# How the analysis works and how far to trust it > The signals behind the AuthorText score, how they add up, what the checks on real texts showed, and where the score is wrong. With the numbers. - URL: https://authortext.com/en/method - Language: en - Site: AuthorText (https://authortext.com/) The score is a weighted sum of text statistics. No hidden model decides, and no single signal can decide alone. Below is what is counted, how it adds up, and what happened when we tested it on texts with known authors. The test results are mixed, and we say so. ## What the number is A text goes in, about twenty measurements come out: how much sentence lengths vary, how many stock phrases there are, how dense the linking words are, how rich the vocabulary is, and so on. Each measurement earns a few points when it looks like machine writing. The points are added up. That is all. The score is not a probability that a machine wrote the text. It says how many of the usual traces of generated text are present. ## The signal families Most of the weight sits in the first two families. The rest add small amounts. ### Rhythm of sentences and paragraphs People write short and long sentences in a ragged mix. Generated text tends to keep a steady length. We measure the spread of sentence lengths and of paragraph lengths. Even sentences give up to 32 points, even paragraphs up to 28, and both together add up to 14 more. ### Stock phrases About 150 phrases in four languages, sorted into three tiers by how often real people use them. Phrases that never showed up in our human texts count most: two of them give 24 points. Common ones give almost nothing. ### Linking words The number of connectives per sentence, such as "moreover" or "in addition". A dense run of them gives up to 13 points. ### Repeated templates The same short heading line again and again, or the same sentence shape over and over. Up to 14 points for repeated headings, a few more for repeated patterns. ### Vocabulary and small signs Variety of words, entropy, word length, punctuation, passive voice, readability. Two of them (variety and entropy) can give 15 points, the others between 3 and 5. Unusual symbols and hidden characters are reported and add a few points. ## How the points add up The points are summed. No single measurement can give more than 32, and the sum before trimming can reach about 165, so one odd feature never decides the result. A short text is multiplied by a factor between 0.9 and 1, which keeps long texts from gaining and short ones from being punished twice. The result is capped at 100. Short texts get partial points only. Under 150 words or five sentences the page says the result is unreliable, and it says the same for promotional and social-media style, which the method was never tested on. ## Reading the bands Below 20: no signals. From 20 to 34: weak signals. From 35 to 54: a mixed profile. From 55 to 69: probable AI. From 70: strong AI-like style. In our first calibration a threshold of 50 to 55 worked for "probable" and 70 for "strong". Treat the bands as a way to sort texts for a closer look, never as a ruling. ## What the checks showed First, a small calibration in version 0.4.0. We took 34 confirmed answers of AI services (8 services, 4 languages) and 20 human texts. At a threshold of 50 it found 93.9% of the AI texts with 0% false positives; at 70 it found 60.6%, again with no false positives. It looked excellent. The set was small, and it covered few genres. Then we tested on a much larger corpus: 5808 human texts, 1286 AI texts and 299 mixed ones. The picture was far weaker. AUC, the chance that a random AI text scores above a random human one, was 72.9%. With the threshold set so that only 1% of human texts get flagged, the method found 14.4% of the AI texts. At 5% flagged humans it found 29.2%. So the first result was too rosy, and the second one is the number to remember. We keep both on this page on purpose. ### By length Texts of 80 to 299 words are close to a coin toss: AUC 55.6%, recall 4.4%. From 300 words it is better: AUC 76.7%, recall 16%. ### By language Recall at 1% false positives: Russian 22.7%, Ukrainian 19.1%, Portuguese 8.5%, English 6.0%. The Portuguese and Ukrainian language resources have not been checked by native speakers yet. ### By model Newer models slip past more easily. At 1% false positives the method found 19.7% of gpt-oss-120b texts and 6.7% of GLM-5.3 texts. Early web chats were easier to spot than today's models. ### Mixed texts A human text edited by a model was flagged in 0.0% of cases at the strict threshold and 9.7% at the loose one. Machine translation: 2.0% and 4.0%. The method does not really separate such cases, which is why the page speaks of signs and not of "percent AI". ## Where it gets things wrong False positives come mostly from careful, even prose. News texts are the worst case: AUC 55% and 14.6% of human texts flagged at the loose threshold, so for news the signal does not work and the page should abstain. Scientific articles reached 10.0% false positives at the same setting. Encyclopedia style was fine, 0.5% and 4.3%. Volunteer translations were flagged a bit more often than other human texts, 2.3% against 1.1%. Non-native Ukrainian writing was indistinguishable from native writing at the size we could test, 1.5%. The other direction fails too: a well made modern machine text often passes. And the tool has no idea what you asked a model to do before you pasted its answer. ## How to use a result Read the evidence places, not only the number. If three quoted sentences really look like filler, that tells you more than a 62. Compare with the author's earlier work. Ask for drafts. Talk to the person. A learned scorer is shown beside the main score as a provisional second opinion; it has had far fewer tests, so give it less weight still. ## Related reading - [How to read the numbers in an AuthorText analysis](https://authortext.com/en/articles/reading-the-numbers.md) - [When the detector mistakes a human for AI](https://authortext.com/en/articles/when-detectors-disagree.md) - [Open the checker](https://authortext.com/en/index.md) - Contact: support@authortext.com ## Links - [Home](https://authortext.com/en/index.md) - [plans](https://authortext.com/en/plans.md) - [help](https://authortext.com/en/help.md) - [developers](https://authortext.com/en/developers.md) - [about](https://authortext.com/en/about.md) - [method](https://authortext.com/en/method.md) - [faq](https://authortext.com/en/faq.md) - [privacy](https://authortext.com/en/privacy.md) - [terms](https://authortext.com/en/terms.md) - [articles](https://authortext.com/en/articles/index.md)