HumanizersFor students
Flagged After a Humanizer Pass: Bring the Draft History, Not Another Tool
Whether a humanizer pass moves a detector's reading depends on which detector read it. Pangram's 2026 report puts humanized text at 97.67 percent still detected as AI [1], while a Notre Dame team found fewer than 4 percent of short-abstract rewrites still flagged by Pangram 3.2 and GPTZero [2]. StealthGPT's homepage reports a 98 percent human average on Turnitin over 30 trials, its own measurement [3]. None of them ran on this paper. Turnitin states that it does not determine misconduct [4], so a case is settled by the draft history and a conversation.
Olive Staff & OliveUpdated
Why Did the Rewrite Not Move the Score?
Because the detectors' own published results say a rewrite is still caught: Pangram's report puts humanized text at 97.67 percent detected as AI 1. The number the humanizer advertises was produced by the humanizer. StealthGPT's homepage prints a results table naming Turnitin at a 98 percent human average across 30 trials, with no independent verification 3. The school ran neither test.
Each set of numbers is a real measurement, and neither one is an audit. The Pangram 4 technical report was written by Pangram Labs with the University of Maryland, and it counts a Mixed verdict as a catch too: 98.83 percent of humanized text reads as Mixed or AI in that evaluation 1.
A rewrite that slipped past one detector last term tells nobody what this one read.
The counter-evidence runs the other way. A 2026 University of Notre Dame study ran 642 AI-labeled abstracts through Undetectable AI and found fewer than 4 percent still flagged afterwards 2, and those were abstracts of 25 to 500 words 2, well short of a full paper; the shortest of them give a detector little to read. The sentence-level mechanics are laid out in what a humanizer does to a paragraph.
What Does the Flag Decide, and What Does It Not?
By the vendor's own account, nothing. Turnitin's post on false positives says the company makes no determination of misconduct, and that the call rests with the educator and institutional policy 4. The percentage is one input to a meeting held at the school.
The error rates are published and small, which is exactly why they matter in a single case. In a 2023 post Turnitin put its sentence-level false positive rate at around 4 percent, meaning one highlighted sentence in 25 may be human-written 5, and reported under 1 percent at document level for papers marked at 20 percent AI writing or more 5. Applying that rate to a campus's human-written papers is an assumption, and even on that assumption a low rate still means some papers wrongly flagged.
| The report shows | What it settles |
|---|---|
| A percentage of AI writing | Nothing on its own 4 |
| Highlighted sentences | Around 4 percent may be human-written 5 |
| A flag after a rewrite | Nothing about the rewrite |
Somebody at the school still has to read the paper and ask about it.
The rewrite question and the integrity question come apart here. A paper can read as machine-written and have been typed by the student who submitted it. Another can read as human and have been generated wholesale. The report is evidence about text, the case concerns a person, and only one of those two things can be read off a percentage.
Open the Draft History Before the Meeting Is Scheduled
Open it early, because the version history is the one artifact that existed before the flag did. Google Docs keeps the record under Version history. Google states that browsing earlier versions requires permission to edit the file, so a document shared as view-only has not shared its history 6.
Deadline pressure makes two things easy to get wrong. Google warns that deleting version history is permanent and cannot be undone, so nobody tidying a document before a hearing touches it 6. The record also proves less than it looks like it proves. A 2026 framework in Frontiers in Artificial Intelligence reads process documentation as evidence of engagement rather than proof of independent authorship, since process records can themselves be fabricated 7.
What to have ready before the conversation, offered as a record rather than as a verdict:
- An editable copy or a downloaded export, since a view-only share carries no history 6
- The assignment prompt, plus the submitted file and its timestamp
- Notes, outlines or a source list made before the draft existed
- A plain account of every tool that touched the file, spelling and grammar checkers included
A file with no history at all is a different conversation, and what a Google Docs record can and cannot show works through that case.
What Should the First Message to the Instructor Say?
Keep it short and dated. Name the assignment and state which tools touched the draft, at which stage, then attach the version history. Ask for a meeting instead of arguing the percentage by email. The person reading that message still holds the decision, by the vendor's own description 4.
Honesty over the humanizer is the hard part, and it decides how the rest of the meeting goes. An account that leaves out the rewrite, and is then contradicted by a browser history or a subscription receipt, costs a student the one thing still worth having, which is credibility. What to say to a professor after an AI accusation sets out the wording line by line.
Honesty about the rewrite is not a confession of misconduct, and a first message concedes nothing that has not been established.
A message sent the day the report appears is a different document from one sent three weeks later, after an inquiry has opened. Both end up in the same file. Running the paper through a second tool inside that window is the move that turns an argument over a score into an argument over conduct.
Is Another Humanizer a Way Out?
No, and not because the next tool fails. The University of Southern California's integrity office states, as its policy page read in September 2026, that AI-produced material presented as a student's own work counts as plagiarism whether copied verbatim, near-verbatim or paraphrased 8. A rewrite changes wording. It does not change what a submission claims about authorship.
At a school with that policy line, the Notre Dame result becomes a finding on the institution. The same study that measured near-total evasion on short abstracts also reported that honest AI editing carried a higher sanction risk than humanizer-assisted evasion 2. Read as advice, that is a recommendation to hide. Read as a measurement, it describes a system rewarding the wrong behavior, and the repair belongs to the people who write the policy.
