Free tools / Accusation Defense Kit
Flagged by an AI detector? Build your response.
A detector said your writing was AI. You wrote it. This page gives you the three things that actually move an academic-integrity conversation: a letter you can fill in and send, a checklist of the evidence to gather before you send it, and a citation pack with real sources you can attach. No account, no email address, nothing to pay. Everything below runs in your browser — nothing you type is sent to us.
Do these four things first — before you reply to anything
- Stop editing the file. Its revision history is your best evidence, and every change you make after the flag weakens it. Do not rewrite the flagged passages hoping the score drops.
- Export what you already have — version history, drafts, notes, sources, research browser history. Do it today; some platforms trim old revisions.
- Ask, in writing, what the accusation is based on. Which tool, what score, which passages, and what the written policy says about how a score may be used. The letter below asks all four.
- Read your institution’s academic-integrity procedure before any meeting, so you know whether this is informal or the first step of a formal process, and whether you may bring an advisor.
And the one thing not to do: do not confess to end the discomfort. An admission closes the case permanently. The discomfort of an unresolved accusation does not.
The evidence checklist
Tick what you can produce. Everything you tick is written into the letter below automatically, in the letter’s own sentence. Nothing here is stored — this is a working surface, so copy or download before you close the tab.
Process evidence — the strongest kind
Understanding evidence — the part almost nobody offers
Procedural ground — get these before the meeting
The appeal-letter builder
Fill in what you know and the letter rewrites itself as you type. Blanks stay as visible
[placeholders] so you can finish it later.
The tone is deliberately calm and procedural: it asserts what you did, offers evidence, asks four questions
in writing, and cites the research — without accusing anyone of anything.
Your letter
Read it before you send it. Cut anything that isn’t true of your situation, and add the one or two specifics only you know — the section you rewrote three times, the source you argued with. A letter that sounds like a template gets read like a template.
The citation pack
Attach the ones that fit your case. Every entry links to a primary or reputable source, dated, so anyone reading your letter can check it. Cite figures precisely — an overstated number is the fastest way to lose the room, and the accurate numbers are damaging enough.
1. The vendor’s own numbers — Turnitin
Turnitin’s chief product officer, Annie Chechitelli, has publicly stated a document-level false-positive rate of under 1% for documents flagged at 20% AI or above, and a sentence-level false-positive rate of about 4%. Turnitin’s guidance also states that the AI indicator should not be the sole basis for a misconduct decision. If your case rests on a score alone, this is the single most useful citation you have: it comes from the company that sells the tool.
Source: Turnitin — “AI writing detection update from Turnitin’s Chief Product Officer” (2023).
2. Detectors penalise non-native English writers — Liang et al., 2023
Seven leading detectors were run over TOEFL essays written by non-native English speakers: 61.3% average false-positive rate, and 97.8% of those essays were flagged by at least one detector, against near-zero rates on comparable US-student writing. The result inside the paper that matters most: when the researchers used ChatGPT to make the same essays sound more natively fluent, the false-positive rate fell to 11.6% — the detectors are scoring fluency, not authorship. Use this if English is not your first language.
Source: Liang, Yuksekgonul, Mao, Wu & Zou, “GPT detectors are biased against non-native English writers”, Patterns 4(7), 2023 — arxiv.org/abs/2304.02819.
3. No tool tested was reliable — Weber-Wulff et al., 2023
A peer-reviewed evaluation of 14 AI-detection tools found that none exceeded 80% accuracy, that the tools were biased toward classifying text as human-written, and that accuracy fell further on machine-translated or lightly edited text. Use this when the argument is about detectors in general rather than one vendor.
Source: Weber-Wulff et al., “Testing of detection tools for AI-generated text”, International Journal for Educational Integrity 19(26), 2023 — edintegrity.biomedcentral.com.
4. A university cancelled the contract and published why — Washington State, February 2026
WSU’s Provost’s Office terminated its Turnitin AI-detection contract in February 2026. The memo to instructors gives the reason in numbers a panel will recognise: between 2023 and 2025, 33% of Review Board cases involving AI allegations ended in a finding of not responsible because the detection result had been submitted without any other supporting evidence. It also names the disparate impact on neurodivergent students and students writing in a second language. This is the strongest citation in the pack for the specific argument “a score on its own should not decide this”, because it is an institution saying so about its own cases.
Source: WSU Office of the Provost — memo to instructors, February 2026 (PDF); context via the WSU Faculty Senate.
5. Detection switched off outright — Curtin University, from 1 January 2026
Curtin disabled Turnitin’s AI writing detection across all campuses and study periods from 1 January 2026, while keeping originality checking. Useful as a current, forward-dated example that this is live institutional practice rather than a 2023 news cycle — and, if you are in Australia, that a peer institution has already made the call.
Source: Curtin University — “Update on Turnitin AI-Detection Tool”.
