Why AI Detectors Flag Non-Native Writers (False Positives)

You wrote every word of your paper. You cited your sources, you revised it twice, and you submitted it with a clear conscience. The report came back marked as likely AI, and now you are trying to prove a negative about your own writing. If English is your second language, this is not rare bad luck. It is a documented pattern.

And here is the uncomfortable truth: an AI detector false positive is far more likely to land on a non-native English writer than on a native one. That is not because non-native writing is worse. It is that these tools measure the statistical texture of language, and clear, careful second-language writing happens to share features with the machine text they were trained to catch.

It is not a fringe complaint from frustrated students. It is the finding of peer-reviewed research, it is why several major universities pulled back from these tools, and it is why courts and administrators have started taking detector output with real caution. This is what the evidence actually says, and what to do if a detector gets you wrong.

What the Stanford study found about detector bias

The strongest evidence comes from a 2023 study by Liang and colleagues at Stanford, published in the Cell Press journal Patterns and titled, plainly, "GPT detectors are biased against non-native English writers."

The researchers ran essays through seven widely used detectors. For essays written by non-native English speakers, taken from a TOEFL corpus, the detectors flagged around 61% as AI-generated on average. For essays by native English writers, the false-positive rate was roughly 5%. Nearly one in five of the non-native essays, about 19.8%, was unanimously flagged as AI by every detector in the test. These were human essays in every case.

The mechanism is the interesting part, because it explains why the bias is baked in rather than accidental. Many detectors lean on a signal called perplexity, which is a measure of how surprising a text's word choices are to a language model. Writing full of common words and predictable phrasing produces low perplexity, and low perplexity reads to the detector as machine-like. Second-language writers often use a more limited, more conventional vocabulary, not because their ideas are simple but because they are writing carefully in a language that is not their first. The detector sees the smooth surface and calls it artificial.

That is the core of the fairness problem. The very habits that make careful second-language prose clear and correct are the habits these tools were built to punish.

Why non-native writing triggers AI detector false positives

It is worth sitting with why this happens, because understanding it takes some of the sting out of a bad flag.

A native speaker writing quickly will scatter idioms, unusual word choices, odd rhythms, and the occasional grammatical shortcut. That unpredictability registers as high perplexity and reads as human. A non-native writer, or frankly any disciplined academic writer, tends to favor precise, standard constructions and a controlled vocabulary. Large language models also favor precise, standard constructions and controlled vocabulary. The detector cannot tell the difference between a careful human and a careful machine, because on the specific axis it measures, they look alike.

None of this means the writing is lower quality. Clarity is a virtue in academic prose, not a defect. It simply means the detector is measuring the wrong thing and drawing a confident conclusion from it.

That confidence is the danger. A tool that outputs a crisp percentage feels authoritative, even when the number rests on a signal that systematically disadvantages a whole population of genuine writers.

Universities that stepped back from AI detectors

Institutions noticed the problem early, and several acted on it.

In August 2023, Vanderbilt University disabled Turnitin's AI detector, and issued an unusually detailed explanation for doing so. It noted that while the claim of a false positive rate of about 1% would not be significant in relation to an entire paper, given that there were some 75,000 submissions per year, the fact that it had been shown to produce biased results against non-native English writers, coupled with the opaqueness of its processes, meant that around 750 papers might end up being erroneously flagged. Michigan State, the University of Texas at Austin, Northwestern, and the University of Pittsburgh took similar steps. Southern Methodist University stepped back in December 2023, and the University of Waterloo followed in September 2025 after finding human-written text scored as fully AI.

Turnitin's guidance is also cautious. Its AI indicator downgrades numbers from 1 to 19%, using an asterisk rather than a number, since false positives are more widespread on the lower end of the scale. It also explicitly warns that the score should not be used alone to make decisions regarding academic integrity. That is a vendor telling institutions not to treat its own output as proof, which is a strong hint about how much weight a single number deserves.

If your school still uses these tools, that is not a reason to panic. It is a reason to know your rights and keep your evidence, which we come back to below.

Real cases, laid out plainly

A few disputes have made the news, and they are worth understanding precisely, because the details matter and the headlines often blur them.

In the case of Orion Newby at Adelphi University, a student essay was flagged as fully AI-generated. A federal judge reviewed the matter and found the finding "without merit," ordering it removed from the student's record, as reported in February 2026. We break down what that ruling did and did not establish in the Newby v. Adelphi ruling. At UC Davis, William Quarterman was flagged by GPTZero and cleared his name using his Google Docs version history, which showed the essay being written over time. That was an administrative resolution, not a lawsuit.

One point needs to be stated plainly, because a lot of online summaries get it wrong. None of these outcomes were legal judgments against Turnitin the company, and no student in these cases "sued Turnitin and won." The disputes were with schools over how a flag was used, and the lesson is about process and evidence, not about a detector vendor losing in court. Framed accurately, the takeaway is empowering: a detector score is a claim, not a verdict, and legitimate writers have successfully contested it.

