How to Prove You Didn’t Use AI (When You Really Didn’t)

Do not edit the flagged document. Not a sentence, not a comma, and not through a humanizer, a rewriter, or any paraphrasing tool. If the words are genuinely yours, the file exactly as it stands is your single most valuable document, and every change made after the accusation becomes something you will have to account for later. Reworking authentic writing in pursuit of a friendlier reading turns a defensible position into a suspicious one.

Now the part you should weigh before trusting any of the rest. My background is in computational linguistics and natural-language processing, and I also build the commercial AI-writing tool this site sells. My business gets easier every time a reader concludes that detectors cannot be relied on, and that is a conclusion parts of this page argue for. So do not take the argument on my authority. Every load-bearing claim below links to a source you can open yourself, and the two studies that carry the most weight are peer-reviewed and freely readable.

The instruction in the first paragraph is the one a company in my position has an obvious incentive to skip, which is exactly why it comes before anything else.

What happens next is procedural, and the process is slower and more negotiable than it feels in the first hour. At most institutions an AI flag is not a finding. A flag opens a conversation with the instructor, and only some of those conversations become a formal integrity referral.

It also helps to know that false flags at institutional scale are scheduled events rather than freak ones, because one university has published the arithmetic. Vanderbilt put the sum in its August 2023 notice explaining why it had disabled Turnitin’s AI detector: hold the error rate at a single percent, apply it to a year of student submissions, and the wrongly-flagged count lands near 750. Shrink the population to the few hundred essays one department reads in a year, and that same one percent still delivers wrongly-flagged papers annually rather than never. So the reassurance here is not that everything will be fine. It is that a flag on genuine work is a predictable output of an instrument with a nonzero error rate, and the sum saying so was published years before anybody opened your file.

Ordinary preserved evidence is what closes cases like this. William Quarterman, a UC Davis senior whose take-home history midterm GPTZero flagged, was given a failing grade and referred to the campus office handling student judicial affairs. His Google Docs edit history cleared him roughly a month later, the investigator writing that “I believe you most likely wrote the text you submitted” (USA Today, April 2023). Nothing in his file had been assembled after the accusation.

One instinct to resist immediately: arguing that the percentage was low. Treating the number as something to negotiate concedes that the number is valid evidence in the first place, and that concession is the weakest ground available to somebody who wrote the work. If you genuinely did not use AI, the stronger position is the one developed over the next three sections, that the reading should not be carrying evidentiary weight at any value.

Genuinely human writing gets flagged because detection tools do not read authorship. They score surface statistics of the prose, mainly how predictable each word is given the words before it and how much sentence rhythm varies across a passage. Machine text tends to sit low and flat on both measures. So does careful, formal, heavily-revised human academic writing, which is the uncomfortable overlap at the centre of the whole problem. None of that is specific to one vendor: whether the reading came from Turnitin’s indicator, from GPTZero, or from the AI-detection feature inside Grammarly, it is a statistical judgement about finished text, and it runs into the same overlap. A separate question is whether Grammarly’s editing tools themselves can put your writing on the wrong side of that judgement; that has its own mechanics and its own vendor documentation, covered in does Grammarly get flagged as AI.

The best-documented consequence of that overlap falls on people writing in a second language. In 2023 Liang and colleagues at Stanford published a controlled test in Patterns (DOI 10.1016/j.patter.2023.100779, with the preprint openly readable). They ran authentic student writing, produced with no machine involvement whatsoever, through seven publicly available detectors. On 91 TOEFL essays by non-native English speakers, those seven detectors labelled 61.3% as AI-written. On 88 essays by United States eighth-graders, taken from the Hewlett ASAP corpus, the same seven detectors averaged a 5.19% false-positive rate. Both figures matter, and nearly every page covering this topic quotes only the first. The gap between them is what makes the finding usable: on writing no model touched, second-language students absorbed close to twelve times the false-flag rate of the native-speaking control group.

If English is not your first language and you are tempted to leave that out of your response because it feels like an excuse, leave it in. Disclosing it is not a plea for leniency, and it concedes nothing about the quality of your writing. It points at a documented property of the instrument that produced the flag. Concealing the one fact with a peer-reviewed number attached to it discards the best-evidenced argument available to you.

