Why AI Detectors Flag Real Human Writing

AI detectors don't detect AI. They score predictability.

An AI detector has never seen your ChatGPT window, your drafts, or your keystrokes. It sees the finished text and nothing else, so it cannot answer the question everyone thinks it is answering. What it computes instead is how statistically ordinary the writing is: how closely each word follows the word a language model would have picked, and how little that pattern shifts from sentence to sentence. Text low on both measures gets scored as AI. That is the whole mechanism.

The problem is that statistically ordinary and machine-written are not the same category. They overlap enough that detectors do catch raw, unedited model output. But they are not identical, and the gap between them is where real students get flagged.

What is perplexity, in plain language?

GPTZero, which popularized the term in this context, defines perplexity as "a measure of how likely an AI model would have chosen the exact same set of words as found in the document." Low perplexity means the text is easy to predict. High perplexity means it keeps surprising the model.

Feed a language model the fragment "In conclusion, this essay has" and ask what comes next. Most of its probability lands on a small cluster: demonstrated, shown, argued, explored. Continue with "demonstrated" and the model was not surprised, so perplexity at that position is near zero. Continue with "run out of patience" and it spikes. A detector does this at every word position in your document and averages the surprise.

Language models produce low-perplexity text by construction. Generating means repeatedly sampling from a probability distribution over the next word, and most sampling settings favor the higher-probability options. The output is, on average, the text the model already expected. That is not a flaw a detector discovered. It is how generation works.

What is burstiness, and why does even writing look suspicious?

Burstiness is the second axis. GPTZero defines it as "a measure of how much writing patterns and text perplexities vary over the entire document." In practice you see it in sentence length and rhythm: whether a piece surges and stalls, or holds one steady pace from the first line to the last.

Human writing is bursty because human attention is. You labor over a definition, then dash off an aside. A twenty-eight-word sentence gets followed by a four-word one. A model has no attention budget to run down, so it holds register, clause depth and sentence length remarkably steady.

Two passages carry the same three facts. The first: "The library extended its opening hours in response to sustained demand from students during examination periods. The change was welcomed by undergraduates who had previously struggled to find quiet study space in the evening. Staff reported that usage of the upper floors increased in the first month after the new schedule began." Sentence lengths: sixteen, eighteen, eighteen.

The second: "The library stays open later now. Students asked for it, over and over, through three exam seasons in which the only quiet room on campus shut at six. It worked. Upstairs filled up within a month." Sentence lengths: six, twenty-two, two, six. Same facts, and neither version is wrong. But the first holds a near-constant pace and a near-constant clause shape, and the second does not. That difference is burstiness, and it is what the score reads.

You can run that check on your own draft: count the words in each sentence of a paragraph and look at the spread. If nearly every sentence lands within a few words of the average, your text is low-burstiness, whoever wrote it.

The experiment that shows detectors score style, not authorship

If a detector scores style rather than authorship, then changing only the style of a document, while holding the author and the argument fixed, should change the verdict. That experiment has been run. A 2023 study in the journal Patterns, by Weixin Liang and colleagues at Stanford, put seven widely used GPT detectors over two sets of unambiguously human writing: 91 TOEFL essays and 88 US eighth-grade essays from the Hewlett Foundation's essay scoring dataset.

On the eighth-grade essays the detectors were close to flawless. On the TOEFL essays, written by non-native English speakers, they misclassified more than half, at an average false positive rate of 61.22%. All seven unanimously flagged 18 of the 91, and 89 of the 91 were flagged by at least one. The unanimously flagged essays had significantly lower measured perplexity than the rest.

Then the researchers ran the controlled version. They took the eighth-grade essays, human-written and correctly classified, and used ChatGPT to rewrite only the vocabulary, with the prompt "Simplify word choices as if written by a non-native speaker." The arguments, the structure and the authors were unchanged. Misclassification as AI rose from an average of 5.19% to 56.65%, and measured perplexity fell. Run in the other direction, with a prompt asking for word choices closer to a native speaker's, the TOEFL essays' average false positive rate fell from 61.22% to 11.77%.

