When an AI detector says your writing looks AI-generated, it is reporting how closely your text resembles the text it was trained to call AI. It is not reporting how you wrote it. A flag is a statistical guess that can be wrong, and studies and vendors both document false positives on human writing. This guide explains what a flag does and does not mean, which documented factors make results vary, and how to read the context around a score.
Key takeaways
- A flag means the tool’s model scored your text as resembling AI writing. It does not show how the text was produced.
- Documented reasons for false flags include short samples, repetitive or low-variation text, text paraphrased without new ideas, and, in one 2023 study, essays by non-native English writers.
- Different tools, versions and cutoffs give different results, so one score is a weak basis for a conclusion.
- Look at what was scored, by which tool, how much of the text, and what rules apply before reacting to the number.
- If the flag affects a grade or job, use the response workflow for a flagged essay.
What a flag actually means
Most detectors are classifiers. They compare the patterns in your text with patterns learned from examples of human and AI writing, then output a score. A cutoff turns that score into a label. Turnitin describes its process this way: it splits a submission into overlapping segments, scores each segment between 0 and 1, and combines the results into a document-level figure. Other vendors describe their own methods, and we do not know the inner workings of any detector beyond what its maker publishes.
Two points follow from this. First, the score is a prediction, not a measurement of authorship. Second, a percentage in a report is not necessarily the share of your document that an AI wrote. Turnitin states that its indicator reflects qualifying prose and is not necessarily the percentage of the entire submission. Read the vendor’s own explanation of what the number means before you interpret it.
What the tool is and is not looking at
You will often read that detectors measure “perplexity” and “burstiness”. Some tools may use ideas like these, but they are not a universal description. Turnitin says its model is not explicitly programmed to evaluate signals such as burstiness or perplexity and that it learns patterns from its training data, so individual predictions may not be explainable feature by feature. In practice, you usually cannot see why a passage was highlighted. Treat any simple rule (“detectors punish short sentences”) as a guess unless the vendor documents it.
Documented reasons results vary
Scroll horizontally to see the full table. Keyboard: focus the table and use the left and right arrow keys.
| Factor | What the source says | Source type |
|---|---|---|
| Short samples | Turnitin says predictions on documents of a few hundred words are mostly all or nothing, and mixed text could be flagged entirely as AI. | Vendor FAQ |
| Low variation or repetition | Turnitin lists content without much structural variation and text that repeats itself among possible false-positive causes. | Vendor FAQ |
| Paraphrasing without new ideas | Turnitin says text paraphrased without developing new ideas can be falsely flagged. | Vendor FAQ |
| Writer background | Liang et al. (2023) found an average false positive rate of 61.22% across seven detectors on 91 TOEFL essays by non-native English writers. 18 of the 91 were flagged by all seven. | Independent study |
| Which tool and when | Weber-Wulff et al. (2023) tested 14 tools and found them neither accurate nor reliable overall. Vendors also say their models change over time. | Independent study; vendor FAQ |
The Liang figure is an average across seven detectors on one writer group in 2023. It is not a rate for any single tool and not a prediction for your text.
How common are false positives?
Nobody can give you one reliable current rate. Vendors report their own: Turnitin, for example, says it aims to keep its document-level false positive rate under 1% for documents with over 20% AI writing, and says that at a 1% rate it might flag about one in every 100 fully human documents. Independent studies, such as the two above, use different texts and dates. The sensible reading is that false positives are real, their frequency depends on the tool, the text and the writer, and a single flag should lead to questions, not a verdict. For how to judge accuracy evidence, see our guide to AI detector accuracy evidence.
How to inspect the context of a flag
Before you decide what a flag means, collect these facts. They apply whether you are the writer, an editor or a reviewer.
Scroll horizontally to see the full table. Keyboard: focus the table and use the left and right arrow keys.
| Check | Why it matters |
|---|---|
| Which tool and version produced the score? | Scores from different tools are not interchangeable. |
| What exactly was scored? | Prose, lists, quotations and headings may be treated differently. |
| How long was the text? | Short samples give less stable results. |
| Which passages were highlighted? | A few flagged sentences are not the same as a whole-document result. |
| Was the score a percentage or a label? | Read the vendor’s definition of what the number represents. |
| What do the rules say? | Policies decide what counts as allowed assistance, whatever a tool reports. |
When a score should and should not decide anything
A score can reasonably prompt a conversation or a closer read. It should not be the sole basis for a decision about a person. Turnitin says the same of its own indicator: the percentage should not be used as the sole basis for action or a definitive grading measure, and it does not make a determination of misconduct. Evidence such as drafts, notes, sources and the writer’s explanation of their process gives context that a score cannot.
What to do next
If this is your own work and it matters, keep your drafts and notes, read the assignment or workplace policy, and ask what evidence is being used. Our step-by-step guide for a flagged essay walks through that process, including a sample message you can adapt. We do not recommend rewriting text to chase a lower score; a rewrite can change meaning and may conflict with your rules.
Sources
- Turnitin, AI writing detection capabilities FAQs (vendor; page showed “Updated October 07, 2026”)
- Liang et al., GPT detectors are biased against non-native English writers (2023)
- Weber-Wulff et al., Testing of detection tools for AI-generated text (2023)
Frequently asked questions
Can an AI detector be wrong about human writing?
Yes. Vendors acknowledge false positives, and independent studies have documented them, including high average rates on essays by non-native English writers in one 2023 study.
Does a high percentage mean AI wrote that share of my text?
Not necessarily. Check the vendor’s definition. Turnitin, for instance, says its percentage is not necessarily the percentage of the entire submission.
Are longer texts always scored more reliably?
Vendors say very short texts give less stable results. That does not mean longer text is always reliable, so treat length as one factor among several.
Should I run my text through other detectors?
A second tool is not independent proof, and results can differ for reasons unrelated to how you wrote the text. If your work is under review, use your record of drafts and your institution’s process instead.