If you wrote it yourself and a detector said otherwise, start here. The tool did not find evidence about you. It produced a probability estimate about a piece of text, based on patterns it associates with machine-generated writing. Human prose can carry those patterns for entirely ordinary reasons, which is why this happens to people who did nothing wrong. Clear structure, careful editing, and a conventional topic all push writing toward the regularity these systems tend to notice. There is also documented evidence that some detectors flag writing by non-native English speakers far more often than comparable native-speaker samples, and Turnitin's own guidance is that a score should not be the sole basis for action against a student. This article covers what a flag actually represents, which writing habits tend to trigger one, what evidence helps if you are accused, and what to avoid doing in the meantime.
What a flag actually represents
A detector did not verify who wrote your essay. It analyzed textual patterns and produced a likelihood.
Broadly, tools in this category look for statistical characteristics associated with machine-generated text, such as unusually predictable phrasing, limited variation across a document, or features a trained classifier learned from labeled examples. The specific signals and how they are weighted differ by product, and commercial systems including Turnitin do not publish their full model architecture. Anyone describing exactly how a given detector reaches its conclusion is inferring.
What matters for you is the shape of the reasoning rather than its internals. These systems infer authorship from properties of the text. Human writing can have those properties. That gap is where false positives live, and no amount of confidence in the interface closes it.
Writing habits that can trigger a flag
None of the following are mistakes. Several are things you were explicitly taught to do.
Writing to a formula. Five paragraphs, topic sentence, three supporting points, restated thesis. Structure taught for clarity produces regularity, and regularity is what these tools tend to notice.
Editing until it is smooth. Each round of tightening pulls sentences toward more standard phrasing. Careful revision can make a document read as more conventional than the first draft did.
Writing about a well-covered topic. Photosynthesis, the causes of the First World War, supply and demand. Conventional subjects carry conventional vocabulary.
Writing under a tight word limit. Compression favors efficient, standard phrasing and removes the asides that would otherwise vary the texture.
Writing in a discipline that demands precision. Technical, legal, and medical prose is deliberately regular because ambiguity is dangerous. That regularity can resemble the patterns detectors associate with AI.
We describe these as things that can contribute rather than causes, because no published research isolates any single habit as the trigger, and the detectors are not open to inspection.
If English is your second language
This part has real evidence behind it and deserves stating plainly.
A study published in Patterns evaluated seven widely used GPT detectors against two sets of known human writing. On US eighth-grade student essays, the detectors classified the human samples far more accurately. On the other set, they flagged more than 61 percent of TOEFL essays written by non-native English speakers as AI-generated. The TOEFL essays pre-dated widespread generative AI use, so the flags were errors rather than detections.
The proposed explanation follows from how the detectors examined in that study worked. Those systems responded to how predictable the language was, and second-language writing often draws on a somewhat narrower range of vocabulary and sentence structures. That is a description of learning a language, not a deficiency in the writing.
Vendors are aware of the finding. Turnitin says it took second-language learners into account when assembling its training sample, along with users from non-English-speaking countries and less common subject areas. That is a reasonable response to the research. It is not independent confirmation that the problem is solved, and no public test settles the question either way.
Related situations can behave similarly. Drafting in one language and translating can change sentence patterns and vocabulary in ways that affect detectors, particularly when the translation is heavily polished. That does not mean translated work is always flagged, but it may behave differently from the draft you started with.
Tools you may not have counted as AI
Basic spelling and punctuation correction is unlikely to move a result much, because it changes very little text. The boundary is blurrier than it used to be, and knowing where your own tools sit matters.
Generative rewrite, clarity, tone, and rephrase features produce new sentences rather than correcting existing ones. If you accepted those suggestions across a document, that text was generated, whatever menu it came from.
Machine translation output is produced by a model, and heavily normalized translations can read differently from original composition.
Paraphrasing tools deserve particular care after a flag, for reasons covered below.
What a flag can and cannot support
Compare it with a similarity match. When a similarity check fires, it points to source material: a URL, a document, a page that can be opened and compared line against line. An AI writing report gives a percentage and, above the reporting threshold, highlighted passages. What it does not produce is an external source or any other independently verifiable exhibit showing who wrote the text. That difference is why these cases are so hard to contest, and why being on the wrong end of one feels so unfair.
The vendors say as much. Turnitin advises that its AI score should not be the sole basis for action against a student, and does not attribute a score at all between 1 and 19 percent. What Turnitin's report actually contains covers the reporting rules in detail.