Sit with what that experiment holds constant. Same writers. Same ideas. Same paragraph order. Only the words changed, and the verdicts flipped in both directions. A detector score is a measurement of vocabulary and rhythm. It is not, and structurally cannot be, a measurement of who did the writing.

Which kinds of real human writing score like a machine?

Once you know the score is built from predictability and evenness, you can predict who gets flagged without running a test.

Formulaic academic prose is first in line. Literature reviews, methods sections and lab reports are written to a template on purpose, in a register where deviation is a fault rather than a virtue. "The present study examines" is a low-perplexity opening precisely because it is the phrase the field has agreed on.

Writers drilled in the five-paragraph essay come next. Topic sentence, three supports, transition, restate. That structure produces paragraphs of near-identical length with near-identical internal shape, which is a textbook low-burstiness signature. Students who internalized it because it was the reliable way to pass an exam are penalized for having learned it well.

Technical, legal and scientific registers are flagged for the same reason. In a contract or a lab protocol, reaching for a synonym is an error: the same object must be called the same name every time. Consistent terminology is exactly what a perplexity model reads as predictable.

Heavily edited text drifts the same way. Every pass of a grammar or style checker nudges a sentence toward the more conventional phrasing, which is by definition the more predictable phrasing. The University of Nebraska-Lincoln's Center for Transformative Teaching warns instructors that sections rewritten using Grammarly or Word's suggestions can come back flagged at moderate to high confidence. Polishing is not misconduct, but it walks your text toward the threshold.

The same mechanism catches anyone writing in a language whose vocabulary they are still building, since a smaller pool of active word choices is a more predictable one, and what to do when that lands on your own submission is covered in why essays get flagged as AI.

Why the writing school rewards is the writing detectors flag

There is an uncomfortable symmetry here. Open almost any essay rubric and you will find instructions to write clear topic sentences, use explicit transitions, hold a consistent formal register, keep parallel structure, and avoid slang and digression. Every one of those instructions reduces perplexity, burstiness, or both.

So the student who follows the rubric most faithfully produces the most machine-like statistical profile in the class, and the student with a loose, idiosyncratic voice produces the least. Neither fact says anything about who used a chatbot.

Vocabulary compounds it. Academic English has a stock register, and both trained students and language models reach for it, because both learned it from the same corpus of published writing. To see how much of that register is sitting in a draft, an AI word checker lists the specific phrases doing the damage. Replacing them with your own words is good editing regardless of what any detector thinks.

How a 1% error rate becomes thousands of flagged papers

Every claim about detector accuracy is a rate, and rates mislead badly at institutional scale. This is the base-rate problem, and two universities have published the arithmetic themselves.

Turnitin's stated document-level false positive rate was 1%. Vanderbilt University submitted 75,000 papers to Turnitin in 2022. One percent of 75,000 is 750 papers that could have been incorrectly labeled as containing AI writing, in one year, at one university, taking the vendor's own number at face value. Vanderbilt published that sum in August 2023, when it disabled the detector.

Washington State University ran the same calculation in a February 2026 provost's office memo canceling its Turnitin AI detection contract. In Fall 2024 alone, Turnitin analyzed 148,547 assessments at WSU. At a 1% false positive rate, the memo states, "almost 1,485 assessments were likely flagged by Turnitin as AI generated when they were not."

Now take the arithmetic to the vendor's own volume. Turnitin's press release marking the first anniversary of its AI writing detector, in April 2024, reported over 200 million papers reviewed, about 11% of which carried at least 20% AI writing. Set that 11% aside and apply the universities' one percent to what remains: roughly 1.8 million papers. The rate stays reassuringly small. The count does not.

The WSU memo also records what a flag costs once it enters a formal process. Between 2023 and 2025, it says, 33% of all Academic Integrity Hearing Board cases at WSU involving alleged AI misuse ended in a finding of not responsible, because the AI detection had been submitted without any other supporting evidence.

Why detectors can't simply turn false positives off

The obvious response is to make the detector stricter. That does not help, because every detector sits on one dial: a false positive at one end, a missed AI document at the other. Moving away from one error moves you toward the other.

