Why honest writing gets flagged
Detectors measure statistical smoothness: predictable word choices, uniform sentence rhythm, even paragraph structure. Machines write that way by default. The problem is that some humans do too, and some genres demand it. The detector is not reading minds; it is reading texture, and human texture overlaps machine texture at the smooth end. Every false positive story starts in that overlap.
Who carries the risk
Non-native English writers top every study. Writing in a second language pushes toward safer constructions and higher-frequency vocabulary, which is precisely the machine profile. Liang et al. measured false positive rates above 50% on TOEFL essays across commercial detectors, against near-zero on native-speaker essays from the same prompt pool. Formal and technical writers come next: legal prose, lab reports, documentation and academic register are formulaic by requirement. Heavy grammar-tool users blur the category honestly, since machine rewriting of human drafts is machine writing in the statistical sense. And short samples put everyone at risk, because below a couple hundred words every detector is guessing.
The playbook: contesting a false flag
Step one, in writing, ask for specifics. Which tool, which version, what score, and the report itself. Vague accusations dissolve under specific questions, and you are generally entitled to the answers.
Step two, assemble process evidence. Google Docs or Word version history, drafts, outlines, notes, search history if you are comfortable sharing it, and earlier writing of yours in the same voice. Incremental version history is the single most persuasive artifact in these processes.
Step three, cite the science. The Liang study on non-native bias if it applies to you. Turnitin’s own guidance that scores are not sole evidence. Vanderbilt’s public decision to disable detection over unverifiable false positive rates. Asking what false positive rate did the institution validate before relying on this tool is legitimate and frequently unanswerable.
Step four, escalate calmly. Integrity processes have appeal layers, and documented process beats a percentage at nearly every one of them. Keep everything in writing, keep your tone administrative, and remember that the burden of proof sits with the accusation.
The playbook above is what a wrongly accused person will correctly do to you. Build your process so it survives: disclosed tools, inconclusive bands honored, process evidence requested before any action, scores never standing alone. Our detector states its uncertainty on the dial precisely so it cannot be quietly converted into a verdict machine.
A template for responding to an accusation
When the email arrives, calm specificity beats outrage. A structure that works, adapted to your own voice:
Paragraph one, state your position without hedging: I wrote this work myself, and I can document the process. Paragraph two, request the specifics in writing: which tool produced the score, which version, what number, and a copy of the report, plus the institution's documented false positive rate for that tool. Paragraph three, present your evidence inventory: version history with timestamps, drafts and outlines, notes, and prior writing in the same voice, offered for comparison. Paragraph four, cite the ground rules: the vendor's own guidance that scores are not sole evidence of misconduct, your institution's integrity procedure, and, if it applies to you, the published research on elevated false positive rates for non-native English writers. Close by asking for the next step in the formal process and keep every exchange in writing.
Two cautions. Do not rewrite your essay to score lower and resubmit, which reads as consciousness of guilt and destroys your best evidence. And do not lean on a clean score from a different detector as your defense, because the same uncertainty that makes the accusation weak makes your counter-score weak. Process evidence wins these cases; competing percentages prolong them.
The arithmetic of being wrongly flagged
One more lens that clarifies the stakes: base rates. Imagine a course of two hundred essays where ten students actually used AI, and a detector with a generous 95% accuracy in both directions. It catches roughly nine or ten of the guilty, and it also flags about ten of the hundred ninety honest essays. The accusation pool is now half innocent, which means a flagged student, before any other evidence, carries roughly even odds of having done nothing wrong. That arithmetic is why score-only enforcement fails mathematically, not just ethically, and why every credible policy demands corroborating evidence. It is also why we print the inconclusive band on the dial: the honest representation of a probability includes the part where it tells you nothing.
Why our own tool draws the gray zone
Scores between 40 and 69 here are labeled inconclusive, and the verdict copy under every reading says what the score cannot prove. That costs us the appearance of decisiveness that other tools sell. It buys the thing this article is about: a tool that is structurally harder to use against an innocent writer. The mechanics behind the bands are in how AI detectors work, and the accuracy record of the most consequential institutional tool is in the Turnitin accuracy guide.