AI humanizer detector

An arms race with fingerprints

Once detectors appeared, so did tools promising to defeat them. A humanizer takes the polished, predictable output of a chatbot and makes it less predictable: it swaps common words for rarer ones, varies sentence length, adds contractions and filler, and sometimes introduces deliberate small errors. More aggressive tools tamper with the characters themselves, inserting invisible Unicode between letters or replacing Latin letters with identical-looking Cyrillic or Greek ones, so that detectors reading the text as tokens see gibberish.

Each trick removes one kind of evidence and adds another. GPTTrace looks for both.

What survives humanizing

Structure

Rewriting at the word or sentence level rarely changes the shape of the answer: an introduction that restates the question, three balanced points, a tidy conclusion, paragraphs of nearly equal length. GPTTrace’s paragraph-uniformity measure and the structural signs from Wikipedia’s guide often still fire after a rewrite.

Hidden characters

Zero-width spaces, word joiners and soft hyphens are invisible on screen but present in the text. The highlighted copy below each result shows them as red labels. Scattered inside ordinary words, they are a strong sign of deliberate obfuscation; see the invisible character detector for details.

Thesaurus damage

Synonym swapping produces phrases no fluent writer would choose — “profound learning” for “deep learning”, “colossal information” for “big data”. A reader notices these at once even when a detector doesn’t. If a text has several such phrases, ask the writer what they meant.

Statistics

Humanizers aim to raise sentence-length variation, and some overshoot into an erratic pattern. The deep scan’s neural model, trained on a benchmark that includes paraphrased and adversarially modified text, is more robust to these rewrites than word lists alone.

A word of caution

Many people edit their own writing heavily, use a thesaurus or write in a second language, and their text can look “humanized” for innocent reasons. As with every detector result, use the highlighted evidence to start a conversation, not to settle one.

Homoglyphs: letters that only look English

Some humanizers replace Latin letters with look-alikes from other alphabets — a Cyrillic “а” or “о”, a Greek “ο” — that render identically but are different characters to a computer. Search engines, plagiarism checkers and detectors then fail to match words. A quick test is to search the text for a common word you can see on screen: if your browser’s find function can’t find “report” in a sentence that clearly contains it, the letters have been swapped. Pasting the text into a plain-text editor with a non-Latin-aware font sometimes makes the substitutes visibly different. Text with homoglyphs has almost always been deliberately manipulated, whoever wrote the original.