Turnitin has been unusually open about where it set that dial. Its chief product officer, Annie Chechitelli, described the tradeoff plainly in an interview with BestColleges: the company estimates it finds about 85% of AI writing and lets roughly 15% go by, specifically to hold its false positive rate below one percent. That is a deliberate decision to miss AI rather than accuse humans, and it still does not take false positives to zero.

Watch the guardrails tighten around that decision. In June 2023 Turnitin acknowledged a higher incidence of false positives on documents where less than 20% of the writing was flagged, with Chechitelli telling Higher Ed Dive that "real-world use is yielding different results from our lab." The company doubled the minimum submission length for an AI score, from 150 words to 300, and began printing an asterisk next to any score under 20%. It went further a year later: the University of Georgia's Center for Teaching and Learning records that as of July 8, 2024, scores between 1% and 19% are replaced by an asterisk entirely, "to indicate that the score is less reliable."

Read that sequence for what it is. A vendor that first marked its low scores as unreliable, then stopped showing them at all, and doubled its minimum input length along the way, is telling you the signal is weakest on exactly the material students submit every day. UGA draws the instruction that follows: Turnitin scores should not be used as a definitive measure of misconduct. For the published numbers behind the product, how accurate Turnitin is works through them.

What the people who build detectors say about their limits

OpenAI built a detector for its own models and then withdrew it. The AI Text Classifier launched on 31 January 2023 and was retired on 20 July 2023 as "no longer available due to its low rate of accuracy." The numbers were public from the start: on a challenge set of English texts, it correctly identified 26% of AI-written text as likely AI-written, and incorrectly labeled human-written text as AI-written 9% of the time. OpenAI said at launch that it is impossible to reliably detect all AI-written text.

GPTZero is similarly direct in its own documentation. "AI detectors work on probabilities," it says, and a high AI-likelihood score "means the text shares patterns common to AI writing. That is very different from being able to say that it was definitely produced by an AI." It adds that its results belong inside a wider review process alongside human judgment.

Independent testing has not been kinder. Debora Weber-Wulff and colleagues tested 12 publicly available detection tools plus Turnitin and PlagiarismCheck for the International Journal for Educational Integrity in 2023, and concluded that "the available detection tools are neither accurate nor reliable and have a main bias towards classifying the output as human-written rather than detecting AI-generated text." The University of San Diego Legal Research Center's guide to AI detectors sets Turnitin's stated false positive rate of less than 1% against a Washington Post test of the tool that produced a much higher rate of 50%.

There is a theoretical result too. Sadasivan and colleagues at the University of Maryland showed that a recursive paraphrasing attack dropped a zero-shot detector's AUROC from 96.5 to 25.2, below the 50 that random guessing would produce. They also proved a bound: as the text a model produces converges on the text humans produce, the performance of even the best possible detector approaches random. Detection gets harder as models improve, for mathematical reasons rather than engineering ones.

Institutions have drawn the conclusion. Vanderbilt disabled Turnitin's AI detector in 2023, Yale's Poorvu Center states that "the AI detection feature of Turnitin is currently disabled," and WSU's 2026 memo names UC Berkeley, Colorado State, Indiana, Michigan State, Oregon State and the University of Washington as R1 peers that had already stopped using it in integrity cases.

What a high score actually tells you about your prose

All of this leaves a detector score as a style report rather than a verdict, and a style report is something you can run on yourself first. Humanit's free AI detector takes up to 300 words per check with no account and no cost, and it returns subscores showing which signal is pushing the number up, whether that is predictable phrasing, uniform sentence rhythm, or stock vocabulary.

If the writing is genuinely yours and it still scores high, the score is telling you something about the prose rather than about you: it is unusually even. The fixes are the ones this article has been describing. Vary your sentence lengths. Cut the stock academic phrases. Let one paragraph run short. Put a specific detail where a generalization sits. Each of those raises the exact properties a detector measures, and each makes the writing better to read.

What to do once a flag has already landed on work you wrote is a different problem, handled in what a detector flag actually means.

If the text is AI-assisted and your institution's policy permits you to edit and rewrite AI output, rewrite it properly. Rewriting AI text into your own voice is a genuine editing task, and no tool can promise you a particular detector outcome, because detection is probabilistic and the thresholds belong to your institution. The honest version of this advice does not change: know your course policy, do the work, and check it yourself before you submit.