A growing number of tools promise to take AI-generated text and make it undetectable — run your draft through, and the result slips past Turnitin. What that pitch leaves out is what a humaniser actually changes, and what it doesn't.
These tools work on the same statistical signals a detector looks for: how predictable each word choice is, and how uniform sentence length and rhythm are across a passage (the same measures covered in our guide to how AI detectors actually work). A humaniser varies sentence length, swaps in less predictable phrasing, and breaks up the uniform rhythm that gives AI text away statistically. What it doesn't do is check whether anything the original draft claimed is actually true. It's a rewriting layer, not a fact-checking one.
If the AI tool that produced the original draft invented a statistic, misattributed a finding, or cited a paper that doesn't exist, a humaniser doesn't catch any of that. Checking a claim against a real source was never the job it's built for. It reworks the sentence around the error and hands it back with the same mistake, now in unfamiliar phrasing that's harder for you to spot as wrong, since you didn't write either version and the second one doesn't even match the first. Every layer between you and the actual source is a layer where an error can hide, and a humanised AI draft has two.
Several UK universities' academic misconduct policies make this point explicitly: it's undisclosed AI use they define as misconduct, not a specific detector score. De Montfort University's policy, for example, lists submitting AI-generated content without acknowledgement, or using AI to "substantially amend or improve" work without declaring it, as intentional misconduct — a rewritten, "humanised" version is still undisclosed AI use, whatever score it gets. York St John's policy treats careless or inappropriate AI use as misconduct under its plagiarism and fabrication provisions in much the same way, regardless of how the text was produced. Passing a detector was never the actual bar.
The exact process varies by institution, but a tiered pattern is common. A first offence often means losing the marks for that piece of work, sometimes with a capped-pass resit, and a module can be failed outright. Repeat offences, or misconduct treated as more serious (final-year or postgraduate work, for instance), escalate toward suspension or enforced withdrawal. The most serious cases, including essay-writing services and comparable attempts to pass off someone else's work as your own, can mean expulsion, and under some policies a degree can even be revoked after graduation if serious misconduct is discovered later. None of this hinges on whether a detector was fooled at the time.
A humaniser is solving the wrong problem. The actual risk was never a detector catching a particular writing style. It's submitting words, facts, or citations that aren't genuinely yours or genuinely checked. Finding and verifying the real paper behind a claim removes that risk at the source, rather than disguising it one more rewrite downstream.
The honest takeaway: a humaniser changes how AI-generated text scores, not whether it's genuinely yours or genuinely accurate. Those are the two things that actually matter, both to your university and to the argument you're trying to make.