What to do if you are wrongly flagged

If a detector flags work you wrote, the situation is recoverable, and the writers who resolve it well tend to do the same handful of things.

Keep your drafts and version history. Google Docs and Word both track editing over time, and that timeline is the single most persuasive piece of evidence you can offer, exactly as it was for the UC Davis case. Ask, politely, what score triggered the flag and how the reviewer is interpreting it, since a suppressed or borderline number is weak ground for any decision. Point to the research and the institutional caution above if you need to make the fairness case in writing. Our step-by-step guidance on how to appeal a false AI-detection flag walks through the full process.

Going forward, the goal is not to trick the tool. It is to write in confident, natural English so your own careful prose is less likely to be misread, and to disclose any AI assistance the way your institution requires. Our AI text humanizer is built for exactly that, and if you write in English as a second language, our dedicated guide for AI humanizer for ESL researchers covers the workflow in more depth. The aim is fairness for the work you have already done, not disguise.

Before and after: fixing the flat texture that trips a detector

Theory is easier to trust when you can see it on a real paragraph. Here is a results passage an ESL researcher drafted with an AI assistant to smooth the English. It reads cleanly, and a detector still flagged it.

Before (flagged as likely AI):

"The results demonstrate that the intervention significantly reduced recovery time by 27% across all cohorts (Chen et al., 2021). This substantial improvement clearly highlights the effectiveness of the proposed approach. The findings confirm the hypothesis and underscore the importance of early intervention."

Three sentences, almost identical in length, every clause marching at the same pace. The vocabulary is safe and predictable. That evenness is what a detector reads as low perplexity and low burstiness, and it is the same texture that flags a lot of genuine non-native writing.

After (edited in the writer's voice):

"Recovery time fell by 27% across the three cohorts (Chen et al., 2021), a bigger drop than we had expected (n = 214). The effect held, though it was weaker in the oldest group. That pattern points to a real benefit from early intervention, but with a sample this size we would not push the claim further than the data allows."

Read them back to back and the second one sounds like a person who did the work.

What changed, and why it matters. We kept every load-bearing fact: the 27% figure, the (Chen et al., 2021) citation, the cohort count, and the sample size. Nothing about the evidence moved. What we added was variation. Sentence lengths now swing from long to short to long, which raises burstiness. Word choices are less predictable, which raises perplexity. We also put the hedging back: "points to," "we would not push the claim further." AI drafts strip that caution out, and so do over-smoothing tools, which is one reason polished non-native English drifts toward the AI profile.

This is also a quality edit, not a trick. The after version makes a more measured, more defensible claim than the before version, which overreached with "demonstrate" and "confirm the hypothesis." That is the point we keep making: the edit that lowers your false-flag risk is usually the same edit that would survive peer review. An academic humanizer can do this first pass while protecting your citations and numbers, but you still read every line.

One caution. No single edit guarantees a clean score, and detectors keep changing (Turnitin added humanizer detection in August 2025). Aim for writing that reflects your own reasoning, understand what perplexity measures, and disclose AI help per your journal or university policy.

Frequently asked questions

Q: Why do AI detectors flag non-native English writers?

Most detectors measure how predictable a text's word choices are, and careful second-language writing tends to use common, standard vocabulary that reads as low perplexity. That statistical smoothness looks machine-like to the tool, which produces an AI detector false positive even though a human wrote every word. The bias is built into what the detector measures, not into the quality of the writing.

Q: What did the Stanford study find about detector bias?

The 2023 Stanford study in Patterns tested seven detectors and found they flagged roughly 61% of non-native TOEFL essays as AI, compared with about 5% for native writers. Nearly 20% of the non-native essays were unanimously flagged by every detector. All of the essays were written by humans.

Q: What should I do if wrongly flagged as AI?

Keep your draft history from Google Docs or Word, since a visible writing timeline is the strongest evidence you have. Ask what score triggered the flag, cite the research on false positives, and follow a clear appeal process. A detector score is a claim you can contest, not a final verdict.

Q: Have universities stopped using AI detectors?

Several have limited or disabled them. Vanderbilt disabled Turnitin's AI detector in 2023 over false-positive and bias concerns, and Michigan State, UT Austin, Northwestern, Pittsburgh, SMU, and Waterloo took similar steps. Many institutions that still use detectors now treat the score as one signal rather than proof.

Q: Does writing in simpler English increase your risk of being flagged as AI?

To a degree, yes. Detectors read simpler, more predictable vocabulary as low perplexity, the same statistical signal they associate with AI, which is one reason clear ESL writing draws false positives. The fix is not to inflate your vocabulary artificially, but to restore natural sentence variation and your own hedging and voice.

Q: Can I fix a false AI flag by editing my own writing instead of using a humanizer?

Often you can. Reading your draft aloud, varying sentence length, and putting back the limits you actually mean will raise burstiness and make the text sound like you wrote it. A tool can speed up that first pass, but the goal is genuine voice, and you should still disclose any AI help under your journal or university policy.

A humanizer built for ESL writers

Write in clear, natural academic English and reduce the false positives that hit non-native writers hardest, with meaning and citations preserved.