Institutions have acted on this. In August 2023 Vanderbilt disabled the tool and published its reasoning, including a risk estimate: the university had submitted roughly 75,000 papers in 2022, and if the 1 percent false positive rate stated at launch had applied across that volume, around 750 papers could potentially have been flagged in error. That figure is an estimate of exposure, not a count of confirmed cases. Some other institutions have since limited or disabled AI detection as well.
And detectors disagree with one another on the same text often enough that treating any single output as settled is hard to justify.
How to respond if you are accused
Find out what the flag actually was. A percentage, a set of highlighted passages, or an instructor's impression are three different things, and your response depends on which one you are answering.
Gather your process first. Version history in Google Docs or Word, dated drafts, notes, outlines, photographs of handwritten planning, browser history for sources, timestamps on saved files. This is the category of evidence that speaks to authorship, and it is worth assembling before you reply.
Offer to walk through the work. Explaining why you structured an argument a particular way, what you cut, and where a source came from demonstrates something a score cannot measure. In practice it is often the most persuasive thing available.
Reply in writing, factually. Something close to this is usually enough for a first response:
Dear [Instructor],
Thank you for letting me know about the AI detection result on [assignment]. I wrote this work myself and would like to help resolve the question.
I have version history for the document covering [dates], along with my outline and notes, and I am happy to share all of it. I am also glad to meet and talk through how I developed the argument.
Could you tell me which passages were flagged, so I can address them specifically?
Thank you, [Name]
Ask what your institution's procedure is and whether an appeals process exists, and keep the exchange in writing.
What not to do
Do not rewrite the essay and resubmit it. It changes the document under discussion after the fact, it is difficult to explain, and rewriting tools carry their own risks that have nothing to do with honesty.
Do not admit to something you did not do in order to end the conversation faster. Records of these outcomes can persist.
Do not rely on a second detector as your defense. It is another probability estimate with the same limitations as the first.
If this is affecting your wellbeing, your institution's student support or counseling service exists for exactly this kind of situation, and using it is not an admission of anything.
What to keep from now on
Leave version history switched on. If a course requires you to write inside a platform that records the process, such as Turnitin Clarity's composition space, understand that paste events and revisions inside that space can appear in the writing report. A paste event is contextual information rather than evidence of misconduct, and there are ordinary reasons to paste from a notes app or a reference manager. What helps is keeping the drafts and source material that explain your workflow, so the record has an explanation attached to it.
Keep outlines and notes rather than deleting them when the work is done.
Read your own draft through a detector before submitting and look at the flagged passages rather than the total. Unfox is an AI detector that reports results at the sentence level, and checking your own draft before submitting covers what to look for in any tool you use. The point is not a low number. It is knowing in advance which sentences a reader might misread.
None of this is a fair burden. It is currently a practical one.
FAQ
Why was my essay flagged as AI when I wrote it myself?
Because the tool estimated a likelihood from textual patterns rather than verifying authorship. Formulaic structure, heavy editing, a conventional topic, a tight word limit, and precise technical prose can all produce the regularity these systems associate with machine writing. A flag describes the text, not you.
Do AI detectors flag non-native English speakers more often?
Research published in Patterns found that seven widely used detectors flagged more than 61 percent of TOEFL essays by non-native writers as AI-generated, while classifying a comparison set of native-speaker essays far more accurately. Turnitin says it accounted for second-language writers when building its training sample, though no public test confirms the issue is resolved.
How do I prove I did not use AI?
Through process rather than counter-analysis. Version history, dated drafts, outlines, notes, and the ability to talk through your reasoning address the actual question. A second detector report is another probability estimate and carries the same weakness as the first.
Can I be penalized based on an AI score alone?
That depends on your institution's policy. Turnitin advises that its score should not be the sole basis for action against a student, and several universities have written similar guidance into their procedures. Ask what your institution's policy says and whether an appeals process exists.
Does using Grammarly get you flagged?
Basic spelling and punctuation correction is unlikely to affect a result much, since it changes little text. Generative rewrite, clarity, and tone features are different, because they produce new sentences. If those were applied across a document, that text may register the way other generated text does.
Should I rewrite my essay to lower the score?
Not after an accusation. Modifying the document afterwards changes the artifact being discussed and is hard to explain. Before submitting, revising specific sentences in your own words is reasonable. Running the whole essay through an automated rewriter is a different thing and carries its own problems.