6. A regulator’s position — TEQSA (Australia)
Australia’s Tertiary Education Quality and Standards Agency states plainly that any score, high or low, is insufficient evidence by itself to allege misconduct. If your institution is Australian this is binding sector guidance rather than an opinion; elsewhere it is still a national regulator reaching the same conclusion as the vendor.
Source: TEQSA — Academic Integrity Toolkit, detecting AI-generated text.
7. The arithmetic of a 1% error rate — Vanderbilt
Vanderbilt disabled Turnitin’s AI detector and published its reasoning: with roughly 75,000 papers submitted annually, even the vendor’s own claimed 1% false-positive rate implies about 750 students wrongly flagged per year at one university. This is the citation for the moment someone says “but the error rate is only 1%”.
Context and the wider institutional list: our own roundup of universities that disabled AI detection. Secondary counts now put the total at 50+ institutions across the US, Canada, UK, Australia and South Africa — if you cite that number, cite it as a secondary tally, not a verified census.
8. Our own measurement — 861 human sentences through GPTZero, May 2026
We ran 861 verified human-written sentences — from pre-LLM PubMed abstracts, Wikipedia, ESL learner writing and news journalism — through GPTZero one sentence at a time. 13.8% came back classified AI or mixed; worst on professionally edited news prose (19.8%) and on writing by people learning English (16.0%). The raw data is published.
The caveat, which you must include if you quote this: that is not the number GPTZero publishes, and it is not measuring the same thing. GPTZero states a false-positive rate of no more than 1% at document level; we tested at sentence level, deliberately the harder case, because a short span gives the classifier less signal. Both can be true at once. Cite ours as what it is — a sentence-level worst case on text nobody disputes was human-written.
Sources: our 861-sentence GPTZero false-positive study (raw CSV published), the wider false-positive picture, and our quarterly re-test of whether AI detectors are actually getting better.
Questions people actually ask at this point
How can I prove I did not use AI?
With process, not prose. The strongest single artefact is the document’s own revision history — real writing shows dozens of sessions with false starts and deletions rather than one paste. After that: dated drafts, annotated sources, research history, messages about the assignment, timestamped photos of handwritten notes. The other half is demonstrating you understand the work: offering to talk through your argument or write a comparable passage under observation. Stop editing the file now — every edit after the accusation degrades the best evidence you have.
What should I do if a professor accuses me of using AI?
Freeze the file, export your evidence, ask in writing for the specific basis of the accusation, and read your institution’s procedure before the meeting so you know whether you are in an informal conversation or a formal process and whether you may bring an advisor. Do not confess to end the discomfort: an admission closes the case permanently. The letter above does the asking for you.
Can professors actually prove you used AI?
A detector score is a probabilistic classification, not proof of authorship — Turnitin says so, and TEQSA says any score alone is insufficient to allege misconduct. Washington State University found that 33% of its 2023–2025 Review Board AI cases ended in not responsible precisely because the detection result arrived without other evidence. Where proof exists it comes from other things: a submission history inconsistent with the work, an inability to discuss one’s own argument, an admission.
How often does Turnitin falsely detect AI?
Turnitin’s own chief product officer states under 1% at document level (for documents flagged at 20% AI or more) and about 4% at sentence level. Independent research finds higher, unevenly distributed rates: 61.3% average across seven detectors on TOEFL essays by non-native writers (Liang 2023); none of 14 tools above 80% accuracy (Weber-Wulff 2023). Our own sentence-level GPTZero test returned 13.8%. Nobody, including the vendors, claims zero.
Is this free, and do you see what I write?
The kit costs nothing, needs no account and asks for no email address. The builder and the checklist run entirely in your browser — there is no server call, nothing you type reaches ToHuman, and closing the tab discards it. Copy or download your letter before you leave.
Should I rewrite the flagged text so the score drops?
Not while you are under investigation. Editing the file after the flag damages your revision history, which is the best evidence you have, and it reads as consciousness of guilt if it is discovered. If you wrote it, the answer is evidence, not rewriting. We sell a rewriting tool and we are still telling you this — see below.
Who made this, and what we are not claiming
ToHuman makes a tool that rewrites AI-assisted drafts to read more naturally. We built this kit because our own research on detector false positives kept being read by people in the middle of an accusation, and what they needed was not a rewriting tool.
So, plainly: if you wrote the work yourself, do not put it through a humanizer. It changes nothing about the accusation, it damages your version history, and it is the wrong answer to the question you have been asked. Nothing on this page promises that any rewrite will change a detector’s verdict; those classifiers belong to other companies and are retrained without notice, which is the argument we make at length on our GPTZero bypass page.
If what you actually have is an AI-assisted draft you want to sound like your own writing before you submit it — a different situation from this page — the tool on our homepage is free for 2,500 humanized words a month with no card required. And if you want the longer background reading, our full checklist for anyone falsely accused of using AI covers the process end to end.
This page is general information about an academic process, not legal advice. If your institution has moved to a formal hearing with a sanction attached, ask whether your students’ union, an ombudsperson, or a student-advocacy office can attend with you — most institutions have one and most students never ask. Last reviewed 2 September 2026.