Will AI detectors flag my writing?
It can, even if you wrote every word. A detector grades style, not authorship. It scores how predictable your word choices are and how much your sentence rhythm varies, so the plain prose you were taught to write is what it reads as machine-made, and second-language writers are flagged most. You cannot argue a percentage down. Keep the drafts and the version history, and you have something that answers it.
What an AI detector actually measures
Two questions this page does not answer. Whether your own use of AI was allowed is a policy question rather than a detection one, and if a chatbot drafted any of what you are handing in, the references it gave you are a bigger risk than any score.
A detector does not read your mind, and it does not compare your essay against a database of known AI essays. It has no record of what ChatGPT told you. Instead it runs your text through a statistical model and scores how "machine-like" the patterns look. Two signals do most of the work.
- Perplexity measures how predictable your word choices are. GPTZero describes it as a measure of how likely an AI model would have chosen the exact same set of words as the ones in your document. Low perplexity (very predictable wording) reads as AI.
- Burstiness measures how much your sentence rhythm varies. Human writing tends to mix long, winding sentences with short, blunt ones. AI text is often more even. Low variation reads as AI.
Notice what neither of these checks: whether the ideas are yours, whether you understand the topic, or whether you actually typed the words. A detector is grading style, not authorship. That gap is the root of every problem below, and it is why the things that actually give an essay away are a different list entirely.
Why honest writing gets flagged
Here is the uncomfortable part. The writing habits schools spend years teaching, clarity, simplicity, consistent structure, are the same habits that lower your perplexity score. If you write in plain, direct sentences and avoid flowery detours, you can look "predictable" to a model, which is exactly what these tools associate with machine output.
The clearest evidence is a 2023 Stanford study by Liang, Yuksekgonul, Mao, Wu, and Zou. The researchers ran essays through seven popular detectors and found that the detectors consistently misclassified non-native English writing as AI-generated, while native writing was accurately identified. This is not a small bias. When the team tested real essays written by humans for the TOEFL English exam, the detectors incorrectly labeled more than half of the essays as AI-generated (an average false-positive rate of 61.3%), and at least one detector flagged 97.8% of them. The same tools correctly identified more than 90% of essays by U.S. eighth-graders as human.
The reason is mechanical, not malicious. Second-language writers often use a narrower band of common words and simpler constructions, which produces exactly the low-perplexity signal a detector treats as a red flag. If you write in your second language, or you just write plainly, you are statistically more likely to be flagged, even when every word is yours.
The detectors themselves are not reliable
It is tempting to assume that if a tool exists and a university bought it, it must work. The track record says otherwise.
The most telling example is the company with the most to gain. OpenAI built its own AI Text Classifier, then shut it down on July 20, 2023, citing the tool's low rate of accuracy. By OpenAI's own numbers, the classifier correctly identified only about 26% of AI-written text as "likely AI-written", and it could still flag human writing as machine-made. The maker of ChatGPT could not reliably detect ChatGPT.
Turnitin, the tool most students will actually face, has the same problem in quieter language. Turnitin has acknowledged that its detector produces a higher incidence of false positives in real-world use than in its own lab, particularly when a document contains less than 20% AI writing, the exact situation a student who lightly used AI for grammar would be in.
Why a "1% false-positive rate" is not reassuring
Turnitin points to a low false-positive rate, but small percentages get large at the scale schools operate. Vanderbilt University worked the math on its own campus: applied to the 75,000 papers it submitted in 2022, around 750 student papers could have been incorrectly labeled as having some AI writing. Vanderbilt disabled Turnitin's AI detector for the foreseeable future, concluding it does not believe AI detection software is an effective tool for identifying AI-written work. Vanderbilt is not alone: the University of Pittsburgh's Teaching Center also turned the detector off, telling instructors that current AI detection software is not reliable enough to act as proof. When the customers turn the product off, that tells you something.
What detectors cannot see
Two facts make the false-positive problem worse, not better.
First, detection accuracy collapses once AI text is edited or paraphrased. So the students most likely to get caught are the honest ones who wrote in a flaggable style, while anyone deliberately gaming the system can slip through. The tool punishes the wrong people.