The reason the gap exists is worth one sentence, because a panel will ask. As a writer’s command of English strengthens, error density drops and the prose grows tighter and more uniform, which is the same low-variance texture a classifier was trained to read as synthetic; I published a measurement of that convergence in what actually changes as English proficiency grows. Skilled second-language writing is pushed toward the profile other detectors penalise. That is a property of the detector, not a description of the writer.

The errors also appear to land unevenly along lines beyond language. Common Sense Media’s September 2024 report The Dawn of the AI Era, drawn from Ipsos fieldwork that March and April, found 20% of Black teenagers saying a teacher had wrongly flagged their schoolwork as AI-generated, against 7% of white teenagers and 10% of Latino teenagers. Read the base before you cite it: that question reached 771 teenagers rather than the full sample, and the answers are what teenagers reported about themselves rather than verified accusations. Cite it as exactly that, and it still belongs in any honest account of who ends up having to defend themselves.

There is no single false-positive rate for AI detectors, and the strongest independent evidence says the category as a whole is not dependable enough to carry a disciplinary decision. A false positive in AI-writing detection is exactly the situation this page addresses: text a person wrote, labelled by software as machine-written. If you produced the writing yourself, this section is the part of your file that does the most work, because it is the only part that does not depend on anybody believing you.

In 2023, Debora Weber-Wulff and colleagues published an evaluation of fourteen detection tools in the International Journal for Educational Integrity (DOI 10.1007/s40979-023-00146-z). Across all fourteen, the authors could not identify one that was simultaneously accurate and resistant to modest editing of the input. When an independent team tests an entire product category and finds no dependable member of it, the correct reading is a category-level warning rather than a shopping problem you can solve by switching vendors.

OpenAI reached a similar conclusion about its own field. The company withdrew its AI Text Classifier in July 2023 on accuracy grounds, and its guidance for educators states plainly that tools of this kind do not reliably distinguish machine-written from human-written work. The organisation with the most to gain from a working detector is on record saying the working detector does not exist.

Institutions have acted on this. Vanderbilt University published its reasoning in August 2023 for switching off Turnitin’s AI-writing detector campus-wide, and did the arithmetic out loud: at Vanderbilt’s submission volume, an error rate of merely one percent still works out to something like 750 papers a year wrongly identified. Curtin University in Australia reached the same decision from the other side of the world, switching off Turnitin’s AI writing detection across all its campuses and study periods from 1 January 2026, and naming false-positive risk to its multilingual and neurodivergent students as the reason. Be precise if you quote Curtin: only the AI-writing feature went, and its text-matching for plagiarism stayed switched on. Those two decisions are directly verifiable and both belong in a written appeal. A tracker maintained by GradPilot , last updated in July 2026, lists 24 institutions that have stepped back, and the entries on it are not all the same policy: Yale is recorded as having disabled the tool, Johns Hopkins as treating such tools as advisory and not endorsed, and UCLA as having deactivated it or never switched it on. That roster and those characterisations are the tracker’s own, not something verified here, and anyone putting a school into a formal document should check that school’s current guidance first.

A reader who has just been accused has every reason to discount a commercially-interested source, so here are measured numbers rather than a promise. We run a free detector at our detection page, and on human writing it holds up: across 15,542 human-written passages it stayed clean on 99.8% (a 0.2% false-positive rate) and it flagged none of the non-native TOEFL essays or the native-English student essays most at risk of a wrongful flag. On the writers this page exists for, its error rate is zero. A detector score is a signal, not proof (nobody should be accused on a percentage alone), but a measured 0.2% false-positive rate is exactly what makes a clean read on your own writing worth trusting.

Turnitin’s own product behaviour is the most useful item in your file if Turnitin produced the flag and the work is genuinely yours, and most students it flags are never told about it. Since 8 July 2024, Turnitin’s published guidance on AI writing detection has withheld every reading in the 1% to 19% range: the report shows an asterisk where the figure would sit, attributes no percentage, and highlights no sentences. The stated reason is that the company’s own testing found a higher incidence of false positives across that range. A genuine 0% still displays as 0%, so this is not a rounding habit. It is the band immediately beneath Turnitin’s 20% threshold going unreported, and that band is where a great many flagged students sit. Read plainly, the whole of it is the vendor declining to make an AI claim about the work at all. Reports generated before July 2024 can still carry a numeral, so check the date on yours.

Two further lines in that same guidance belong in a written response. Turnitin scopes its headline accuracy claim to the top of the scale, describing it as “keeping our false positive rate ... under 1% for documents with over 20% of AI writing,” which leaves the withheld band outside the number the company advertises. And the guidance sets a limit on how the output may be used: the reading “should not be used as the sole basis for adverse actions against a student.” Neither line asks you to dispute a percentage. Both come from the company that produced it.

