimensions, and both start with the same act: giving the instrument something you actually wrote.

Frequently Asked Questions

Is the score stored or shared? Your submission is analysed on our servers and kept privately alongside the result so you can retrieve the breakdown. It is not published, and it is not used to train anything. The privacy page covers exactly what happens to text you paste.

Does the score change if I paste the same text twice? No. The measurement is deterministic. The same text against the same baseline produces the same number. It will only change when the baseline is re-measured, and the results page shows the measurement date so you can tell.

Why does my score drop when I paste a very short email? Short, task-shaped writing has less room for your habits to show. A three-line scheduling email is constrained by the task more than by you. The tool flags results under the reliability bar as indicative so you do not over-read them.

Which languages have a measured baseline? English, French, Spanish and Japanese. For any other language the comparison is hidden rather than computed against the wrong centroid.

Can the score tell me which AI model writes most like me? Not directly. It compares you to the average of five models. The per-model spread is real, especially on expressiveness, and we published it separately so you can see where the individual models sit.

Ready to see your own number? Measure your Voice Distinctiveness Score with anything you have written, free and without an account, and read the six-dimension breakdown before you write another tone label.