How AI text humanisation works
Measure, rewrite, check. What happens to the text step by step and where the rewrite stops.
GPTRevealer first measures the text — how evenly long its sentences are, how many phrases typical of AI it contains and what repeats in it — and from that works out the machine feel on a scale from 0 to 100. Then a language model rewrites it to suit the chosen document type, keeping the facts, numbers, names, quotations and the order of ideas, and adds small human irregularities to the finished text. The result is measured again and goes through four checks; anything that fails is repaired once. Machine feel is our own measure of style, not a detector score, so humanisation does not guarantee that the text passes a detector.
What happens before the rewrite
First the language of the text is identified — Czech, Slovak, English or German. It decides which list of phrases typical of AI is used, which common words are not counted as repetition and which rules the rewrite follows.
Then the text is measured. The language model gets the results as part of its instructions: it knows how regular the sentence rhythm is, which phrases to remove and what repeats in the text. The same measurement runs again after the rewrite, so the analysis shows the original and the humanised text side by side.
What humanisation measures
Three properties of the text and the machine feel built from them. Everything is calculated on our side, without a language model, so the same text always gets the same numbers.
Machine feel
An overall number from 0 to 100, shown before the rewrite and after it. Half of it comes from sentence uniformity, 30 % from phrases typical of AI and 20 % from repeated words and collocations.
Limit:It is a measure of style, not a detector score. A low number does not mean the text will pass a detector, only that it has fewer of the properties we can measure.
Sentence uniformity
How close in length the sentences are. It is calculated from the variance of sentence lengths: the more the lengths differ, the lower the number. AI text tends to have even sentences, while a person mixes short and long ones.
Limit:In a short text with only a few sentences the number is unstable. Some human texts have a regular rhythm too, such as instructions or official letters.
AI phrases
How many times phrases typical of generated text appear, such as “in today's world”, “plays a crucial role” or “in conclusion”. Each of the four languages has its own list. Set clichés count twice as much as single overused words.
Limit:The list only catches known phrases. It does not recognise new or less common ones, and people use the same phrases too.
Repeated collocations
Words and three-word sequences that repeat in the text. Common words such as conjunctions and prepositions are not counted, and neither are numbers and citation details.
Limit:In a technical text, repeating a term can be correct. The check flags it anyway and the rewrite tries to vary it.
How the rewrite works
Four steps, one after another. Deep rewrite is optional and only runs with an account; the demo without signing in skips it.
Rewrite by document type
The language model receives the measurement results and the rules for the chosen document type — an essay is rewritten differently from an e-mail or an academic paper. It has to keep the facts, numbers, names, quotations, the order of ideas and the paragraphs, stay close to the original length and alternate short and long sentences. A long text is rewritten in parts of roughly 3,000 characters, and each part picks up where the previous one ended.
Limit:The model follows the rules, but not necessarily to the letter. That is why a check follows the rewrite.
Deep rewrite
An optional step. The rewritten text is translated into English and back — or, if it is already English, into German and back — which breaks up what is left of the machine rhythm. The result is used only if the word count changes by no more than a fifth. Otherwise it is discarded and the process continues with the ordinary rewrite.
Limit:It takes longer and holds technical terms less well. For a text with precise terminology it is safer to leave it off.
Human irregularities
Small touches a person would write are added to the finished text: a short sentence reacting to the previous one, an aside in brackets, a conjunction at the start of a sentence or a clarification. The changes touch 15–25 % of the sentences; the rest stay as they are.
Limit:Facts and meaning must not change, but the style does shift. With a formal text it is worth reading the result and taking out any irregularities that do not belong there.
Check and repair
The result is measured again and goes through four checks: sentence rhythm, AI phrases, repeated collocations and length, where the word count must not change by more than a quarter. If one of them fails, the model repairs the text once, exactly where the check found the problem, and the measurement runs again.
Limit:The repair happens only once. Anything that still fails after it shows up in the analysis as a failed check.
Writing in your style
Humanisation can keep to the style you write in. From your sample texts in the same language it works out your average sentence and word length, how much you vary the length of your sentences, how many commas, dashes and question marks you use and which linking words you use most often. It also adds short excerpts from three samples to the instructions, so the rewrite moves towards your own voice rather than a generic “human” style.
Every finished humanisation that passes the checks becomes a sample on its own. With the Pro plan you can also upload your own earlier work, each piece at least 250 characters long. There is room for at most twenty sample texts; once it is full, a new humanisation replaces the oldest of the humanisations and uploaded work stays.
What humanisation cannot guarantee
It cannot guarantee that the text passes a detector. Each detector measures something different, and machine feel is only our own measure of style. The rewrite removes the properties we can measure, but another tool may still flag the text — and no AI text detector is one hundred per cent accurate in either direction.
Nor does it change where the text comes from. Humanisation changes the form, not the content: the ideas, sources and conclusions stay the ones you put in. Whether you may submit or publish the rewritten text is decided by the rules of your school, publisher or client, not by a number.
Always read the result. In a technical text, check the terms and citations; in a formal letter, the tone. If you want to know how the text fares with detectors, run it through detection after the rewrite — it shows seven detectors side by side.
Full humanisation with a free account
Longer texts, the deep rewrite, analysis before and after, and a history of your humanisations. The account is free and needs no card.
Open humanisation