You probably got flagged because your writing is easy to predict, not because it reads like a machine. AI detectors can’t see who wrote a text. They check how close your word choices, sentence rhythm and structure sit to the most typical way of writing. Plain, formal, polished or second-language writing sits close to that average, so it gets flagged too.
A detector’s score is a guess, not evidence. The companies that make detectors say so themselves. If you’re accused, your drafts and writing history are what count.
How do AI detectors decide your writing is AI?
They measure how predictable your text is to a language model.
A language model can guess the next word in a sentence. For “I poured a cup of”, it will guess “coffee” or “tea”. GPTZero describes perplexity as “how likely an AI model would have chosen the exact same set of words as found in the document.” If your text keeps using the words the model expected, it looks like AI.
They also look at rhythm. People tend to mix it up: a long sentence, then a short one, then a word nobody would guess. Chatbots keep a steadier pace. GPTZero calls this burstiness, “how much writing patterns and text perplexities vary over the entire document” (GPTZero).
Some detectors also look for phrases chatbots overuse and for formulaic structure. “Firstly… secondly… in conclusion” is easy to guess.
Put together, a detector is asking one question: how close is this text to the average? AI writes close to the average by design, because it picks likely words. But many people write that way too.
Can AI detectors be wrong?
Yes, and the people who build them say so.
- Turnitin says that for documents with over 20% AI writing, its document false positive rate is under 1%. Its sentence-level rate is about 4%: roughly 4 in 100 highlighted sentences may be human-written (Turnitin). Its guidance also says the AI report “should not be used as the sole basis for adverse actions against a student” (Turnitin guide).
- Vanderbilt University turned off Turnitin’s AI detector in 2023. It had sent 75,000 papers to Turnitin in 2022. At a 1% false positive rate, “around 750 student papers could have been incorrectly labeled” (Vanderbilt).
- OpenAI retired its own AI text classifier in July 2023 “due to its low rate of accuracy.” It caught 26% of AI-written text and flagged human text as AI 9% of the time (OpenAI).
There’s also a limit no product can fix. Researchers at the University of Maryland tied how well the best possible detector can do to how different AI text is from human text. As models write more like people, that gap shrinks, and detection gets closer to guessing.
What in your writing gets flagged?
Usually it’s writing that is simple, tidy or heavily polished.
Simple vocabulary. In a Stanford study, detectors wrongly flagged 61% of essays by non-native English writers on average. When the researchers had ChatGPT make the same essays’ word choices richer, that fell to 12%. When they had it simplify essays by US eighth graders, those essays got flagged more often (Liang et al., 2023). The detectors were reacting to how complex the language was, not to who wrote it.
Formulaic structure. School essays, cover letters, lab reports and status updates follow templates. Templates are predictable.
Heavy rewriting with AI tools. Fixing typos changes little. When a tool rewrites whole sentences, it replaces your word choices with the ones it finds most likely. That is exactly what detectors look for. The more a tool rewrites, the more your text reads like its output.
Words that are spreading everywhere. A study of more than 15 million PubMed abstracts found “delves” used 28 times more often in 2024 than expected, and estimated that at least 13.5% of 2024 abstracts were processed with LLMs (Kobak et al., 2025). People pick up these words from reading AI text, then get flagged for using them.
Em dashes come up a lot here. They’re not proof of anything. Writers used them long before chatbots.
Who gets flagged more often?
Non-native English writers, and anyone whose writing is plain and regular.
The same Stanford study ran 91 TOEFL essays through seven popular detectors. All seven flagged 18 of them as AI. At least one detector flagged 89 of the 91, or 97.8%. Essays by US eighth graders were classified almost perfectly (Liang et al., 2023).
James Zou, the study’s senior author, explained why: detectors “typically score based on a metric known as ‘perplexity,’ which correlates with the sophistication of the writing — something in which non-native speakers are naturally going to trail their U.S.-born counterparts” (Stanford HAI).
If you write in your second language, you probably use common words and safe sentence patterns. That’s normal. It’s also what detectors read as AI.
Why does one detector say 100% AI and another say human?
Because each one uses a different model and a different cutoff.
The number often isn’t the share of your text that AI wrote. GPTZero’s headline probability, for example, is how likely it thinks the whole document was AI-generated. It says a 50% score could otherwise be read either as a coin flip or as half the text being AI (GPTZero). So “100% AI” means the tool is very sure, not that it’s right.
Run the same paragraph through three detectors and you can get three different answers. That disagreement is a good reason not to trust any single result.
What should you do if you’re accused?
Show how you wrote it. Don’t answer one detector’s score with another detector’s score.
- Open your version history. Google Docs keeps it under File → Version history. Word keeps one for files saved to OneDrive or SharePoint (Microsoft). It shows your text growing over hours or days, not appearing in one paste.
- Gather your drafts and notes. Outlines, sources you read, earlier versions, messages where you talked about the topic. Anything with a date helps.
- Offer to talk it through. Explain why you made your argument, where a source came from, what you cut. If you wrote it, this is usually easy.
- Bring the detector makers’ own words. Turnitin’s line about the report not being the sole basis for action, and the Stanford study if English is your second language.
If you wrote in Google Docs, the Draftback Chrome extension can replay the document’s revision history like a video. Teachers use it to check for AI. You can use it to show your process.
Should you change how you write to avoid getting flagged?
No. Writing for a detector makes your writing worse, and it doesn’t make you safe.
The worry is real. In a 2026 YouGov survey of 2,373 UK university students, commissioned by Studiosity, 75% of students who use AI said they worried about being wrongly flagged for plagiarism by detection tools (Times Higher Education).
But once a score becomes the goal, you start chasing the score instead of writing well. Tools that promise to beat detectors rewrite your text toward whatever the detector rewards today. When detectors change, the target moves. Your meaning can get lost along the way.
What makes writing yours is the part a detector doesn’t measure: the detail only you know, the judgment you actually hold, the example from your own week. Keep those in. If you use AI to help you write, start from your own notes and opinions so it doesn’t replace them. I wrote about how in How to Write with AI Without Losing Your Voice.
And keep your drafts. They protect you better than any rewrite.
Frequently asked questions
What percentage of AI is acceptable?
There's no shared standard. Your school, employer or client sets the rule, so read their policy. Low scores are the least reliable: Turnitin marks scores under 20% with an asterisk for that reason.
Does Grammarly get flagged as AI?
Fixing spelling and grammar barely changes how predictable your text is. Letting a tool rewrite whole sentences or paragraphs swaps your word choices for the model's, and that can raise a detector's score.
Do em dashes mean AI wrote it?
No. Some chatbots use them a lot, but writers used em dashes long before chatbots existed. One punctuation mark tells you nothing about who wrote a text.
Can teachers prove you used AI?
Not with a detector score alone. Turnitin's own guidance says its AI report should not be the sole basis for adverse action against a student. Teachers usually also look at drafts, version history and how well you can talk through your work.
Why does ZeroGPT say everything is AI?
Each detector uses its own model and its own cutoff, so the same text can get very different scores. A high score from one tool and a low score from another is common. Neither result is proof.