Price is the axis a second purchase would be betting on, and it has been measured from the detector side. Human, the academic-writing detector at human.olive.is, finds that whether a humanized document is caught depends on the humanizer's price tier more than on which humanizer it is. That is Human's own in-house measurement on its own held-out sets, carrying no figure and no independent audit. This is an estimate from our detector. A flag should be reason to look closer and not verified proof of cheating.
The rewrite is the part to disclose, not the part to double down on.
What is left is unglamorous work that no vendor sells. Write the disclosure. Attach the process record and ask for the meeting. An AI use statement for an assignment sets out the format and the wording, and one sentence naming the tool and the stage it touched may be the whole requirement. None of that depends on which detector the school licenses next year.
Common questions
Does Turnitin detect humanized text?
The published figures come from other vendors and from research groups, and they disagree. The Pangram 4 technical report puts humanized text at 97.67 percent detected as AI, and at 98.83 percent when a Mixed verdict is counted as a catch 1. A 2026 Notre Dame study found close to the opposite on short abstracts, with fewer than 4 percent of rewrites still flagged 2. Both are real measurements of different detectors on different material. Neither one predicts what a particular school's tool returned on a particular paper.
Can a flag be removed by rewriting the paper again?
No. The report already exists and is attached to the submission that produced it. A later rewrite creates a second document with its own history and leaves the first one untouched. It also changes the question a reader is asking, from what the detector read to what the student did after seeing the result. What a student can change is the record of how the first draft was made, plus an early conversation with the instructor.
Does using a humanizer count as cheating on its own?
The institution's own policy settles it, together with what the humanizer was run over. The University of Southern California's integrity office states that AI-produced material presented as a student's own work is plagiarism whether it was copied, near-copied or paraphrased 8, which covers machine-written text put through a humanizer. Running a humanizer over writing a student produced themselves is a different question, and whether the policy covers it is worth checking in the policy's own wording rather than assuming either way.
What if the paper was written without AI and the humanizer was only used to be safe?
Say so early and show the record. Turnitin's own position is that the report does not determine misconduct and that the decision rests with the educator 4, so the meeting is where this gets settled. A Google Docs version history carries timestamps that predate the flag, and Google notes that earlier versions are visible only to people who can edit the file 6. A 2026 framework warns that process records show engagement rather than proof of authorship 7.
Why do two detectors disagree about the same paper?
Because they are different models, trained on different material and set at different thresholds. The published figures show the gap: Pangram's own 2026 report puts humanized text at 97.67 percent detected as AI 1, while a Notre Dame team testing Pangram 3.2 and GPTZero on short abstracts found fewer than 4 percent of rewrites still flagged 2. Treat any one score as a reading rather than as a fact about the paper, whichever product produced it.
What should a student bring to an academic integrity meeting?
The draft history in a form the instructor can browse, since Google states that earlier versions are visible only to people who can edit the file 6. The assignment prompt and the submitted file, with its timestamp. Any outline, notes or source list made before the draft. A short written account of which tools touched the paper and at which stage. A competing score from a detector nobody at the school licenses does not answer the report the committee is actually holding.
References
- 1.Pangram 4 Technical Report Pangram Labs and University of Maryland (arXiv), 2026. arxiv.orgHumanized text labeled AI-generated 97.67 percent of the time, and Mixed or AI-generated 98.83 percent of the time.
- 2.Why AI Detection Fails for Academic Integrity Karr, Khvatskii, Hua and Chawla, University of Notre Dame (arXiv), 2026. arxiv.orgFewer than 4 percent of rewrites still flagged after Undetectable AI, measured on 642 abstracts of 25 to 500 words, and the finding that honest AI editing carried a higher sanction risk than humanizer-assisted evasion.
- 3.StealthGPT StealthGPT, 2026. stealthgpt.aiThe vendor's own results table, naming Turnitin at a 98 percent human average over 30 trials, published with no independent verification.
- 4.Understanding false positives within our AI writing detection capabilities Turnitin, 2023. turnitin.comTurnitin states that it makes no determination of misconduct and that the decision rests with the educator and institutional policy.
- 5.Understanding the false positive rate for sentences of our AI writing detection capability Turnitin, 2023. turnitin.comSentence-level false positive rate of around 4 percent, and a document-level rate under 1 percent for documents marked at 20 percent AI writing or more.
- 6.Find What's Changed in a File Google Docs Editors Help, 2026. support.google.comBrowsing earlier versions requires edit permission on the file, and deleting version history is permanent and cannot be undone.
- 7.Designing AI-Resilient Assessment in Higher Education: A Four-Pillar Conceptual Framework Frontiers in Artificial Intelligence, 2026. frontiersin.orgProcess documentation should be read as evidence of engagement rather than proof of independent authorship, since process records are not immune to fabrication.
- 8.Academic Integrity & Generative AI University of Southern California, Office of Academic Integrity, 2026. academicintegrity.usc.eduAI-produced material presented as a student's own work counts as plagiarism whether copied verbatim, near-verbatim or paraphrased, as the policy page read in September 2026.
8 sources, numbered by first appearance.
General guidance for teachers, administrators and students. What holds at one institution, on one assignment, may not transfer to another.
Human checks whether a piece of writing can be verified as written by a person, and shows the sentences behind the result. It does not report a probability that a person used AI, it does not check for plagiarism, and no number it produces stands for a student's honesty. A flag should be reason to look closer and not verified proof of cheating.