False positives: what an AI flag actually measures
If you're asking why your writing is flagged as AI when every word is yours, start with what the detector actually saw. It didn't watch you write. It saw finished text and estimated how closely it matches the patterns language models leave behind: predictable word choices, even sentence rhythm, stock phrases.
That makes a flag a guess about style, not a record of what happened. Careful, conventional human writing often fits those patterns, which is why false positives happen. A free AI detector that marks each sentence, like ours, at least shows you which lines tipped the score instead of handing you one number. For the mechanics behind it, see how AI detectors work.
Why AI detectors flag your writing: six common causes
Six things make human writing look predictable to a detector. You'll probably recognize at least one.
1. A formal, formulaic register
School and work writing rewards safe phrasing: “furthermore”, “plays a significant role”, topic sentences that restate the prompt. Language models learned from huge amounts of exactly that kind of text, so they reach for the same phrases. Write in the expected register and your sentences become easy to predict. Easy to predict is what a detector scores as AI.
2. Templates and the five-paragraph essay
Structure counts too. An introduction ending in a three-part thesis, three body paragraphs of about the same length, and a conclusion that restates the thesis is also ChatGPT's default shape. If a rubric taught you that shape, your essay can look machine-made at the paragraph level even when every sentence is yours.
3. Grammar and rewrite tools smoothing your text
Grammarly can rewrite whole sentences, and QuillBot is built to paraphrase them. When you accept a full-sentence rewrite or run a paragraph through a paraphraser, that sentence carries the tool's word choices instead of yours. Detectors pay attention to this: Turnitin, for example, says its model also looks for AI text reworked with paraphrasing tools such as QuillBot. Your own sentences, once reworded by the same tools, can pick up the same traits.
Spelling and comma fixes are one thing. Heavy rewrites can be exactly what the flag is pointing at.
4. Writing in a second language
This cause has the strongest evidence behind it. In a 2023 Stanford study published in Patterns, seven widely used detectors labeled more than half of 91 TOEFL essays by non-native English writers as AI-generated, an average false positive rate of 61.22%. Essays by US eighth graders were classified almost perfectly, and 89 of the 91 TOEFL essays were flagged by at least one detector.
The researchers traced the gap to vocabulary. A smaller, safer range of words makes text more predictable. When they had ChatGPT simplify the word choices in the US students' essays, the average share flagged as AI rose from 5.19% to 56.65%.
5. Short texts
A paragraph or two gives a detector very little to go on, so a few generic sentences can swing the result. Turnitin raised its minimum from 150 to 300 words based on its own testing, and our detector won't score anything under 40 words for the same reason. A short discussion post or abstract is the easiest kind of text to misjudge.
6. Common topics and familiar wording
Millions of people have written about social media and teenagers, or the causes of the First World War. The obvious points, and the obvious ways to phrase them, are the ones a model has seen most. In a well-known example, Ars Technica showed detectors labeling part of the US Constitution as AI-written, largely because that text appears so often in the data these models learn from.
My essay was flagged as AI but I didn't use AI: what to do
Stay calm and reply in writing
A flag feels like a verdict, but even the companies that build detectors say it isn't one. Turnitin tells instructors its model may not always be accurate and should not be the sole basis for adverse actions against a student. Reply politely, say plainly that you wrote the work yourself, and ask to talk it through. Keep the message short. Your evidence will do the arguing.
Gather evidence of how you wrote it
The best proof that you wrote something is the trail you left while writing it. Collect what you have before any meeting.
| Evidence | Where to find it | What it shows |
|---|---|---|
| Version history | Google Docs keeps it automatically; Word keeps it for files saved in OneDrive or SharePoint | The text growing over hours or days, with deletions, typos and rewrites |
| Drafts and outlines | Older files, notebooks, photos of handwritten plans, emails to yourself | How your argument changed before the final version |
| Notes and sources | Highlighted PDFs, library loans, browser history, your reference manager | The reading behind your claims |
| Feedback along the way | Comments from a tutor, the writing center or a classmate | Other people saw the work in progress |
Don't create anything new after the fact. A backdated “draft” will hurt you far more than a missing one.
Ask which tool flagged it and what the score was
Ask which detector was used, what score it gave and which passages it highlighted. The details matter. A borderline score is weaker evidence than a high one, and Turnitin itself marks scores under 20% with an asterisk because its testing found more false positives in that range. If the highlighted passages are the formulaic parts, such as your introduction or closing summary, point that out.
For how much weight any score can carry, read our guide on whether AI detectors are accurate.
Offer to explain your work in person
A conversation is often your strongest evidence. Offer to walk through your argument, explain why you picked your sources, and say what you cut and why. Someone who wrote an essay can usually discuss it in detail without notes. If it helps, offer to write a short paragraph on the topic in front of your teacher.
Know your school's policy and appeal process
Read your institution's academic integrity policy before the meeting, especially the parts on AI and on appeals. Look for what counts as evidence, who decides, and how long you have to appeal. If the decision goes against you, ask for it in writing and use the formal process. A student advocate, ombudsperson or academic adviser can help you prepare.
Don't run a flagged essay through a humanizer to “fix” it. Changing the text after the fact doesn't show you wrote it, and it can make your situation worse. The LumenWrite humanizer is for drafts where you're allowed to use AI. If you're accused of using AI when you didn't, your writing process is what clears you.
How to make your writing sound more like you
None of this guarantees a detector will agree with you, and you shouldn't have to write for a machine. But the habits that make writing less predictable also make it better to read.
- Use specifics. “My bus was late four times in March” says more than “public transport is often unreliable”. Names, numbers, places and titles are hard to predict.
- Bring in your own examples. Something from class, your job or a conversation. A model can't invent your experience, and readers can tell when it's real.
- Vary sentence length. Follow a long sentence that builds an argument with a short one. Like this.
- Cut stock phrases. “Plays a significant role”, “a wide range of”, “serves as a reminder”. Say the plain thing instead.
- Take a position. Model-written text balances everything. If one argument is stronger, say so and say why.
- Let paragraphs be uneven. Not every paragraph needs the same length or a neat closing line, and you can usually drop the one that repeats what you just said.
- Keep your process visible. Write in Google Docs, or in Word with the file saved to OneDrive, so version history builds up as you go.
If you're a student checking an essay before you hand it in, our page on the AI detector for essays explains what the score means for coursework and how to read the sentence highlights.