Hub · AI detection & flagxiety
How do you prevent a false AI flag — and what if you already have one?
A detector score is a statistical guess, not a finding of fact — and the error rate lands hardest on non-native English writers. You cannot stop a detector from producing that guess. You can make sure that when it does, a record of how you actually wrote the document already exists.
The protection that works is not a better score. Detectors disagree with each other, scores change between versions, and «passing» one tool does not mean passing the one your institution uses. What holds up is process evidence that accumulates while you write — version history at minimum, or a cryptographically signed record of the writing process. An accusation against a documented process usually ends the conversation, not starts it.
How do you prevent a false AI flag?
Short answer: keep evidence, not a score. Write where version history is kept, keep your outlines and earlier drafts, and for high-stakes work keep a signed, append-only record of the writing process — proof that exists before anyone asks for it.
Prevention is not a defence assembled afterwards; it is evidence that accumulates while you write. Version history is the floor — Google Docs, Word autosave, or any editor that keeps earlier drafts, outlines and notes, because the goal is showing the document grew over time. For work where the stakes are real, the Diglot Authorship Certificate goes further: an append-only, cryptographically signed record of your writing process, verifiable by anyone you share it with. For the short version of what to keep and from when, work through the authorship checklist.
Running your draft through a detector before you submit is not prevention. It can lower anxiety, but detectors disagree with each other, scores change between versions, and «passing» one tool does not mean passing the one your institution uses. Committees find a documented writing process far more persuasive than any counter-score from another detector.
What do you do if you have already been flagged?
Short answer: do not reply defensively in the first hour. Ask for the full detection report and your institution's threshold policy, gather your version history and drafts, then appeal in writing under the formal academic-integrity procedure.
| Step | What to do | Why it matters |
|---|---|---|
| Do not send a defensive reply yet | Nothing you write in the first hour helps. Acknowledge receipt if required, and ask for time to prepare a response. | Panic emails become part of the record. |
| Request the full detection report | Not the score — the report. Which passages were flagged, which tool, which version, what threshold. | A score without a report is not evidence anyone can examine. |
| Ask about the threshold and policy | What score triggers review? Is a detector score alone sufficient for a finding under your institution's policy? | At most institutions it is not — their own rules require more. |
| Gather your process evidence | Version history (Google Docs, Word autosave), earlier drafts, outlines, notes, search history, messages discussing the work. | The goal is showing the document grew over time. |
| Submit a written appeal and ask for human review | Present the evidence calmly, cite the false-positive research on this page, and request review under the formal academic-integrity procedure. | Not an informal chat. |
The long version with templates: how to appeal a false AI accusation.
Why do detectors misfire on non-native English writers?
Short answer: detectors score statistical uniformity — how predictable the word choices are and how much sentence rhythm varies. Careful formal writing has low variety by design, and English written across two languages carries translationese, so both land where machine output lands.
Detectors estimate whether text is machine-like by measuring statistical uniformity: how predictable the word choices are (perplexity) and how much sentence rhythm varies (burstiness) — the wider practice of inferring authorship from such measurements is stylometry. Careful formal writing has low variety by design — and writing produced between two languages carries translationese, which looks statistically similar to machine output. Nobody is «detected»; a resemblance is scored. The constant low-grade fear this creates has a name: flagxiety, and it is expensive well before anyone is actually accused.
What research can you cite in an appeal?
Short answer: six findings, each with a named source and a year. Liang et al. («GPT detectors are biased against non-native English writers», Patterns, 2023) found detectors falsely flagged 61% of non-native TOEFL essays; Weber-Wulff et al. (2023) tested 14 tools and called them «neither accurate nor reliable»; OpenAI retired its own classifier in July 2023; Vanderbilt University disabled Turnitin's AI detection in August 2023.
Every entry in the table is a verifiable, named source. The compendium is kept current as new research and rulings land — last reviewed 23 August 2026.
| Finding | Source | Year |
|---|---|---|
| AI detectors falsely flagged 61% of TOEFL essays written by non-native English speakers. | Liang et al., «GPT detectors are biased against non-native English writers», Patterns (Cell Press) | 2023 |
| Across 14 detection tools tested on human, machine, and lightly edited text, researchers concluded the tools are «neither accurate nor reliable», with accuracy dropping sharply on paraphrased content. | Weber-Wulff et al., «Testing of detection tools for AI-generated text», International Journal for Educational Integrity | 2023 |
| OpenAI retired its own AI-text classifier six months after launch, citing its «low rate of accuracy». | OpenAI, AI classifier sunset notice | July 2023 |
| Vanderbilt University disabled Turnitin's AI-detection feature, writing that the tool's false positives and opacity made it unsuitable for misconduct decisions. | Vanderbilt University, Brightspace announcement | August 2023 |
| Synonym-spinner paraphrasing produces detectable «tortured phrases» («counterfeit consciousness» for «artificial intelligence») that have led to journal retractions — a separate failure mode from AI detection, often conflated with it. | Cabanac, Labbé & Magazinov, «Tortured phrases» research | 2021 |
| Litigation over AI-detection accusations has reached US courts, beginning with a Massachusetts case over a high-school AI-cheating finding. | RNH v. Hingham Public Schools | Filed 2024 |
The current cycle of AI-detection litigation is tracked in AI detection lawsuits 2026: what ESL writers need to know.
Which detector or situation are you dealing with?
Short answer: detector-specific guides for Turnitin, Copyleaks, GPTZero and Originality, plus the situations that produce flags on their own — patchwriting, neurodivergent writing styles, and the appeal itself.
- Does Turnitin detect AI? What the score means
- Turnitin AI detection — how accurate is it?
- Turnitin similarity score — what's actually safe?
- Copyleaks flagged you — what to do
- GPTZero vs Turnitin vs Originality — accuracy compared
- How common are false positives?
- How to appeal, step by step
- Accused of using AI on work you wrote
- The lawsuits: what ESL writers should know
- Detectors and neurodivergent writers
- Patchwriting — the other reason ESL writers get flagged
Flagged as AI — questions
How do I prevent a false AI flag before anyone accuses me?
Write where history is kept (Google Docs, Word with autosave, or Diglot), keep outlines and notes, and for high-stakes work use the Authorship Certificate — an append-only, cryptographically signed record of how the document was written that exists before anyone asks.
Can a detector score alone prove I used AI?
No. An AI-detection score is a statistical estimate that your text resembles machine-generated patterns — it identifies no source, no tool, and no act. Most institutional policies require more than a score to sustain a misconduct finding, and vendors themselves caution against using scores as sole evidence.
Why was my writing flagged when I wrote every word myself?
Detectors flag statistical uniformity: predictable word choices and even sentence rhythms. Careful, formal writing — especially by non-native speakers who learned English through study — has exactly those properties. Liang et al., writing in Patterns (Cell Press) in 2023, found that AI detectors falsely flagged 61% of TOEFL essays written by non-native English speakers. Your writing being flagged says more about the method than about you.
Should I run my essay through detectors before submitting?
It can lower anxiety, but treat results loosely: detectors disagree with each other, scores change between versions, and «passing» one tool does not mean passing the one your institution uses. Better long-term protection is process evidence — drafts, version history, or a cryptographic record like the Authorship Certificate.
What evidence actually wins appeals?
Process evidence. Version history showing the document growing over hours and days, earlier drafts with your own corrections, notes and outlines, search history. Committees find a documented writing process far more persuasive than any counter-score from another detector.
Diglot is a bilingual writing editor with authorship proof built in — start for free, no credit card required.
