Will humanised text pass an AI detector? What the studies show
Sometimes yes, sometimes no, and there is no way to guarantee it. Independent studies show that detectors often get it wrong after humanisation or simple edits, but the result varies with the tool, the detector and its version, and detectors keep learning to recognise humanised text. No AI text detector is a hundred percent accurate in either direction. Humanisation therefore makes no sense as insurance against being found out, only as an edit to the style of a text the author reads through and stands behind.
What the studies say
A peer-reviewed 2024 study by Perkins and colleagues tested six detectors on generated texts that the authors then edited in simple ways. The accuracy of all six fell after the edits, and the authors advise against deciding on academic misconduct on the strength of a detector alone. More recent work by a team at the University of Notre Dame, not yet peer-reviewed, reached a similar conclusion for texts that had gone through a commercial humaniser.
Other results point the other way. A 2026 study found that detectors mostly kept flagging text from QuillBot's paraphraser, although the result varied considerably from one attempt to the next. So neither holds reliably — neither that humanisation always gets through, nor that a detector always spots it.
Why the results differ
Each detector measures something different, and a humaniser only changes what it can measure itself. When a rewrite breaks up the sentence rhythm and throws out worn-out phrases, a detector that relies mainly on those will show a lower number. A detector that judges the text as a whole with a different language model may keep flagging it.
Detectors also change. In August 2025 Turnitin announced a feature meant to recognise text edited by a humaniser, and there is still no independent check of how well it works. A result from last month may not hold today, and a tool that promises permanent undetectability is promising something outside its control.
What follows from this
For authors: humanisation is not insurance. If the rules require you to declare your use of AI, a rewrite changes nothing about that, and if you cannot explain a text, you will not defend it even with a low score. It makes sense as an edit to the style of a text you stand behind — read it, correct it and, where needed, declare it.
For teachers and editors: a low score does not mean a person wrote the text, and a high one does not mean AI did. A detector result is a starting point for a conversation, not proof. What usually decides it is whether the author can explain the content and show how the text came about.
FAQ
Does DetekceGPT promise that humanised text will pass?
No. Our humanisation measures its own machine feel, that is sentence rhythm, AI phrases and repetition, and shows how it changed. It is not a detector score, and we do not promise to get past a detector.
Does it help to humanise a text several times in a row?
Each further round takes the text further from the original wording and raises the risk that the meaning shifts. It will not reliably get you past a detector, and a text you no longer recognise yourself is a worse outcome than a higher score.
Should I check humanised text with a detector?
It shows you which passages still read as machine-made, and that is useful. But do not take a low number as permission — what decides is the rules and whether you can defend the text.
Try it on your own text
Paste a text or upload a file and look at the score and at the specific sentences that came out suspicious.
Run a detection