Paraphrase, Rinse, Repeat: What a Humanizer Actually Does to Your Sentences

A humanizer has one job: make text look less like a language model wrote it. Not better. Not clearer. Less detectable. Those are different targets, and the gap between them explains almost everything strange about humanizer output.

Here’s what is happening under the hood, from the crude to the clever.

Level one: synonym swapping

The oldest trick predates generative AI entirely. Plagiarists have used “article spinners” for years to dodge plagiarism checkers, replacing words with thesaurus entries until the text no longer matches its source.

The results can be spectacular. In 2021, researchers documented what they called “tortured phrases” in published scientific papers — paraphrasing artifacts like “counterfeit consciousness” for artificial intelligence and “profound learning” for deep learning. The phrases were so odd that they became a way to spot papers that had been run through a spinner.

Modern humanizers are more careful, but the instinct survives: swap common words for less common ones, because less common words raise perplexity.

Original	Swapped
The results show a clear trend.	The outcomes exhibit an unmistakable tendency.
This approach has limitations.	This methodology possesses constraints.
Students often struggle with citations.	Learners frequently grapple with attributions.

Constructed examples, typical of the pattern.

Each swap is defensible on its own. Stack enough of them and you get prose that sounds like it was translated into English by a committee.

Level two: sentence surgery

Detectors look at rhythm as well as vocabulary, so humanizers break up the rhythm. Long sentences get split. Short ones get merged. Fragments appear. A rhetorical question gets bolted on. Contractions get sprinkled in to sound casual.

Done well, this is just editing. Done mechanically, it produces a recognizable texture: variation for its own sake, with short punchy sentences arriving at random rather than where emphasis is needed.

Level three: planted imperfection

Some tools add errors deliberately — a lowercase letter where a capital belongs, a missing comma, an “it’s” where “its” should be. The logic is that models rarely make these mistakes, so mistakes look human.

It sometimes moves the score. It also means you are paying a service to make your writing worse, and then handing that writing to someone who will grade it.

Level four: model-on-model rewriting

The more sophisticated humanizers don’t use rules at all. They use another language model, prompted or fine-tuned to rewrite text in a way that scores as human on detectors. In effect, the detector becomes the training signal.

Academic work has shown how effective this can be: a 2023 study found that running AI text through a dedicated paraphrasing model dramatically reduced detection rates across several detectors, while largely preserving meaning.

The catch is that this is the easiest technique for detector companies to fight, because they can generate unlimited training examples by running the same humanizers themselves. Which is how the arms race keeps turning.

What it costs

Every humanizer technique trades something away:

• Precision. Synonyms are rarely exact. Technical terms are the first casualties.
• Voice. Mechanical variation flattens whatever personality the original had into a generic “casual” register.
• Meaning. Aggressive rewriting can quietly change claims, especially hedges and qualifiers.
• Trust. Text that has been humanized often has a texture experienced readers learn to recognize, detector or not.

The irony is hard to miss. A humanizer’s output is optimized against a statistical model of human writing. It is not optimized for an actual human reading it.