Put Humanizers Under the Disclosure Clause, Not a Banned-Tools List

Under disclosure and misrepresentation of authorship, beside the definition of plagiarism. Pangram Labs defines a humanizer as a tool that rewrites AI-generated text to evade AI detectors ¹. That's a description of concealment, not of a new kind of misconduct. The conduct a policy already names is handing in generated work as a student's own. A humanizer is the method. A method doesn't need its own chapter.

An existing plagiarism definition can reach the case already. The University of Southern California's academic integrity office treats material created by generative AI and presented as a student's work as plagiarism, whether it was copied verbatim, near-verbatim or paraphrased ². A humanizer hands back the paraphrased version of exactly that.

What an older template can lack is the disclosure half. The TeachAI sample policy puts it in one line: if a teacher or student uses an AI system, its use must be disclosed and explained ³. The same guidance tells teachers not to use detection tools at all, on the ground that their accuracy is questionable ³. A school that keeps a detector has already departed from that advice, which is one more reason its policy needs the evidence row in the table below.

A humanizer sits inside that one line. Running a draft through one is a use, and what comes back is a draft that has to be declared.

The fourth row keeps the first three usable. The wording for it is worked through in the AI clause in an academic integrity policy.

Because the category is far larger than any list a handbook will carry. NBC News reported in 2026 that Turnitin keeps a list of 150 tools that adjust text so a detector will not flag it, some charging as much as 50 dollars for a subscription ⁴. Cursive, an academic integrity software company, counted 43 humanizer sites drawing 33.9 million visits in a single month ⁴.

The products also make claims about the policy itself. Undetectable AI's homepage sells a rewrite that will "Score as human written on AI detectors and improve readability" ⁵. BypassGPT goes further. It tells students its tool helps them "submit essays and assignments without breaching any academic guidelines" ⁶, which is a statement about a student's own institution, made by software that has never read its rules.

A student who believes it has been told, by a product page, that the pass is allowed. The policy is what answers back, and it answers longest when it describes the behaviour rather than the brand, because products come and go and the sentence about concealment does not. What that sentence has to cover is what a school actually sees when a paper has been rewritten.

Three sentences do the work. The first defines the conduct and the second sets the standard of evidence. The third says what happens next. Start with the definition: presenting AI-generated or AI-rewritten text as a student's own work, or concealing an AI tool used at any stage, violates the integrity policy. That wording covers a humanizer without naming one. It also covers the tool that launches next term.

The evidence sentence is the one that gets tested. A detector reading is one input into an integrity process, and it is never on its own a finding of misconduct. Undetectable AI's own ethics page states that "A single AI text detection score alone is NOT enough to rule out cheating or deception" ⁷, and the same page says the company has never condoned cheating and never will ⁷. So a site selling evasion says the same thing about a score's limits that this clause does.

Then the consequence: a suspected case opens with a conversation and a request for the drafting record, not with a sanction issued off a percentage. An older template leaves that third sentence out, and it is the one that hands a marker a next step on a Tuesday afternoon that is neither ignoring the paper nor filing a charge.

On the abstracts one study tested, it flagged the honestly edited text more often than the humanized one. A 2026 University of Notre Dame study ran AI-labelled text through a commercial humanizer and found fewer than 4% of the rewrites were still flagged ⁸. Its blunter finding, on a cohort of 642 abstracts of 25 to 500 words, is that honest AI editing carries a higher sanction risk than humanizer-assisted evasion ⁸.

One study on short passages doesn't settle it, though, and detection of rewritten text is not weak everywhere. The Pangram 4 technical report, from Pangram Labs and the University of Maryland, says the model reads humanized text as AI-generated 97.67% of the time, and its per-system recall across 13 commercial humanizers runs from 92.78% to 99.70% ⁹.

The two studies measured different things. The independent one tested Pangram 3.2 and GPTZero against a single humanizer ⁸. The 97.67% is Pangram 4 measured by Pangram and its academic co-authors, not an outside test ⁹. So the gap is between an outside test and a vendor's co-authored report, and between versions, as much as between detectors.

Neither 4% nor 99.70% can become the standard in a policy. Neither number describes the paper sitting on a desk this afternoon. What a policy can set is the point at which a reading starts a conversation, and who looks at the drafting record before anyone writes to a student. It can also say what gets recorded when the answer turns out to be nothing.

Any detector a policy leans on should publish where its threshold sits. A committee drafting the evidence sentence can take the operating points Human publishes at human.olive.is as a worked example: accusation-safe flags a paper above 15% machine content, standard above 6% (the default), sensitive above 2%. Human measured those points in-house, on its own held-out sets, not under an independent audit. Take a reading of 8% machine content. It flags at the standard point and stays unflagged at the accusation-safe point, which is the kind of line a policy can write down. This is an estimate from our detector. Treat a flag as a reason to look closer, not as a finding. The percentage describes the text of the paper and the point it crossed, never the student who handed it in.

Four sentences change, and not one of them names a product. The disclosure clause gains the words rewrites or paraphrases existing text. In the plagiarism definition, rewritten goes in beside quoted and paraphrased. The evidence paragraph gains one line: a detector reading is not a finding. And the declaration students sign gains a question about which tools touched the draft, and at which stage.

That last edit does more work than the other three. A student who has already written down that a tool rewrote two paragraphs is concealing nothing. Where the course allows the tool, the case leaves the integrity process and goes back into marking, where the question is whether the paragraphs are any good. Disclosure is cheaper than adjudication for everyone who has sat through one.

The clause wording itself is set out in a school AI policy template with a detector clause, and the line a student is asked to sign is in an AI use disclosure statement. A mid-year revision has to reach students somewhere they will read it. After that, review both once a year, with the plagiarism definition open beside the declaration form, and check whether either has picked up a brand name since the last revision.