Second, detectors offer almost no transparency. You get a percentage with no explanation of which sentences triggered it or why. You cannot inspect the reasoning, and neither can your professor. A number with no audit trail is a weak basis for a serious accusation, which is why James Zou, one of the Stanford researchers, warned that we should be very cautious about using any of these detectors in classroom settings.
That weakness has now been tested outside a university's own process. In February 2026, a court ruled that Adelphi University's plagiarism finding against a student was without merit and ordered his record expunged, after a detector marked his essay as fully AI-written and the university acted on it. A flag, on its own, did not survive scrutiny.
What actually gives an essay away
Detectors are what students worry about, and prose style is what they think a marker notices. Both are the weakest signals in the room. Researchers at the University of Pennsylvania who study AI text found that humans perform barely better than chance at identifying AI writing, with the gap narrowing only under deliberate training and incentives. A professor who feels that a paragraph sounds like ChatGPT has a hunch. What turns a hunch into a case is the short list of things that survive a second look, and none of them is a writing style.
A citation that points at nothing. This is the one a marker can confirm in minutes, and it catches more students than every other tell combined. A 2025 study in JMIR Mental Health tested GPT-4o across simulated literature reviews and found that about one in five citations were entirely fabricated, and more than half were either fake or carried bibliographic errors such as wrong or invalid DOIs, with close to a third invented on the least-researched topic it tried. A reference is a factual claim with a yes-or-no answer: they search the title, find nothing, check the DOI, find it dead, and there is no innocent reading of that. The mechanism is why AI makes up citations, the check itself is how to check if a citation is real, and the one pass to run over a draft you already have is how to get ChatGPT to cite real sources.
A real source attached to a claim it does not make. The subtler cousin: a genuine paper summarised slightly wrong, or a specific statistic pinned to it that the paper never reports. The reference checks out and the sentence in front of it does not, which anyone who opens the paper sees at once. The University of South Carolina's integrity office puts the responsibility where it lands: you are fully responsible for the information you submit based on a generative AI query. Reading the source before you cite it is the only fix.
An essay that engages with nothing in particular. AI prose is fluent and weightless: it defines terms, surveys both sides, and never grips the question that was set or the reading that was assigned. A 2025 University of East Anglia study compared 145 student essays with 145 from ChatGPT and found that human essays used over three times as many engagement features, the rhetorical questions and personal asides that make writing sound like somebody. This proves nothing by itself. It is what makes an instructor read more closely, and a close read is where the citations get checked.
No trail at all. Real writing leaves outlines, abandoned paragraphs, comments and saved versions. Google Docs revision history and Word's tracked versions record how a piece grew, and a finished essay pasted in one go reproduces none of it. Most of the time nobody asks. When somebody does, the absence is conspicuous and the presence ends the conversation, which is why the trail cuts both ways: it clears the student who did the work as cleanly as it exposes the one who did not.
Put those together and the pattern is worth keeping. The tells that hold up are about substance, and the only ones that "make your essay undetectable" advice addresses are the ones about style. You can paraphrase prose until a detector shrugs while the invented reference is still invented, the source still says something else, and the engagement the assignment asked for is still missing.
What to do if you are wrongly accused
If a detector score gets pointed at your work, do not panic, do not confess to something you did not do, and do not edit the submitted file. A flag is not proof, and you have a stronger case than the score suggests. The one thing worth doing straight away is preserving the trail that answers it: drafts, outlines, notes, and version history from Google Docs or Word, saved somewhere separate while it still exists.
The rest of it, how to reply to the email, what to ask for in writing, what an academic integrity meeting is actually like, and how the advice changes if you did use AI, is a guide of its own: what to do when you are accused of using AI on an essay. It is also worth being sure each reference is genuine before it matters, which is what checking that a citation is real walks through, because a source that does not exist is the accusation nobody has to argue about.
The key point is simple: you cannot prove a negative by arguing about a percentage, but you can show the trail of work that produced your essay. A drafting history that records every change as you go is the cleanest version of that trail, and what it does and does not settle is how to prove you did not use AI. Keep that trail as you write, and an accusation becomes much easier to answer.