What to do when your writing is flagged
A practical sequence: work out what the score actually says, get a second measurement, and assemble the record of process that the text itself does not contain.
A detection score is a probability produced by a classifier with a measurable error rate. It is not a match against a source document, and it is not a record of how the text was made. Knowing precisely what the number is makes the rest of this easier.
Everything below is about establishing facts. What follows from those facts depends on your institution or publisher and their rules, which we are not in a position to characterise.
1. Find out what was actually measured
Different tools report different quantities, and the first question is what the number refers to.
- Which tool produced it, and what does that tool publish about its own false positive
rate
- Document or passage. A document percentage says what proportion of text was flagged.
A per-passage view says where
- How much text was scored. Short submissions, or documents heavy with tables, equations
and references, may contain far less qualifying prose than the page count suggests
- Where the threshold sits. A flag means a score crossed a line somebody chose
If the score came from Turnitin, what a Turnitin AI score means covers how to read that specific report.
2. Get an independent measurement
Detectors disagree with each other for structural reasons: different training data, different thresholds, different segmentation, different weighting of the underlying signals. We set them out in why AI detectors disagree.
A second score tells you whether the first reflects something in the text or something in that tool's calibration. Wide disagreement is itself informative, and usually means the text is genuinely borderline rather than that one tool is broken.
Ours is free and unlimited, scores each paragraph separately rather than only the document, and publishes its false positive rate: 0.8% at our public setting, 0.4% at a stricter one, measured on pre-LLM academic writing.
3. Look at which passages were flagged
Per-passage scores are more useful than a document figure, because the pattern carries information.
Clustered in one section suggests something specific about that section, and it is worth looking at what is different about it. Methods sections, structured abstracts and anything written to a template are regular by design, and regularity is what detectors read.
Spread evenly across everything suggests a general property of the writing rather than a specific passage. Formal register, consistent sentence length and low vocabulary variation all produce this, and they are properties of a great deal of careful human writing. See why AI detectors flag human writing.
4. Assemble the record the text does not contain
Finished text carries no information about how it was produced. That is a limitation of the method rather than a gap in any particular tool, and it means the record has to come from somewhere else.
What exists, usually without anyone having planned for it:
- Version history. Google Docs keeps revision history automatically. Word tracks changes
when enabled. Both show the document developing over time
- Earlier files. Outlines, notes, partial drafts, anything with a timestamp
- Reference material. Sources you read, in a manager or a browser history
- Correspondence. Supervision notes, feedback, messages discussing the work as it
developed
- Related work. Earlier pieces you wrote, which establish how your writing normally
reads
This is also the practical argument for keeping drafts as a habit. The record is easy to have and impossible to reconstruct afterwards.
5. Know the specific limits of the number
Facts that are worth having to hand, each from the tool vendors themselves:
- Turnitin publishes a false positive rate of under 1% of documents and around 4% of
sentences, and states that documents scoring below 20% show a higher incidence of false positives
- Turnitin's own analysis found 54% of falsely flagged human sentences sit directly beside a
sentence scored as AI written, and 26% two sentences away, so wrong flags cluster around genuine ones rather than scattering
- Detection rates vary sharply by which model produced the text, so a tool's headline accuracy
figure depends mostly on which generator it chose to test against
- Non-native English writing is flagged at higher rates, repeatedly, in published
evaluations. See why non-native English writing gets flagged
What a score cannot establish
A score describes how text reads. Several different situations produce similar scores: text a model generated, text a person drafted and a model edited, text a model drafted and a person edited, and text a person wrote in a formal, regular register. The words carry no record of which.
A model asked to reproduce a human passage verbatim produces output identical to the human original. No detector can separate those, because there is nothing in the text to separate.