detectai.media

Falsely accused of using AI? What to do

A calm, evidence-led guide to contesting a false AI-writing accusation: make non-use the more plausible account and put the burden back on the accuser.

By The DetectAI team
• 7 min read
Contents

The burden is on the accuser to show misconduct, not on you to prove your innocence. You may not be able to prove non-use with absolute certainty, but a genuine drafting trail and process record can make human authorship the overwhelmingly plausible account, while the detector’s output is, in the words of one independent study, “a simple claim without verifiable evidence.”

Start from what a detector score actually is

Before assembling a defence, understand what is being held against you. A post-hoc detector produces a probability, not a finding. As Weber-Wulff et al. (International Journal for Educational Integrity 2023) put it, that output is “a simple claim without verifiable evidence,” which means a student accused on it alone “would have no possibility for a defence.” Your task is not to disprove the score. It is to show that the score does not discharge the accuser’s burden, and that human authorship is at least as plausible as the alternative. Why these tools single out particular writers in the first place is its own subject, set out in our companion piece Why AI text detectors falsely accuse real writers; here the concern is what to do once it has happened.

The evidence you can present, ranked honestly

Not all corroboration is equal, and it helps to be honest about which is which.

The strongest is an organic drafting trail: Google Docs or Word version history, tracked changes, or git commits that show the document growing incrementally over time. It demonstrates process, which a detector cannot produce and which is hard to fake convincingly after the fact. Keeping that trail as you work is the single most useful habit here: the record you build now is the evidence you would produce later. Its honest limits: some people write in a single sitting or by hand and have no trail, it can in principle be simulated, and many people simply never keep it.

Next strongest are contemporaneous process artefacts: notes, outlines, annotated sources, and reference-manager entries dated around when you worked. They do not show the document forming keystroke by keystroke, but they show the work behind it.

Weakest on its own is the ability to discuss or reproduce the work, an oral defence. It is informative to a human assessor, but it is biased against people who articulate poorly under pressure, who are disproportionately the same writers detectors over-flag. Offer it as support, not as your foundation.

What to ask for when you are accused

It is reasonable to ask for the basis of the claim, and doing so is ordinary evidentiary hygiene, not obstruction. Ask for four things in writing: the detector’s name and version; the score and the threshold it was judged against; the raw report, meaning whether the flag is document-level or sentence-level and which passages drove it; and the policy basis for treating the score as evidence at all. The threshold matters because a score is uninterpretable without its operating point: RAID (Dugan et al., ACL 2024) had to fix and disclose a false-positive rate precisely because the same tool reads very differently at different thresholds.

Stand on the institutional record

You are not arguing against a settled consensus that detectors work. You are pointing to a record in which the leading vendor, multiple universities, and a national exam-board council have already conceded the score is not, by itself, proof.

  • Turnitin, the leading vendor, states that its score “does not make a determination of misconduct,” even while the company claims a false-positive rate of less than 1%.
  • Vanderbilt University disabled Turnitin’s AI detector in 2023, noting that a 1% false-positive rate across its roughly 75,000 annual papers would wrongly flag around 750 of them.
  • The University of Pittsburgh recommends against the tools and disabled Turnitin’s detector.
  • Rice University’s Honor Council “will not use detector software as the sole or primary evidence in an adjudication.”
  • MIT Sloan publishes guidance titled “AI Detectors Don’t Work.”
  • The University of North Florida cites both insufficient accuracy and a FERPA privacy risk.
  • The Joint Council for Qualifications, which oversees the UK exam boards, requires detection to form part of a holistic approach rather than a sole judgement.
  • OpenAI withdrew its own AI Text Classifier in 2023 “due to its low rate of accuracy” (a 26% true-positive and 9% false-positive rate), the maker of a leading model conceding the tool was not fit for consequential use.

Taken together, the vendors who build these tools and the institutions that deploy them already decline to treat a detector score as proof of misconduct.

Use the tool’s own measured failure rate

If you know which tool flagged you, its independently measured error rate is part of your answer. In Weber-Wulff et al.’s testing, GPT Zero produced a false-accusation rate of around 50% on human documents. In Liang et al. (Patterns 2023), Originality.ai produced a 76% false-positive rate on non-native English essays. A score from a tool with a documented double-digit, or even fifty-percent, false-positive rate on people like you is not the strong evidence it is being presented as, and naming that rate puts the operating point back into the conversation.

The honest limit

One practical caution that fits the moment you are in: now that you have been accused, do not alter, rewrite, resubmit, or delete the work or its draft history. The original file and its version trail are your strongest evidence, so changing anything now looks like tampering and weakens the one thing that actually helps you. Preserve everything as it stands, make copies, and keep the argument on the evidence and the tool’s known failure rate, not on your prose.

This is contestation, and how far it goes depends on what you kept. With a complete drafting trail, human authorship becomes the overwhelmingly plausible account, strong enough to defeat an uninterpretable score; with little or no record, you can still undercut the accusation even if you cannot fully settle it. Either way the burden sits with the accuser, not you: they must show misconduct, and a score its own makers say is not a determination of misconduct does not discharge it. The power asymmetry is real: you are often answering to an institution that controls the process, and what recourse exists varies from one to the next. What the evidence changes is not the asymmetry but the footing. A person contesting a flag with a drafting trail and the tool’s own disclosed failure rate is holding inspectable evidence, while the accusation rests on a claim its own makers say is not a determination of misconduct.

Sources

#text#false-positives#how-to
Last updated
26 June 2026
Category
Reliability