Can AI detectors spot “humanised” text?

Humaniser tools promise to make AI-written text undetectable. The user pastes in an essay from AI, the tool swaps words, varies sentence length and adds the odd imperfection, and the result is meant to read as if a person wrote it. They’re widely used by students, by job applicants and by paper mills.

So do they work?

What a humaniser actually changes

A humaniser rewrites the surface layer of the text. It changes word choice, rhythm of sentences and punctuation as those are the signals that simple detectors rely on. That’s why a basic detector can be fooled.

What a humaniser doesn’t change is the structure underneath: how the argument is built, how ideas follow on from each other, and the patterns a language model leaves in what it chooses to say. A detector trained on humanised text learns to look past the rewording to those deeper patterns.

Why section-by-section results matter

Real documents are rarely all AI or all human (they are hybrid). A student could write an introduction, generate the middle with AI, run it through a humaniser, then edit the conclusion themselves. If an ai detector gives one overall percentage, it hides this.

A better approach is one that marks each paragraph and sentence. The user can then see where the AI-written parts are, whether they come in blocks or are scattered through, and what to ask the writer about.

How AI Aware handles it

AI Aware’s text detector is trained on AI text that has been humanised and edited, as well as on clean AI and human created writing. It highlights the sections that are most likely to be AI-written, and it tells you when text appears to have passed through a humaniser. On our latest held-out tests, it is 98.6% accurate on text.

You can check a document with AI Aware’s AI text detector. Your first 3 checks are free, with no sign-up.

Comments

Popular posts from this blog

How schools and universities can use AI detection fairly