Falsely accused of using AI? What to do, step by step

A detector reading is a claim with a source, and you are entitled to see the source. Ask for the items below before you respond to the substance of the accusation.

Turnitin's FAQ states that the indicator "is not visible to students" and that an instructor can download the report as a PDF and share it (Turnitin, AI writing detection capabilities FAQs, read 9 November 2026). Ask for that PDF. The page on how Turnitin's AI detection works explains what each part of it means.

Detectors read the finished text. You have something they never see: how the text came to exist. The strongest evidence is anything created before the accusation, with timestamps you did not control after the fact.

GPTZero, writing for instructors who see a flag, recommends exactly this: "Ask students to show their process (through edit history or in-person assessment)" and "check for patterns across multiple assignments rather than relying on a single instance" (GPTZero, educators page, read 9 November 2026). The same page says that "like all AI detectors, GPTZero isn't perfect, and short texts or highly edited passages can make detection harder". When the vendor's advice to your instructor is to ask for process, bringing the process is the answer.

Present it plainly. A one-page timeline of the assignment, with the dates of each draft, each source read and each revision, does more than a long defence of the prose.

Every major vendor has written a sentence you can quote. Turnitin states that the percentage "should not be used as the sole basis for action or a definitive grading measure by instructors" and that Turnitin "does not make a determination of misconduct", read 9 November 2026. GPTZero's educators page recommends using its tool "as a conversation starter", read 9 November 2026.

No detector score should be the only evidence in a misconduct case.

The research says why. Liang and colleagues ran seven GPT detectors over 91 essays by non-native English writers and found an average false positive rate of 61.22 percent, with 89 of the 91 essays flagged by at least one detector; the same detectors passed nearly all essays by US eighth graders (Liang et al. 2023, Patterns00130-7), read 9 November 2026). Weber-Wulff and colleagues tested 14 detectors in 2023 and wrote that a student accused "only on this basis would have no possibility for a defence", because the tools return a label without evidence (Weber-Wulff et al. 2023, International Journal for Educational Integrity, read 9 November 2026).

Institutions have said the same. Harvard's Faculty of Arts and Sciences tells its instructors that "it would be inadvisable to count on automated methods for GAI detection" and that FAS does not plan to license such a tool (Harvard Office of Undergraduate Education, AI guidance, read 9 November 2026). Vanderbilt University disabled Turnitin's indicator in August 2023, calculating that a 1 percent false positive rate on 75,000 submissions a year would be about 750 wrongly flagged papers (Vanderbilt University, guidance on AI detection, 16 August 2023, read 9 November 2026). If your institution has a comparable statement, cite it.

Because detectors score predictability, not authorship. Turnitin names the human text most likely to trigger a false positive: "content without a lot of structural variation, text that literally repeats itself, or text that has been paraphrased without developing new ideas", and adds that short documents are scored "all or nothing" as a single segment, read 9 November 2026. Liang and colleagues showed the mechanism when enriching the vocabulary of the TOEFL essays cut the false positive rate from 61.22 percent to 11.77 percent: the detectors were reading vocabulary range.

If you write in a second language, write to a template, or write short plain reports, you fit the profile. That is a fact about the tool, and the guide on AI detector false positives sets out the published rates for each group. It is also an argument you can make: the detector flags writers like you at a published rate, and here is the study.

Do not rewrite the work and resubmit it before the conversation. That looks like concealment and destroys the evidence that the original was yours. Do not run the essay through a humanizer to "fix" the score; that alters work you are claiming as your own, and it produces the pattern that detectors are trained to catch. Do not rely on a different detector's lower reading as proof, since detectors disagree by design, which is the point made on whether AI detectors work.

Do keep every file as it was, write the timeline, quote the vendor statements, and ask for the hearing your policy provides. The guides hub holds the rest of this series, including what detectors measure and what Turnitin states about its own indicator.