The move this enables is not “my number was low, therefore I am innocent.” Arguing the size of the number accepts the number as evidence, and once it is accepted, a slightly higher reading next term convicts you. The move is to ask what the reading is being treated as. If the tool’s own manufacturer will not display a figure in the band beneath its threshold, and independent testing of the whole category found nothing dependable in it, then the reasonable question for an instructor is what evidentiary weight a percentage is meant to carry here, and what your institution’s written policy actually states the threshold to be. In a great many cases there is no written threshold at all, only a number somebody read as guilt. Our breakdown of what Turnitin’s AI indicator actually measures covers the mechanics if you need them for a written response, and there is a parallel accuracy breakdown of GPTZero’s published performance if that was the tool involved.

Ask for the specific flagged sentences in writing. A percentage cannot be answered. Sentences can, especially when you can point to the moment each one entered your draft.

Evidence of authorship is evidence about process, and if you genuinely wrote the piece, the process left a trail without you doing anything deliberate. Everything in this section is about preserving and exporting a record that already exists. None of it involves creating a record, reconstructing one, or adjusting a timestamp. Fabricating evidence is a far more serious offence than the one you are answering, it is trivially detectable in document metadata, and it turns a survivable misunderstanding into an unsurvivable one. If the trail is real, preserve it. If a particular trail is not there, say so plainly and rely on the rest.

It is worth separating evidence that holds up from evidence that only feels persuasive. A strongly-worded denial, a character reference, and a second detector returning a friendlier score all feel like proof and none of them are: the first two say nothing about authorship, and the third invites the panel to keep treating detector output as the deciding input. What holds up is dated, external and boring.

An appeal is a sequence, and starting at the top of it usually backfires. Most institutions run three tiers: an informal conversation with the instructor, then a departmental or integrity-office review, then a formal hearing with a panel and a written record. The overwhelming majority end at the first tier, and if you did not use AI they end there faster when you arrive with documents instead of denials.

Can you be expelled for AI you did not use? The ceiling exists, but first-flag cases seldom reach it. Where a single flag is upheld, the common outcomes are an informal resolution, a rewrite or resubmission, or a penalty confined to that one piece of work. Suspension and expulsion generally require a repeat finding, or a case carrying something more than a detector percentage. Sanction bands are written locally, though, so the range that can actually be applied to you sits in your own institution’s integrity policy rather than in any article. Read it early, and read it before you agree to anything.

The letter carrying those attachments should be shorter than you will want it to be. One page, two at the outside, covering what you were told, what you are attaching, what each attachment shows, and what you are asking for. Keep the register flat and procedural, and leave out how the accusation felt, however earned that paragraph would be, because a document written as a record survives being forwarded to people who were not in the room. Close on one specific request: that the finding be withdrawn, or that the assignment be regraded on its merits. A named request gives the reader something they can decide on. A general appeal to fairness gives them nothing to do.

For a formal hearing, bring the exported version history, the browser-history export, your prior papers, printed copies of the two studies, and a one-page summary stating what each item shows. Ask for the specific policy language being applied to you, and ask that a person rather than a score be identified as the basis of the finding. Bring somebody with you if your institution permits a support person, because sitting in that room alone is harder than it sounds.

Involve an education-law attorney when the possible outcome is suspension, expulsion, a transcript notation or a visa consequence, and when the process itself looks irregular. Ask a prospective attorney three things: whether they have handled academic-integrity matters at your type of institution, whether they can attend the hearing or only advise beforehand, and what procedural options realistically remain if the panel finds against you. Before any of that costs money, though, there is a body worth reading. The International Center for Academic Integrity, an association whose members are institutions and the integrity staff inside them, publishes its standards and its fundamental-values framework openly at academicintegrity.org. Those documents are written for the people administering your case rather than for you, which is what makes them quotable: an appeal citing the standards the field sets for itself is asking a panel for consistency, and consistency is an easier thing to grant than mercy.

If you genuinely did not use AI, none of this is about persuasion. It is about replacing a statistical estimate with a documented record of how the work came to exist, and asking that the record be weighed by a human being. That is a reasonable request, a growing number of institutions have already decided it is the right one, and the evidence you need for it is almost certainly still sitting in your drafts folder.