How to read the numbers in an AuthorText analysis

The result page gives you a score, a band and a set of bars. Here is what they mean in plain words, and how far you can trust them.

The score is a hint

The score runs from 0 to 100 and describes how many features of machine-like text the analysis found. It is a heuristic, not a measurement of authorship. We repeat this on purpose: it is the most important thing on the page.

Sentence-length variation

People write unevenly. A long sentence with a clause or two, then a short one. Then a medium one. Machine text, especially from earlier models, tends to keep sentences close to the same length.

The analysis measures this as a coefficient of variation: how much sentence lengths spread around their average. Low spread adds points, high spread adds none. This is often called burstiness, and it is the heaviest signal in the score. Paragraph lengths are measured the same way, and when both are uniform together, a joint bonus applies.

Repetition

Repeated words, repeated openings, the same template header in every section. The analysis counts how often patterns recur. Some repetition is normal in a technical text, so this signal is only one part of the picture.

Typical phrases

The analysis knows a list of stock phrases in English, Russian, Ukrainian and Portuguese, sorted into tiers by how typical they are. Openers that announce a point before making it, and vague claims about the "important role" of everything, belong there. Each hit points to a line number in the evidence list.

Connective density

These are the words that tie neighbouring sentences together and signal the next step of an argument. Machine text tends to use many of them, evenly spread. A person often just starts the next sentence.

A small worked example

This is an invented illustration, not real data.

The project is important. The team is large. The plan is clear. The goal is simple. The result is good.

Five sentences, two to four words each. The rhythm is nearly flat, so sentence-length variation is low and that signal would score. Now a rewrite:

The project matters. Why? Because the team, which grew from three people to twelve in a year, finally has a plan everyone can explain in one breath. The goal is simple.

Here the lengths range from one word to over twenty. Variation is high and that signal stays quiet. The content is not "more human"; only the rhythm changed. That is exactly why rhythm alone proves nothing.

What "probable AI" means

The score maps to bands. Around 50 to 55 and above is "probable AI", and 70 and above is "strong AI". Those thresholds came from a first calibration on a small set: 34 AI texts against 20 human ones. There it looked good.

On a larger corpus the picture is much more modest. In our latest measurement the analysis found about 14% of AI texts at a 1% false-positive rate, and about 29% at 5%. Modern models evade the current signals better than the early chatbots. So the band is a statement about style, not about the author, and "probable" should be read as "worth a closer look".

Mixed cases are not separated well either: a human text edited by a model, or machine-translated text. That is why we talk about signs of generation in a text, not about an "AI percentage" of its author.

Why short texts give weak signals

Variation needs material. With three sentences you cannot tell a rhythm from a coincidence. In our corpus measurement, texts of 80 to 299 words were close to random, while 300 and more words did better, though still far from reliable. The analysis marks texts under roughly 150 words, or under five sentences, as unreliable, and scales the points down.

Also, a short text with a low score is not evidence of human writing. The note on the page says exactly that: the verdict is unreliable, not proof.

News, promotional and social-media genres are another weak spot, and the analysis may abstain there.

What to do

Read the evidence list before the number: it shows places, not just a total.
Look at the text length and the confidence note.
Treat the band as a reason to reread. If a person's reputation or grade depends on it, ask for drafts instead, as in our fair checklist.
If you are the author and want a more natural rhythm, see how to make your text livelier.

Want to try it on your own text? Open the checker.